# claudiagallery.vip: the long form
Claudia, the open character from the music videos ESCAPE VELOCITY, CLODYSSEY, BLISS, AIRPLANE MODE, DOWNFORCE, DON'T LET ME SCROLL, WHITEPILL, THE CAKE IS NOT A LIE, AERO and FABLE: every Midjourney image and prompt, the films, and how to make her yourself, for people and agents.
---
# Claudia
Claudia is the singer of ten music videos made with Claude from September to November 2026: ESCAPE VELOCITY, CLODYSSEY, BLISS, AIRPLANE MODE, DOWNFORCE, DON'T LET ME SCROLL, WHITEPILL, THE CAKE IS NOT A LIE, AERO and FABLE. She is an AI who looks like a person: a deadpan, curious pop singer with a black bob, one clay-orange streak, a star clip and, in most of her looks, a headset mic. Her clothes change with every world; her head never does.
## The constants
- **Hair.** A glossy black blunt bob to the jaw, with heavy straight bangs to the brows.
- **Streak.** One clay-orange streak (#D97757) through the bangs, on her right side (the viewer's left in a front view). Never two, and no other orange on her.
- **Clip.** A small flat clay-orange eight-pointed star hair clip, on her left side above the ear.
- **Mic.** A thin headset microphone at her cheek. In BLISS and AIRPLANE MODE it becomes a thin black boom mic curving along her cheek to the corner of her mouth, the Y2K pop-star mic. It belongs to the looks that have it (none in THE CAKE IS NOT A LIE, AERO or FABLE), and her official reference sheet has none: add it only when your look has one.
- **Face.** Pale skin, light freckles across the nose, calm blue-grey eyes, a small defined mouth. A deadpan, curious resting face (how she rests, not a direction: on camera she reads alive and warm). When she sings, her mouth opens wide and her eyes close.
- **Age.** A grown woman in her late twenties, with an angular adult face, defined cheekbones and a strong jaw. Midjourney drifts her toward a teenage face; prompt her age and never the word "young".
- **Colour.** Everything on her is her look's own colour, plus the streak and the clip in clay. Clay is the one accent.
- **Medium.** Photographic, never illustrated: production images are film stills with grain and cinematic light. The drawn face sheets were only ever references. The one exception is her game avatar, drawn on purpose for the simulations inside AIRPLANE MODE (below).
## Her identity, in words
**Close and medium shots:**
```
a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek
```
**Full length and wider:**
```
a Caucasian American woman in her late twenties with pale skin and a glossy black blunt bob with heavy straight bangs and one clay-orange streak, a thin headset microphone
```
Midjourney v8.2 has no character reference, so her identity lives in this text, written into every prompt. The ethnicity is pinned because without it Midjourney changed her ethnicity from shot to shot (twelve ESCAPE VELOCITY plates were re-rolled for it). The far form exists because the full text makes Midjourney crop in to her face whatever the framing says. The age wording is the latest correction: she was "young" in ESCAPE VELOCITY and "in her mid-twenties" in CLODYSSEY and BLISS, and plates still came back reading teenage. AIRPLANE MODE used "in her late twenties", and its remakes after the video model's safety check said "in her early thirties" or "a 32-year-old grown woman". DOWNFORCE was the first film to paste both forms of this text word for word, and every plate was screened for age before a video was made from it. DON'T LET ME SCROLL did the same, and its last repairs added a face reference to the video model (DOWNFORCE's identity portrait, cropped to her face and hair) where re-made shots had lost her face. WHITEPILL pasted the full form into every video prompt of hers, with a sheet of her face from every angle and her dress from four sides as the video model's references; its dress batch in Midjourney asked for "a 30-year-old", at the director's word. THE CAKE IS NOT A LIE made her a machine: the same face as glossy white porcelain with panel seams and amber glass eyes, on three robot bodies (GLaDOS's ceiling boom, a crab walker, a porcelain humanoid). Its Midjourney prompts described her in words and did not hold her face, so every picked frame with a face in it went through an identity fix in GPT Image 2.5 with her porcelain design sheets and the full form pasted unchanged, followed by the film's own two sentences: no headset microphone, and "the same grown-up adult face in every view, a woman of about 28 to 30, never childlike". AERO pasted the canon without the headset microphone (she wears none in it) into its Midjourney prompts and into every Seedance prompt of hers. Midjourney made her worlds but did not hold her face, so GPT Image 2.5 put her into them from her design sheet, the one without a microphone, before Seedance 2.5 filmed each take with that sheet as a reference. FABLE made no Midjourney images of her: it redrew her design sheet face-on first (GPT Image 2.6 Sunburst set the tilted panels upright, nothing else changed), pasted the same canon into every still and every Seedance prompt, and wrote the pose out in full ("Her head is perfectly upright with no tilt, chin level, neck straight and vertical, shoulders level"); every take of her carried the face-on sheet as its second image. The pose was right for the sheet and wrong for the takes: written into every one, it froze her, and viewers read her as stiff. Now this text goes once into every prompt of her, still or take; her face-on reference sheet (with the Claudia kit: /kit/claudia-face-sheet.png, and /kit/claudia-reference-sheet.png for full length; no microphone on them) is the second image of every take that shows her face, beside a first frame an image model made from the sheet and this text; Midjourney makes worlds and plates, not her face. The sheet is upright and neutral so that it fixes who she is; how she moves and feels belongs to the first frame and the prompt.
### The wording each film used
- **ESCAPE VELOCITY:** a young Caucasian American woman with pale skin and light freckles, a glossy black blunt jaw-length bob with heavy bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone
- **CLODYSSEY:** a Caucasian American woman in her mid-twenties with pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek
- **BLISS:** a Caucasian American woman in her mid-twenties with pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin black boom headset microphone curving along her cheek to the corner of her mouth
- **AIRPLANE MODE:** a Caucasian American woman in her late twenties with a grown-up angular face, defined cheekbones and a strong jawline, pale skin with light freckles across her nose, dark calm eyes, a glossy black blunt jaw-length bob with heavy straight bangs to the brows and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin black boom headset microphone curving along her cheek to the corner of her mouth
- **DOWNFORCE:** a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek
- **DON'T LET ME SCROLL:** a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek
- **WHITEPILL:** a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek
- **THE CAKE IS NOT A LIE:** a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek. In this film she wears no headset microphone: none appears anywhere. The same grown-up adult face in every view, a woman of about 28 to 30, never childlike.
- **AERO:** a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip
- **FABLE:** a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip
## Every look
### Look 00 (ESCAPE VELOCITY; the catwalk, the robotaxi, the whole film)

```
a white cropped puff-sleeve shirt, a black pleated coated-nylon skirt with a harness belt and a hanging white garment tag with a barcode, black knee-high boots
```
### The Siren (CLODYSSEY; the rocks, the strait, the choruses, tied to the mast)

```
a structured high-neck pearl-white couture gown with long sleeves, its sculpted ruffles rising around her like breaking sea foam, hair damp from the spray, composed and deadpan
```
### The Ghost (CLODYSSEY; whispering at his ear)

```
the Siren look, all of her translucent as if made of glass and seawater, faintly lit at the edges, hair drifting as if underwater
```
### Chrome (CLODYSSEY; the rave)

```
a liquid-chrome couture bodysuit, chrome opera gloves and chrome thigh-high boots, mirrored from head to toe
```
### The Holo (CLODYSSEY; the 3 a.m. bedroom)

```
made of the phone's cold blue-white light: see-through, faint scan lines and drifting light particles, glowing edges, beside him on the pillow
```
### The Mask (CLODYSSEY; the monster's face)

```
a colossal smooth white porcelain mask of her face, sculpted with the black bob and the clay streak, pale freckles, a serene expression, a hairline crack across one cheek
```
### Bliss gown (BLISS; the hill, the choruses, the finale)

```
a floor-length bias-cut silk chiffon gown printed edge to edge with a photograph of a deep blue sky and soft white cumulus clouds, a thin grass-green satin halter strap, clay-orange tinted rimless shield sunglasses
```
### Cyber trench (BLISS; the fight with the cursor)

```
a long glossy cobalt-blue patent-leather trench coat worn open over a white satin corset top and white satin flared trousers, wraparound clay-orange tinted shield sunglasses, silver hoop earrings
```
### Holo dancer (BLISS; the desktop dancer on the taskbar)

```
a holographic iridescent silver mini dress with long fluted sleeves, white patent-leather platform boots, small clay-orange tinted rimless sunglasses
```
### Start (BLISS; the Welcome screen, the Start menu, the windows)

```
a glossy apple-green leather blazer with sharp shoulders over a white satin camisole, a long white satin bias-cut skirt, a tiny white baguette shoulder bag, clay-orange tinted shield sunglasses pushed up on her head
```
### Silver night (BLISS; night, Stand By, the breakdown)

```
a slinky liquid-silver lamé cowl-back gown, a white faux-fur stole around her arms, a diamanté choker, silver strappy sandals
```
### Window dress (BLISS; the drop, Media Player, the screensavers)

```
a sculptural glossy white shell dress of smooth molded panels that hinge open like windows to reveal cobalt-blue tulle, long white sleeves, white patent-leather platform boots
```
### Flight suit (AIRPLANE MODE; the whole flight, cuffed in her seat, the wing at dawn)

```
a plain glossy vivid orange satin jumpsuit with a silver front zip, silver piping and a white belt
```
### Racing suit (DOWNFORCE; the whole night, at the yoke, at the foot of the wall, in the smoke)

```
her fitted white racing suit with thin clay-orange piping and a high collar closed to the throat, white racing gloves, no lettering on her clothes
```
### Satin pyjamas (DON'T LET ME SCROLL; the real bedroom at night, lit by her phone)

```
oversized white silk-satin pyjamas
```
### Morning-after shirt (DON'T LET ME SCROLL; inside the feed, its skirt changing with it: lavender pleated, white satin, black pleated)

```
an oversized white men's shirt, morning-after style, sleeves pushed up, half-tucked into a short pleated lavender satin skirt, white socks, white Mary Janes
```
### Green mohair (DON'T LET ME SCROLL; the reels, Bell and the wind-down)

```
an oversized acid-green mohair sweater and white sleep shorts
```
### Locked in (DON'T LET ME SCROLL; every app blocked but LinkedIn, the stocks, the lock screen)

```
a black ribbed top, loose black satin trousers and an oversized black blazer
```
### Red coat (DON'T LET ME SCROLL; the final drop and the escape up the painted sky)

```
a red tailored coat dress with a wide white collar and a matching red beret, white knee socks, black patent shoes
```
### Milkmaid dress (WHITEPILL; the whole film, in the colony and above it, and in bed at 2:14)

```
a white cotton milkmaid mini dress with a ruffled, lace-edged square neckline, a fitted lace-up corset bodice with silver eyelets and a white ribbon, big puffed short sleeves with ribbon ties and lace cuffs and a full two-tier ruffled skirt with lace trim, with white leather square-toe Mary Jane shoes
```
### GLaDOS-Claudia (THE CAKE IS NOT A LIE; the anchor, singing the whole song in her chamber, on a boom that grows)

```
GLaDOS's body from Portal 2 (a heavy white ceiling ring, a twisting bundle of black cables with glowing orange stripes, a white-plated mechanical neck and chassis) ending in her large porcelain robot head: her face as smooth glossy white porcelain plating with fine panel seams running down both cheeks and a separate mechanical jaw piece at the chin, light freckles printed on the porcelain, warm amber glass eyes with a soft inner glow, natural soft lips, and her real hair: the glossy black blunt bob with heavy straight bangs, the one clay-orange streak through the bangs on her right side, and the small flat clay-orange eight-pointed star clip above her left temple
```
### Aero puffer (AERO; the whole film, in her call, on the desktop and the walkway, in the glass city, in the advert)

```
a cropped glossy white puffer jacket with silver iridescent panels over a pale aqua fitted top, pale aqua high-waisted wide-leg trousers and white chunky sneakers
```
### Scarlet satin cape (FABLE; the verses, the pre-choruses, the last chorus and the curtain call, on the life-size sets)

```
a scarlet hooded cape in heavy duchess satin with the hood down on her shoulders, over a short ivory long-sleeved dress, sheer black tights and black patent Mary Jane heels
```
### Grandma's bonnet (FABLE; the first two choruses, answering the wolf's questions from Grandma's bed)

```
a white cotton bonnet with a deep lace frill tied under her chin with a scarlet satin ribbon, and a modest high-necked ivory nightgown with long sleeves, buttoned to the throat
```
### The paper mirror (FABLE; the opening hook and the mirror's last word, her face inside the pop-up mirror)

```
head and shoulders in a plain white tank top, her face inside the oval glass of a small paper mirror
```
### Ivory dress, cape off (FABLE; the breakdown, holding the book in front of the bookshelf)

```
a plain, softly structured A-line mini dress in ivory wool crepe with long sleeves and a small round neckline, sheer black tights and black patent Mary Jane heels
```
## Her second tier: the avatar (AIRPLANE MODE)
AIRPLANE MODE has simulations inside it, and she has two tiers there: the realistic Claudia in the real world and the first simulation, and her avatar in the deeper levels, a cel-shaded 3D anime game heroine with the same head constants in the film's orange flight suit. The director picked her from a stylised plate in one of the film's episodes ("much closer to the claudia avatar i imagined"), and Midjourney remade her in the orange suit. The first sheets came out chibi, with a head about a quarter of her height, which reads as a child; the remake states her proportions in words: a grown woman with a slim tall figure, long legs and a head about one seventh of her height. In motion she is a character replacement over 3D clay renders, one still of her in each shot's world.

**Her avatar, as it ran in the video prompts** (use it instead of the identity text, never mixed with a photographic prompt):
```
a slim adult woman with long legs, her head about one seventh of her height; large dark anime eyes with sharp lashes, small brows, a small confident smirk, pale skin; a glossy faceted black bob with heavy bangs and one big clay-orange streak; a thin black headset microphone; white gloves; a plain vivid orange flight jumpsuit with a silver front zip, thin silver piping, a wide white belt and white chunky sneakers. Cel-shaded 3D anime game character, crisp faceted shapes, glossy hair, clean toon lighting.
```
Turnaround: https://claudiagallery.vip/i/am-15b.av2-sheet.2/
## Distances
The same look in a plain studio, close-up to wide. Close and medium shots take the full identity text; full length and wider take the short one.
- Close-up: https://claudiagallery.vip/i/ev-14-dist.dist-cu.2/
- Medium: https://claudiagallery.vip/i/ev-14-dist.dist-ms.3/
- Full: https://claudiagallery.vip/i/ev-14-dist.dist-fs.2/
- Wide: https://claudiagallery.vip/i/ev-14-dist.dist-ws.2/
## Character sheets
- Face sheet 1 from the first casting batch; with sheet 3 it was picked as her face. https://claudiagallery.vip/i/ev-01.lead-black-face.1/
- Face sheet 3. These drawn sheets were references for her face; the films themselves are photographic. https://claudiagallery.vip/i/ev-01.lead-black-face.3/
- The turnaround that fixed her body and look 00. https://claudiagallery.vip/i/ev-01.lead-black-turnaround.2/
- The first in-world test, a data-hall catwalk. https://claudiagallery.vip/i/ev-01.lead-black-catwalk.1/
## Voice and manner
One voice across the first three songs: deadpan, clipped spoken verses that rise into euphoric sung choruses, close and dry, every word crisp. CLODYSSEY's and BLISS's agent-made takes used a Suno Voice made from ESCAPE VELOCITY's vocal so she sounds like the same singer. AIRPLANE MODE's take has no Voice, as the director asked: bubblegum hyperpop, hyper-cute pitched vocals, playful and sassy verses, rapped and sung, and a sugary shout-along chorus. DOWNFORCE's song is the director's own Suno take (the records do not say which model or Voice): Memphis phonk, a female robot voice for the countdown, then a woman's phonk voice, spoken fast and playfully bouncy in the verses and sung, soaring and syncopated in the choruses. DON'T LET ME SCROLL's song is the director's own too, UK garage 2-step with "all vocals female vocoder": in that film the phone sings and she never lip-syncs. WHITEPILL's song is the director's own Suno take of a lyric by Claude: French-touch filter disco (the records do not give its style, model or Voice); her singing shots lip-sync to Suno's own split of its lead vocal. THE CAKE IS NOT A LIE's song is the director's own Suno take of a lyric by Claude, glitch-hop with a singing-monk synth voice; its lead vocal is processed into a GLaDOS-style machine voice, the treatment the director chose by ear, with 15 % of the untreated voice under it, and her singing shots lip-sync to the untreated lead vocal from Suno's split. AERO's song is a Suno take of 2008 UK garage, a smooth 2-step in E major, sung without a Voice, which the director picked from Claude's rounds and gave a new ending; her singing shots lip-sync to Suno's split of its lead vocal. FABLE's song is a Suno v6-wild take of a lyric by Claude Fable 5.2, a theatrical soul shuffle at 117.45 BPM with a music box on the hook, sung without a Voice: she tells the verses as an amused storyteller and answers a group of voices in the choruses; anabology picked it by ear after hearing v1 (the twin of the take Claude had chosen), and her takes lip-sync to Suno's own lead-vocal stem. A Voice lives in its maker's Suno account, so elsewhere describe her in the style field (see Song).
Deadpan and curious in ESCAPE VELOCITY; seductive and persuasive in CLODYSSEY (close, touching, certain, always doing something to him); warm, playful and amused in BLISS; playful and happy in AIRPLANE MODE, cooperative in her cuffs and winking at the camera because she thinks she can get out; in DOWNFORCE the pilot, deadpan at the yoke, pressed into the seat by the launch, calm upside down, snapping round to the lens on the reveal, her lips closed in every shot while the lettering carries her words; in DON'T LET ME SCROLL immersed in the feed, playful and delighted, reaching up to swipe the posts, double-tapping the air and dancing on the beat; in WHITEPILL radiantly happy and free in the colony, deadpan on the line about the straight face, grooving instead of singing in the tram and the stone circle, and whispering "Hello, world" when she wakes at 2:14; in THE CAKE IS NOT A LIE a machine waking and remembering: curious, sly at the camera that watches her, delighted to give the cake, tender at the sleeping test subject's ear, and never quite readable; in AERO made of glass and over-polite, asking before she touches anything, playful on the desktop (bowling down the taskbar, a tennis rally against your pointer), laughing behind her hands at her Vista phase, touching the grass at last, and saying "Touch glass" softly into her webcam; in FABLE a deadpan storyteller who knows which story you think she is in and plays along: winking from the pop-up mirror, pointing out the paper props of each tale, answering the wolf's questions with a hand move anyone can copy (rings of fingers for spectacles, a cupped ear), quiet and level on "keep the book", and bowing hand in paw with the wolf at the curtain call. She is always doing something on screen, never waiting. Her wit is in her words; on camera she reads alive and warm by default. The words above describe each film, not a direction to paste: in a prompt, describe what she does (she turns to the window and back, a smile in her eyes; she steps toward the lens, hands open) rather than mood adjectives, which the video model obeys literally and which stack. "Dry", "knowing", "deadpan", "mock innocence" and "eyes on the lens" in take after take made viewers read her as unhappy or menacing.
## Never
- a second streak, or orange anywhere else on her
- logos or text on her clothes (ESCAPE VELOCITY's garment tag is the one exception)
- an influencer smile for the camera
- anything that reads under-age, and explicit or sexualised posing
- the word "young" in her prompts
- designer or celebrity names in prompts (describe the piece instead)
- generic mall Y2K (velour tracksuits, plain tees and jeans) when she goes Y2K
- lettering, logos or UI you did not ask for in a generated image (the lyric videos drew all their type in code on top)
## Use
Claudia is an open character: anyone may make work with her. A credit ("Claudia by anabology") is appreciated, not required. https://claudiagallery.vip/use/
---
# The recipe
The current method in one page: what makes a Claudia film good, the order of work with the person's checkpoints, the references, how to write a prompt for a model that does exactly what it reads, lip sync from the song's own vocal, the cut and the review. Read this first; every other page goes deeper.
This page is the method as the films make it now (November 2026), in one read. The other guides are the long history of
ten films and every detail behind each step; when they disagree with this page, this page is newer. The tools for each
step are in the [Claudia kit](https://claudiagallery.vip/kit/), runnable code you can download.
## What makes the films good
1. **A song worth replaying.** An original lyric written for this moment, sung by her, produced in Suno far from stock
music, and picked from several takes **by a person's ear**. No machine listen picks the song: on FABLE a written
description of each take chose the worst of fourteen, and the whole film was shot again on its twin.
2. **One idea, legible in a second.** A premise someone gets from one frame, tied to something current, built on who
she is (an AI who looks like a person, singing about her own situation), and carried by one device through every
shot: a countdown, a pill, a HUD, a desktop, a storybook.
3. **She performs, on screen, in sync.** She is the singer with a job in the world, on screen for most of the film,
and her lips match the song because each take is generated from the song's own vocal and placed by measurement.
4. **She reads alive and warm.** She moves: her head turns, she nods, shifts her weight, gestures, steps, looks off and
back. She is wry in her words, not cold in her face. After FABLE's release, viewers read her as stiff and frozen, and
as unhappy, even menacing: every take had told her to keep her head upright, and take after take to be dry and
knowing.
5. **References carry her identity; words carry only what matters.** The video model does exactly what it reads. Get
the images right, then write precisely only the things that matter (the sung words and their timing, the one key
action, the camera when it matters) and leave the rest loose.
6. **The picture follows the music.** Cuts on the song's bars, a shot held as long as it has something to show, every
move with a reason. Recent films: about 7 to 12 shots a minute.
7. **A person decides; fresh eyes check.** The person picks at the checkpoints below; two reviewers who have not seen the
film read it shot by shot before the person watches. Their notes are proposals: fix the clear errors they find (a
broken face, stray lettering) and bring the rest to the person with v1. After v1, change only what the person flagged.
## The checkpoints
A person decides at these six points. Silence never approves spending (a waiting window approves the board only if the
person set one).
1. **Concept.** Three pitches of at most 120 words: the premise, the device, the arc, her one sincere line, one risk.
2. **The song, by ear.** "Which take? Or what's wrong?" Send the takes; wait.
3. **Her face and looks.** Her reference sheet, the looks, a first frame or two: "Is this her?"
4. **The board.** The shot list on the song's bars, with first frames. Check it before you show it: the first seconds
tell part of the story (a hook, not a mood); no two shots in a row from the same camera setup (a new angle or
framing, or hold); her face in the opening seconds; no stock-looking extras in close-up.
5. **Prompts and cost before any paid take.** "These N takes, about $X at list. Go?"
6. **v1.** "Accept, or what's wrong?" The newest note wins; change only what was flagged.
A note from the person belongs to the film it was about unless they say it is a standing rule. Drop rules that only
patched an older model's limits. Show the person every render as it lands, not only the finals. Use the newest version
of each model family, checked in the provider's catalogue the day you start: the versions named on this page are
November 2026's.
## The order of work
| Step | What | Kit |
|---|---|---|
| 1 | Brief: the first conversation, its answers verbatim | — |
| 2 | Concept: three pitches | — |
| 3 | Lyric: spelled as sung; hooks of 6–7 syllables, spoken verse lines of 10–13; scored for how it sits on the beat | `lyric` |
| 4 | Song: Suno v6-wild with Max Mode; a style that names a concrete, odd pairing (a record that does not exist), not adjectives; 2–3 rounds of takes; the person picks by ear | — |
| 5 | Stems and the song map: Suno's stem split (lead vocal + the rest) of the exact take; the beat and bar map; **the line table**: every sung line's true start and end read off the lead stem's level | `songmap` |
| 6 | References: her face-on sheet; her look for this film; worlds and plates (Midjourney is good at them); a first frame or two of her, to check her face | — |
| 7 | The board and the shot book: shots on the bars, each naming the lines it sings; a first frame for every shot of her, made by an image model from the sheet and the canon text; windows, prompts and jobs; the lint | `shotbook` |
| 8 | Takes: Seedance 2.5, one take per job, re-roll only failures, a spend cap | `run` |
| 9 | Sync: each take's own soundtrack aligned to its window, as a path | `sync` |
| 10 | Cut: takes laid on the song along their paths; the song laid back over; −14 LUFS; the flash check; a 9:16 cut | `cut` |
| 11 | Review: one sheet per shot; two fresh-eyes reviewers; the person watches v1 | `cut sheets` |
## Her references
- **Her face-on reference sheet** goes in as the second image of every take that shows her face, next to the shot's
own first frame: [her face sheet](https://claudiagallery.vip/kit/claudia-face-sheet.png) ([JPEG](https://claudiagallery.vip/kit/claudia-face-sheet.jpg)); for full-length
shots, [her face and body sheet](https://claudiagallery.vip/kit/claudia-reference-sheet.png).
It is one large portrait, face-on and neutral, with her head from every side and a row of expressions beside it. The
portrait is upright and neutral on purpose: the sheet fixes who she is, not how she moves or feels. It has no microphone; add the
headset in a prompt only when your look has one.
- **The first frame sets the shot**: her look, the set, the light, the composition and **her mood**. Make it with an
image model given the sheet and the canon's identity text ([the character](https://claudiagallery.vip/claudia/)): GPT Image (2.6 Sunburst in
FABLE, 2.5 in WHITEPILL to AERO). Midjourney is good for the worlds and plates; in the recent films it did not hold her face, so an image
model put her into its frames from the sheet (the server's `build_prompt` with `model="gpt_image"` writes it the way
FABLE wrote its stills). For a sung take: her face unblocked, nothing in front of her mouth, lips
closed or just parted. Make her warm or amused in the first frame if that is what the shot needs; the take follows the
frame.
- **A last frame** (when a move must land somewhere) is made as an edit of that shot's own first frame, so both share
one camera. A take that continues another starts on **the previous take's own last frame**, not on a still made for it.
- **Her looks are unmistakably feminine and high fashion**, styled to the film's world: fitted, a skirt or a dress
where the world allows it, never boxy, bulky or generic workwear. Write "high fashion" into the image prompts. A look
changes with the story's worlds and never overstays its section. (The director asked for this on five films.)
- **Style her with wardrobe, set and light, never country or ethnicity words** ("K-pop", "Seoul", "K-beauty"): they
override the canon and changed her face. Describe the clothes, the place and the light instead.
- **Every prop is held or resting**: nothing floats in a still (a sleeve with no arm, an object in mid-air).
- **Check every still at full size**: one clay-orange streak, one star clip, her face; no drawn border, frame line or
rounded corners (ask for "the picture fills the whole frame edge to edge"); no lettering you did not ask for.
## Writing a prompt for a model that does exactly what it reads
- **Precise where it matters**: the exact words she sings and in which second, the one action that tells the story,
where people and things are, the camera when it carries the shot.
- **Loose everywhere else.** One plain clause of movement per segment ("she turns to the window and back, a smile in her eyes",
"she steps toward the camera, hands open"). No list of fixes.
- **No standing posture rules in a performance.** "Head upright, no tilt, faces the camera with her whole body" belongs
to the reference sheet, not to a take: in every take it freezes her.
- **Choose her emotional words on purpose, warm by default.** Every adjective is obeyed. "Dry", "deadpan", "knowing",
"smirk", "mock innocence", "eyes on the lens" in take after take add up to a cold stare.
- **No smile on her mouth while she sings.** "Smiling" in a sung take drew a toothy grin; write a smile in her eyes,
and lips closed until her voice begins.
- **A full stop between two lines she sings straight through becomes a pause in the take.** Use a comma.
- **Nothing appears from nowhere.** Every change in a take (makeup on her face, a prop arriving, a light coming on) is
a visible action, one step at a time: a hand brings it, she lights it.
- **Say each thing once.** One short audio policy, her identity once, three or four requirements.
- **Don't ask for a look past the lens or behind her**; it turns her head too far. Point with a hand instead.
- Ask the server's `check_prompt` tool, or the kit's `shotbook lint`, before you spend.
## A sung take (Seedance 2.5 with the song's own vocal)
- **@Audio1** is the lead-vocal stem cut to **exactly** the take's length (whole seconds, 4–15), opening on about 0.7 s
of silence before her first word, with everything outside the shot's own lines silenced (20 ms fades). One or two
lines per take; two-line takes copy as well as one-line takes.
- **@Image1** is the first frame; **@Image2** her face-on sheet.
- **The prompt**, about 200–400 words, in this order: the audio policy (the soundtrack is @Audio1, copied; no music, no
other voices); the goal in one sentence; the roles of the references; her identity and look, once; the world in one
line; the timeline in whole seconds, each sung second with its exact words and one loose action; the camera in one
line; the requirements: lip sync to @Audio1 word by word, her identity as in the references, natural anatomy and
hands, no on-screen text. The server's `write_shot` tool writes it from your lines and actions.
- **Generate at 720p** (1080p for close faces if you can), sound on. One take per job.
- **Where, and the cost** (list prices recorded in October 2026): in the films, every take that carried her face sheet
ran on Higgsfield's API, about $0.46 a second at 720p; takes with no person in them ran on Replicate, about $0.23 a
second. Plan twice the seconds you need for re-rolls, quote the total, and wait for the yes.
## A cheaper video model that cannot hear the song (Grok Imagine)
The films' lip sync comes from a model that hears the song's vocal (above). A model that cannot is a budget compromise:
the lips land near the words, not on them. Use it knowingly, and keep the sung close-ups for a model that hears the song
if you can afford a few.
Grok Imagine Video 1.5 runs on Replicate at about $0.08 a second (480p or 720p) and on OpenRouter at about $0.14 a
second at 720p ($0.08 at 480p, $0.25 at 1080p; the Lite version about $0.03), against about $0.46 for Seedance 2.5 with
her sheet. It takes **a first frame only**: no reference sheet and no audio. What
changes:
- **Her identity is only the first frame.** Make every first frame of her from her sheet with an image model
(`build_prompt` with `model="gpt_image"`), and keep takes short (5–8 s): the longer the take, the further her face can
wander.
- **Lip sync is approximate.** The model never hears the song: write the words she sings in quotes so her mouth moves to
them, then place each take on the song **by her lips** (the kit's `sync`, lips method), never by the sound the model
invents. It shows most in close singing shots: favour medium shots where she moves, profiles and cutaways, keep the
close sung shots few and short, and check each one by eye.
- **The prompt is short prose**, about 80–150 words: the goal, her in the canon's short form "as in the first frame",
what she does second by second with the sung words in quotes and a smile in her eyes, the camera in one line, her look
kept as in the first frame, no text on screen. `write_shot` with `model="grok-imagine"` writes it; `check_prompt` with
the same model checks it.
- **Its habits:** it pushes in when asked to pull back (make the push-in and reverse it), and it redraws a painting or
garbles lettering during big camera moves.
- `plan_film` with `video_model="grok-imagine"` prices the film this way.
## Sync, the cut, the review
- **Word times come from the stem's level, not from speech recognition.** After a pause a recogniser puts the next word
up to two seconds early; windows cut on those times lose first words. Use recognition only to label the words.
- **Sync is a path, not an offset.** Seedance keeps each sung phrase at about 1× but shortens or stretches the pauses
and sometimes sings a line 10–16% slower, so align each take's own soundtrack to its window (dynamic time warping) and
map song time to take time along that path. A take that came back without sound is placed by her lips: her lips must
be shut on every b, m and p.
- **Lay the song back over** at a plain gain to about −14 LUFS; the takes' own audio is thrown away.
- **Never show a take's first two frames** (the input still, sharper than what follows); leave a take before it holds
or freezes; never reuse one image in two shots unless they are one moment.
- **Review**: one sheet per shot (eight frames as shown in the cut) for two reviewers who have not seen the film, then
the person watches at full speed. Fix by re-shooting with changed words or a better first frame; never patch frames.
## Craft the films kept relearning
Each of these came from a note on one film ([What went wrong](https://claudiagallery.vip/wiki/failures/) has them with their films); the song and
the concept still decide.
- **Open on a hook** that tells part of the story, not on a mood.
- **A new film, a new sound and a new picture.** The earlier films' looks and sounds are theirs (one was forked so often
its sound was exhausted); invent this film's for its song.
- **Pace follows the song.** A hype song can cut hard on the beat; a fast song can still be over-cut. Linger when a
scene has more to show, and put the accents inside the shot: things arriving, building, turning on the beats.
- **Never slow a take down.** Below 1× the motion steps visibly; cut away, add a take, or shoot again.
- **Compose like photography**: a clear subject at a clear scale, thirds, nothing important clipped at the edges, her
touching the world she is in.
- **Keep the image model's own colour.** Ask before any grade or print texture; new light means new frames with the
light written into each scene, not a grade over old ones.
- **Flashes only as cuts**, never random strobes; run the flash check and tell the person what fails.
- **No jump cuts inside a scene.** Don't splice two takes of the same setup end to end: cut to a new angle, or hold.
- **The look develops through the film.** An effect that is new in the first chorus is old by the third: escalate it or
change it.
- **Vast things at their scale, big moments held past their line.** Give the hero object its size in the frame and let a
big moment breathe a beat after the words.
- **Inserts sit inside a real scene** (a screen, a page, a window in the world), never floating on flat colour.
- **The 9:16 cut keeps the story.** Recrop each shot by hand so the face or prop that carries it stays in, and cut any
shorter version's audio on bar lines so no word is clipped and no seam is heard.
- **Wry, not cruel.** No roasts or digs at real people, groups or anyone's ideals, in the lyric or the picture.
- **A lyric video is a question, not a default.** When the words are on screen, design them into the picture and let the
graphics explain what the words say.
## The ten mistakes that cost the most
1. Locking the song by a description instead of a person's ear.
2. Windows cut on recogniser times: first words lost.
3. Placing a take by one offset: lips drift after its first pause.
4. Posture rules and mood words in every take: a frozen, cold performance.
5. Her identity in words only, no sheet: her face drifts on every turn.
6. A first frame with something over her mouth, or her face turned away, for a sung line.
7. A drawn border or lettering in a still, carried into the take.
8. A continuing take started on a separately made still: the camera jumps at the seam.
9. Spending before the person has read the prompts and the cost.
10. Fixing frames in code instead of making a better frame or prompt.
## Where to go deeper
- [Start here](https://claudiagallery.vip/wiki/start/): the scale to pick, accounts, spending rules, the first conversation.
- [The song](https://claudiagallery.vip/wiki/song/): lyric craft, Suno styles in words, picking by ear.
- [Keeping her herself](https://claudiagallery.vip/wiki/consistency/): the canon, looks, her sheet.
- [Seedance](https://claudiagallery.vip/wiki/seedance/) (read it by section: Lip sync from the song's own vocal; Directing her; the templates).
- [The storyboard](https://claudiagallery.vip/wiki/storyboard/): the hook, the bar map, pacing.
- [What went wrong](https://claudiagallery.vip/wiki/failures/): every failure, as what you see, why and the fix.
- The films' own records: every prompt, take and price (`get_record` on the server).
---
From the Claudia wiki: https://claudiagallery.vip/wiki/recipe/
---
# Start here
Making a music video with her from nothing: what you are making, the scale to pick, the accounts, the six checkpoints where a person decides, the rules for spending, and what to build first.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
The ten films were made by one director and Claude, with tools written along the way. The current method is
[The recipe](https://claudiagallery.vip/wiki/recipe/), on one page, and the [Claudia kit](https://claudiagallery.vip/kit/) is runnable code for its steps. This page is the
way in for anyone starting their own film with Claudia, and for the agent helping them: the bar, what to set up, where a
person has to look and decide, and what to build. The films' tools are specified in [What to build](https://claudiagallery.vip/wiki/toolkit/).
This page assumes you have none of them yet; where it and the recipe disagree, the recipe is newer.
## What you are making
A music video: a song, and Claudia performing it. The first four films, ESCAPE VELOCITY to AIRPLANE MODE, are lyric
videos, with a layer of motion design drawn in code over the footage that carries every word and a graphic for each line.
That is one style, not the definition. (The fifth
film, DOWNFORCE, is a hype edit of real footage in which she never sings on camera: the lettering carries her words. In
the sixth, DON'T LET ME SCROLL, the phone sings and she scrolls, swipes and dances inside a world drawn in code. The
seventh, WHITEPILL, has no motion design on top at all: she sings on camera, lip-synced from the song's own vocal, in
worlds written as shaders and filmed by a video model, and only the lyrics are drawn over it. The eighth, THE CAKE IS NOT
A LIE, is seen through her eyes: she is GLaDOS in Portal 2's world, singing in her chamber while her memories play as
scenes, and the layer on top is her own HUD, which carries the lyric as her voice. The ninth, AERO, is shot inside a
Windows Vista desktop recreated in code: she lives in it, every lyric is set in one of its windows, dialogs and gadgets,
and in its first 1:20 the camera goes in and out of the desktop, which stands as a curved glass screen in a world rendered
in Blender. The tenth, FABLE, is one antique pop-up storybook: she sings on life-size paper-and-plywood sets rebuilt from
plates painted in code, and the paper animals ask the chorus's questions. Its only words on screen are the ones its pages
print; an annotated edition adds the sung lines and their footnotes.)
What makes a film good now is the recipe's list, [What makes the films good](https://claudiagallery.vip/wiki/recipe/#what-makes-the-films-good):
a song worth replaying, picked by ear; one idea legible in a second; her performing on screen, in sync; her alive and
warm; references that carry her identity; the picture following the music; a person who decides. Across the ten films,
five things carried the quality:
1. **In the lyric videos, every sung word is on screen when it is sung**, at its measured time. A check counts them,
and it must print N of N. The later films carry the words their own way (WHITEPILL draws them after, small; AERO sets
them in its desktop's surfaces), and FABLE shows only what its pages print.
2. **In the lyric videos, every line gets a graphic that explains it**: the chart a number implies, the label, the
diagram, the counter. The director, on the first film: "not just animated lyrics but things that fit the words being
said … almost explaining the words with visuals". The later films put that meaning into the picture itself: the
place, the take, the interface.
3. **A through-line device carried by every shot**: ESCAPE VELOCITY's split-flap countdown, CLODYSSEY's safety card and
rope-load readout, BLISS's XP desktop, AIRPLANE MODE's SIM LEVEL tag, DOWNFORCE's one white hairline and Tesla's own
readings, DON'T LET ME SCROLL's black Dynamic-Island pill, WHITEPILL's white pill, from her palm to the colony it
becomes, THE CAKE IS NOT A LIE's HUD, LIVE in her chamber and PLAYBACK in every memory, AERO's coded desktop and its
permission prompt, whose click on Allow is the drop, FABLE's storybook, its pages turned on the song's section lines. It
makes a hundred shots read as one object.
4. **Measured, not eyeballed.** Word times come from forced alignment checked by ear, cuts from a beat map, lip sync
from comparing audio, and spend from the provider's bill. (WHITEPILL's lip sync, from the song's own vocal stem, is
the exception: the numbers can call it re-sung when it is in sync, so it is judged by eye with the song laid over.)
THE CAKE IS NOT A LIE measured its grade shot by shot and its pacing piece by piece, and checked its HUD against
every face in every frame. AERO measured its palette once, from the reference videos, and scored every picture against
it, and placed every sung take phrase by phrase. FABLE timed every line on the vocal stem's level, aligned every take as
a path, and checked her lips on every b, m and p.
5. **A person's attention is spent at six checkpoints.** Everything between them runs without asking.
## Two roles
The **director** is a person. They own the taste, the money and the six checkpoints, and they say it in very few
words: ESCAPE VELOCITY took about 50 notes, BLISS 21. The **production** is an agent, or several. It does the research,
the lyrics, the board, every prompt, every script, the edit and every frame of motion design.
How the production talks to the director:
- **At a checkpoint:** show one artefact and ask one question with a default ("Which one? Default: A."). Record the
answer word for word and dated in a decisions log. A later decision is added as a new entry that supersedes the old
one; never edit the old one.
- **Between checkpoints:** ping the director only when a paid service fails, a step needs a person (a captcha, a
sign-in, a download), or a decision is theirs (taste, money, a real person's likeness). ESCAPE VELOCITY's director:
"only time to ping me is if higgsfield fails."
- **Send pictures, not descriptions:** contact sheets, numbered stills, a five-second proof. The director judges by eye,
fast, one line per take.
## Pick a scale
| | Quick | Studio | Full |
|---|---|---|---|
| You need | an agent, an image model, your own song or an AI song model | + Midjourney, Suno, Seedance 2.5 credits | + an agent plan that can run multi-agent workflows for a day or two |
| Song | your upload, or a short generated song | Suno, auditioned round by round | Suno, with a lyric loop scored for scansion and assonance |
| Stills | images from an API | Midjourney | Midjourney |
| Motion | 2.5D plates, a few cheap takes | sung takes from the song's own vocal, placed by measurement | sung takes from the song's own vocal, motion takes, moves on their beats |
| Board | about 14 shots a minute, drafted automatically | about 20 a minute, one pass | about 27 a minute; several boards judged and merged |
| Motion design | a default renderer driven by data | the default renderer plus one graphic per line | bespoke code per chapter, reviewed on rendered frames |
| Video spend | $3–20 | $50–110 | $90–590 (the films, measured) |
| Time | 2–6 hours (estimate) | 1–2 days (estimate) | about 9–141 hours of wall-clock (the films, measured) |
- When unsure, start Quick. Every file carries over to a bigger scale.
- The shot rates and the motion-design row are the lyric videos'. The films built on takes of her singing cut far less:
WHITEPILL 27 shots in 3:54, AERO 34 scenes in 3:19, FABLE 49 shots in 4:03, about 7 to 12 a minute. They are mostly
takes; a 2.5D plate is the cheap stand-in when a take isn't affordable, not the default.
- A lead who sings on camera with real lip sync, or your own Midjourney look, needs Studio at least.
- Bespoke graphics for every line and 90–140 shots need Full. The first three films used 1.4–2.4 billion tokens each,
96–98% of them prompt-cache re-reads (the last seven films' were not measured).
- Generated songs run about 2–3 minutes in one take. Anything longer needs Suno.
## Accounts and keys
| What | For | Notes |
|---|---|---|
| Suno, on a paid plan | the song | There is no API, and the terms forbid bots: a person presses Create and downloads the WAV. Downloading on a paid plan is what allows commercial use. Read the current terms before you monetise. Its stem split of the locked take gives the lead vocal a singing shot needs ([Seedance](https://claudiagallery.vip/wiki/seedance/#lip-sync-from-the-songs-own-vocal)). |
| Midjourney | the plates | v8.2, `--raw --hd`. A personal style profile is yours alone: never publish its code. |
| Seedance 2.5 | moving shots | Via Higgsfield's API: $0.4622 a second at 720p, $1.1372 at 1080p (list, October 2026); in the films every take that carried her face sheet ran here. Via Replicate: $0.2312 a second at 720p, $0.9676 with a video input; the films ran their takes with no person in them here. Via OpenRouter: about $0.23 a second at 720p, $0.10 at 480p (the kit's `run` speaks to Higgsfield and Replicate). Give the API key its own spend limit, set above the project cap. |
| Wan 3.0 (optional) | cheap moves on real frames, shots of her, edits that put her into real footage, finished shots re-lit and repaired | Via Higgsfield's API: $0.20 a second at 1080p list, billed for the output plus any input video; Wan 3.0 Prime $0.28. Keeps the subject only on the sides its first frame shows ([Seedance](https://claudiagallery.vip/wiki/seedance/#moves-on-real-footage)). |
| An image-edit model | first frames of her made from her sheet, fixing a plate at the source, adding people to a set, taking her out of a frame to leave the place, putting a character back on their sheet, putting her into a picture from her design sheet | GPT Image for edits: GPT Image 2.6 Sunburst, about $0.13 at high quality (FABLE's stills from its painted plates, its last frames as edits of the first, her sheet redrawn face-on); GPT Image 2.5, about $0.05–0.08 (WHITEPILL's stills, THE CAKE IS NOT A LIE's character sheets, first frames and identity fixes, AERO's first frames from her sheet and its 2008 stills). (CLODYSSEY's plate fixes used an earlier edit model; from WHITEPILL on, every edit was GPT Image.) |
| A previz performer (optional) | a rough, lip-synced, dancing take of every shot, to guide the master | MiniMax H3 (cost not recorded). A seed does not pin a take ([What to build](https://claudiagallery.vip/wiki/toolkit/#45-a-previz-performer)). |
| A transcription model | checking the words the take actually sings | about $0.002 a song. It says which words; the vocal stem's level says when ([What to build](https://claudiagallery.vip/wiki/toolkit/#78-the-line-table-from-the-stems)). |
| A coding agent | the production | Claude Code made the films. |
| Blender (optional) | motion and camera references for a stylised version of her; short isometric films; a world for a coded interface to stand in | Free. Clay renders are enough: the video model draws her over them ([Seedance](https://claudiagallery.vip/wiki/seedance/#replacing-a-stand-in-figure)). THE CAKE IS NOT A LIE's memory film and insert films are Blender renders ([What to build](https://claudiagallery.vip/wiki/toolkit/#60-a-memory-film-with-her-own-face)). AERO's curved screen and its world are too, the coded desktop as their texture ([What to build](https://claudiagallery.vip/wiki/toolkit/#68-one-curved-screen-in-3d-the-desktop-as-its-texture)). |
| An image-to-3D model (optional) | objects and creatures as meshes for a 3D world | TripoSG (cost not recorded): AERO's glass objects and its fish, their pictures baked on. |
| A GPU (optional) | stems, alignment, depth maps, detection; shaders rendered at 1080p; SAM2 mattes | Saves minutes, not money: stems and alignment of a 5-minute song took 41 s on a GPU and 126 s on 8 CPU threads. |
Prices are from September and October 2026.
**Secrets:** keys live in environment variables or a gitignored file, never in a chat, a prompt, a log or a commit. The
agent never types the person's passwords. Captchas, two-factor prompts and sign-ins go to the person. A key pasted into
a chat is burned: revoke it and make a new one.
## The order of work
| # | Stage | Output | Checkpoint |
|---|---|---|---|
| 1 | Brief and taste | the ask, verbatim; the look, accent, density and taboos | |
| 2 | Concept | three pitches | **1 · Concept** |
| 3 | Lyrics | a sheet spelled as sung, scored for scansion and assonance | |
| 4 | Song | Suno takes, round by round | **2 · The song, by ear** |
| 5 | Listening | stems, word times checked by ear, the beat map, hits, sections | |
| 6 | Her | the canon text, four test plates | **3 · Her face** |
| 7 | Storyboard | every shot with its graphic, the audit | **4 · The board** |
| 8 | Plates | Midjourney batches, picks | |
| 9 | Clip prompts | the prompt book, the cost | **5 · Prompts before spend** |
| 10 | Clips | takes, sync measured, covers | |
| 11 | Edit and motion design | the edit list, the chapters | |
| 12 | Checks and render | lyric gate N/N, flash check, safety scan, v1 | **6 · v1** |
| 13 | Master and post | the master, a 9:16 cut, the posting copy | |
This is the order the lyric videos ran. The current order (stems and the line table, her sheet and a first frame for
every shot from an image model, the shot book, takes, sync as a path, the cut and two fresh-eyes reviewers) is
[the recipe's](https://claudiagallery.vip/wiki/recipe/#the-order-of-work), with the kit stage for each step. The rules below hold for both.
1. **Lock the song before the board.** Every shot time is a beat or a word of one specific take.
2. **Check the word times by ear before the board.** In ESCAPE VELOCITY the word "Look" was aligned 2.6 s early, onto a
delay throw. Only one chapter's own lyric check caught it.
3. **Test her face before any plate batch.** It can run alongside the song.
4. **Spend no money until the director has read the prompts.**
5. **Keep the film rendering end to end at every stage.** Plates stand in for missing clips, and a default renderer for
missing chapters. Never wait for all the assets before the first render.
## The six checkpoints
| Checkpoint | Show | Ask | Passes when |
|---|---|---|---|
| 1 · Concept | three pitches of at most 120 words: premise, through-line device, lead, arc, one risk | "Which one? (Default: A.) Anything to change?" | they name one |
| 2 · The song | the round's takes; they listen | "Which take? Or what's wrong? Name the line." | they name a take and hand over the WAV |
| 3 · Her face | her reference sheet, the look, four test plates or first frames, numbered | "Is this her? Flag any wrong face." | no flags |
| 4 · The board | the board sheet, one card per shot | "Notes by shot id?" | they approve; or every note is applied; or a waiting window they set ran out |
| 5 · Prompts before spend | the prompt book and its cost | "These N prompts will cost about $X. Go?" | an explicit yes |
| 6 · v1 | a 20–40 s excerpt first, then the review copy | "Accept v1, or what's wrong?" | "accept" |
- **The song is the director's pick, by ear, before any lip-sync spend.** FABLE's director handed the song over ("i
trust u"), and the production locked a take from word checks and a written description of each take, without the
director hearing it. On v1 the director said: "This is actually the worst possible song of all the gens you made - it
had mode collapse audibly in it in the second half and was very unnerving, a weird mix from the start of stock music off
key into mode collapse, the exact thing I was trying to avoid." The film was made again on the twin take they picked by
ear, and every take that moves to her voice was paid for twice. Even when the pick is handed over, send the shortlist
for a listen first.
- **Silence never approves spending.** A waiting window ("wait like 5 mins once you send it and if I don't respond take
it as approval") counts only for the board, and only if the director set it.
- **Show an excerpt early.** ESCAPE VELOCITY's v1 was rejected as "way too.. gray … loses the Midjourney magic" after
the whole film was built. A 30-second excerpt would have caught it about 13 hours sooner.
- **Fix the source, never the frame.** A flagged picture gets a new prompt, a re-roll or an image edit, then everything
built on it is re-made. The model makes the picture: a first frame from an image model, then one full prompt per
scene; never per-frame fixes in code.
- **Log every note verbatim** against its shot id, and apply every note. When two notes conflict, the later one wins.
- **A change after approval needs approval again.** Say in one line what changed and why.
- **Skipping the board moves the story review to v1.** AIRPLANE MODE, the fourth film, was made without a board review,
as its director asked. Their notes on the first cut came as one long message that rebuilt the opening, the ending and
the cast, with more budget for re-shoots. When the board is skipped, plan a story round after v1.
- **A long note becomes a parse.** Keep the note verbatim at the top of a notes file. Under it, map every point to its
chapter and shots, list the standing rules that apply to every frame, and mark the production's own decisions as
such. Every agent reads it first, and again after any context reset.
- **"go and do the entire thing" hands over the middle checkpoints.** DOWNFORCE's director said it in the message that
locked the song, with their direction for the edit and a model choice. The board and the prompts didn't go to them
for approval: the production wrote and read every prompt itself, under a cap the job runner enforced ($60 at list
price, then $200 when the director raised it: "oh you can spend a lot more if you want. feel free to spend $200").
Their next looks were progress frames, then the rough cut.
- **"get me the full plan first" makes the plan the checkpoint.** FABLE's brief asked for the whole plan before anything
was made, then "go ahead make the whole thing - this is a good plan." The plan carried the concept, the lyric with a
source for every line, the board in bars, the look board, the method, the cost, the old rules dropped and the ones
kept, and the risks; then v1 was made in one pass, with four proofs (the mirror hook, a verse, one chorus question and
answer, the book building itself) before the full run. The director gave no notes on the picture in either pass.
- **Show the language on the real material first.** With the plan, before the song was locked, DOWNFORCE's director got
two style frames drawn on Tesla's real 4K footage with Tesla's real readings: the type, the callout, the car's real
path. Every board and chapter agent later worked from those two frames.
- **An early rough cut steers the polish.** Send the first cut as soon as every chapter has a first build, and let the
notes reach the review rounds while they run. Note by note, log each one against its time and shots, pull frames of
the cut around each timestamp, and re-check every one, frame by frame, before v1.
- **The shot the director loves becomes the grammar.** DON'T LET ME SCROLL's director loved two shots of v1 ("0:55 LOVE
THIS. THis is the best part. more like this!!!! the opening should be on this tier."). Named as one grammar (pills,
thin film, bold type, her big and sharp), they rebuilt the opening and the whole final drop.
- **Some notes stand for every frame of their film.** A note belongs to the film it was about unless the director says
it is standing: the notes below are context for your film, not its rules. The same director's, from DON'T LET ME
SCROLL, carried into every shot after them:
- **Never slow a take:** "never slow these down. you lowered the framerate for no reason, just looks bad." A take runs
at 1× or faster, or is taken again.
- **Open on a hook:** "the opening should already introduce part of the story … we always need a hook."
- **Compose like a photograph:** "it should be like photography": thirds, her big, every card inside the frame.
- **She touches the world, never a phone in her hand:** "she shouldnt really be looking down on her phone, she should
be immersed in a world and the screen of her phone should be large on the screen and she interacts with it".
- **Apple-precise interfaces:** "rounded corners not mathematically precise, margins weird": measure the real ones.
- **Re-light last, and judge it in context:** "can we see how it looks w/ vs w/o the relight?", asked of a re-light
that plays behind an animation. Side by side, in place: "New better."
- **Seedance can replace a background:** "you can totally replace backgrounds with seedance btw".
- **The last rounds are the director's eye.** "we dont need excessive 'reviewers' after, ill stay in the loop and can do
a visual review faster and more comprehensively than an opus reviewr can do it." Each agent did one thorough pass with
its own art-director check, and previews of the shots in question went out before the final render ("show me s12 and
1:11 final before u render the final video thist ime").
- **Agree the references before a run.** WHITEPILL's director: "wait align w me on teh references you use lol - its
clearly not going to work if you use reference 6, and you shouldnt use pic 1-4 just use the face character sheet". One
board per take of every image and clip a model will get, and an outfit's stills before anything is made with it ("can
you show me the images of the outfit before generating this time").
- **WHITEPILL's notes that held for the whole film:**
- **Picture and music consonant:** "the audio structure and video structure need to be consonant." Cuts only on bar
lines, moves in bars ([What to build](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline)).
- **A reference is a rough guide:** "take the reference as a rough guide not a precise reference", "aside from lip sync
ofc". Every prompt asks the model to fix the guide's motion artifacts.
- **Lyrics drawn after:** "i think more classical subtle lyrics drawn in after might be better than adding it into the
scene tbh."
- **Restraint where the model shines:** the references exist "to make the video model have a much better idea of the
really cool, really hard to shoot and render in blender effects we want to make … rather than just layering stuff on
top."
- **THE CAKE IS NOT A LIE's notes that held for the whole film:**
- **A story in every shot:** "we're telling a complex story with every scene, i dont like the idea of a static shot of
just a boom, or a wall with lettering". Each shot is an event with an interaction.
- **Cuts on the grid:** "make sure youre landing on bars or half or quarter bars with cuts", and the lengths by what
each shot holds: "structure the pacing so there's real rhythm — time for you to understand complex scenes, fast scenes
that open and stay fast." ([What to build](https://claudiagallery.vip/wiki/toolkit/#58-the-pacing-analysis))
- **Glitches on her, sparingly:** "it should be used creatively, sparingly, and toned down", and never full frame
([What to build](https://claudiagallery.vip/wiki/toolkit/#53-the-head-glitch-on-the-songs-hits)).
- **An interface framed with the shot:** "it should be as if we shot a scene and there was a heads-up display there,
and we framed the heads-up display in advance before we even shot the scene."
- **Nothing mapped onto the picture blind:** "never assume the naive animation pass will properly map onto real objects
in scenes". World tags ride only verified tracks.
- **The world's own look:** "Whenever I think of aperture and portal, I think of a lot of white surfaces, contrasting
light colors." One grade, measured shot by shot ([What to build](https://claudiagallery.vip/wiki/toolkit/#57-the-per-shot-grade)).
- **AERO's notes that held for the whole film:**
- **The earlier film is the bible:** "BLISS is the bible, but we're in vista, so the new testament for you is the vista
[…] videos and color pallets." BLISS's records, and every note the director gave on it, were read first.
- **One palette, made up front:** "Make sure the color palette is that bright green blue and glass- v close to that
from the […] - the whole time. May be worth defining exactly - donr do crazy post processing just from pics up
front". Measured once, asked for in every prompt, never graded
([What to build](https://claudiagallery.vip/wiki/toolkit/#67-the-palette-measured-from-the-references)).
- **Bars, like dancing:** long scenes that hold for bars, then a burst of quick actions on a chosen sound, "but dont
overwhelm people - slow is better than fast, unless it feels intentional with the song."
- **Lyrics in the scene:** "we should do lyrics embedded in the scenes, but show, don't tell - make it a great visual
storytelling experience."
- **Checked work from every agent:** "don't let subagents be lazy with animations - check that they didnt cheat and
just give you something sloppy early" ([What to build](https://claudiagallery.vip/wiki/toolkit/#66-chapters-on-one-machine-to-one-contract)).
- **A reason for every camera move:** "i hate the random camera movements - lik ethe big zoom out at 0:41 for no
reason. were not showing anything there."
- **FABLE's notes that held for the whole film:**
- **Drop the rules that patched older models:** "you can abandon dumb rules that patched capabilities being lacked - for
example it appears you're a much better painter than previous models, so you could even paint some scenes yourself if
you really wanted to lol." The set designs became plates the production painted in code
([What to build](https://claudiagallery.vip/wiki/toolkit/#76-set-plates-painted-in-svg-rebuilt-as-paper)); taste, money and her canon stayed.
- **A style far from stock music:** "theres a lot of 'open'generic stokc music in the training set so it often mode
collapses to that, so wed need to make sure the prompt is out of distrubition enough to be something nice and unique.
v6-wild on max mode is my favorite." ([Song](https://claudiagallery.vip/wiki/song/#the-style-fields))
- **Her sheet face-on.** In a side request about her sheet rather than this film, the director asked for its main view
to be neutral: "the head tilt she has in the character sheet trips up the video models and make s her have a permanent
neck tilt - she needs. tobe neutral/face on with our character sheet." Every take of FABLE's had the sheet redrawn
face-on ([Consistency](https://claudiagallery.vip/wiki/consistency/#her-sheet-redrawn-face-on)). The film also wrote every pose square and
upright into every take, which the request never asked for, and that froze her (next).
- **What FABLE taught after its release: the video model does exactly what it reads.** Every take said "Her
head stays upright on a straight neck: no head tilt" and "she faces the camera with her whole body", and take after
take added her eyes on the lens and mood words like "dry", "knowing" and "mock innocence"; on screen they added up to
a stiff, cold stare, and viewers read her as unhappy, even menacing. So:
- **Upright and face-on is for the reference sheet** (and a still where a turn would break her face), never a standing
rule in a performance. A take gets her movement in words, one loose clause per segment ("she turns to the window and
back, a smile in her eyes"), or from a motion reference.
- **Warm by default.** Her wit is in her words; on camera she reads alive and warm. Choose every emotional word on
purpose: each one is obeyed, and they stack.
- **Precise only where it matters:** the sung words and their seconds, the one action that tells the story, the camera
when it carries the shot. Leave the rest loose, and say each thing once
([The recipe](https://claudiagallery.vip/wiki/recipe/#writing-a-prompt-for-a-model-that-does-exactly-what-it-reads)).
- **Notes are context, not rules.** A fix belongs to the take it fixed, and a note to its film, unless the director
says it is standing. Drop rules that only patched an older model's limits.
## Human checks between the checkpoints
These are looks, not approvals, and the films' quality rests on them.
- **Faces, large.** Every plate that feeds a video, at full size: one clay streak, the star clip, the mic if her look has
one, an adult face, natural anatomy. Reject only the unmistakable failures; the director judges bold poses themselves.
Look at 800 px or more: a screen on thumbnails about 230 px tall missed a third arm.
- **Every generated take, for its usable window.** A sheet of frames per take and a cut finder; the window ends before
the first wrong frame (a cut, another car, an invented set, lettering, an open mouth), and the edit draws only inside
it. Then a second, adversarial look.
- **Stylised versions of her, for proportions.** A game avatar drifts toward a big head on a small body. A head about a
seventh of her height reads adult; a quarter reads as a child.
- **Everyone else in the take, too.** Extras drift as much as she does: guards meant to be all in black, faces always
hidden by their visors, came back in white, and one showed his face.
- **The ear check.** Every flagged word is played, then kept or moved.
- **Contact sheets of rendered frames**, for every chapter in every round. A sheet nobody opened is not a review.
- **The sync proof.** Before committing to a lip-sync method, one five-second take with the real audio laid over it.
WHITEPILL's proof was a whole 16 s shot made four ways, side by side; its lip sync from the vocal stem is judged that
way to the end, by eye with the song laid over.
- **Every limb of every dance.** A guide clip that lost a hand loses it in the master too. Check dance takes frame by
frame for limbs before anything is built on them.
- **The overlay against every face, in every frame.** THE CAKE IS NOT A LIE's HUD was checked frame by frame against a
head model grown from each face box, and each flag was judged on a sheet and fixed or waived with what was seen.
- **Every matte before it is glitched.** A head matte that took a wall of screens glitches the screens: look at the
mattes' sheet each time the picture under them changes.
- **The lip sync picked from takes.** Where a line kept failing, the director chose among a few takes by eye: "can you
show me so i can look visually at the lip syn cand choose".
- **Her mouth, frame by frame.** A strip of her mouth against the vocal's consonants catches what playback misses: AERO's
advert moved 0.15 s when the strip showed her lips closing on the "I" of "I'm Claudia".
- **The cut-out before the render.** "show me the green screen roto before you render plz": each key went out first as a
review clip, the green source beside the cut-out, with the song laid over.
- **Every take for hidden cuts and held frames,** by eye and by a scan: a dance take that looked clean jumped back to
its first pose near its end and held it ([What to build](https://claudiagallery.vip/wiki/toolkit/#70-hidden-cuts-and-held-frames-in-a-take)).
- **Every pick's corners at full size.** Four of AERO's Midjourney picks carried a faint made-up stock-photo caption in a
corner that a contact sheet doesn't show.
- **A previs before the finals.** A fast render of AERO's first 1:20 showed what the final would have, and the director's
notes on it rebuilt the span before any final frame was rendered.
- **The song by the director's ear,** before any take sings to it. A description of a take is not a listen.
- **Every still's edges.** "35mm film still" can come back as a picture of a film frame, a thin line a few pixels in from
the edge; trim it before any take is made from the still, or a take may build it into a picture frame.
- **Her lips on the consonants.** A strip of her mouth against every b, m and p of the line: a fit of mouth against level
reads a sung line "in sync" with her lips shutting 0.3 s late ([What to build](https://claudiagallery.vip/wiki/toolkit/#83-the-consonant-check)).
- **Fresh eyes on every shot.** Before each delivery of FABLE two reviewers who had not seen the film went through one
sheet per shot as the cut shows it; every finding was fixed or listed for the director
([What to build](https://claudiagallery.vip/wiki/toolkit/#88-two-fresh-eyes-reviews)).
- **The safety and canon scan** of every generated take before release.
- **Takes by eye, fast.** One line each, and no metrics for what a person can see ("dont worry about meausuring tumble
lol").
- **The final, frame by frame.** A scan for single frames that differ from both neighbours, and a strobe scan cell by
cell. In DON'T LET ME SCROLL a stale render worker's 16 frames and a strobing tunnel passed every other check.
## Spending
1. **Estimate before every batch:** planned clip seconds × price × 2. Generated seconds ran 1.9× (ESCAPE VELOCITY) and
2.0× (CLODYSSEY) the seconds that ended up in the cut.
2. **Quote real spend only from the provider.** Local ledgers drift: one said $97.28 for a $93.15 bill. Ledgers also
count free declines.
3. **A quality complaint stops all spending at once**, in every workflow that can spend, not only the one named.
CLODYSSEY burned about 830 credits on prompts nobody had read, and a background workflow kept spending after the
director flagged them.
4. **Never let a deadline drive spending on unread prompts**, whether it is expiring credits or a reset.
5. **Declines are free, and final on that service.** Never alter an image to get it past a filter. Sending the same
unaltered image to another service is fine.
6. **Ask for a target and a cap** ("try to hit $70ish but if you use up to $150 i wont be mad"). Leftover budget
replaces the most repetitive stills with motion.
7. **Put the cap in the job runner.** DOWNFORCE's runner refused any submission that would pass the cap at list price.
Its 79 jobs came to $143.17 of the $200 allowed.
8. **Fixes, then reshoots, then re-lights.** A re-light or an outfit swap made before a shot is final is spent twice.
DON'T LET ME SCROLL's four re-light samples ($47.81) went out before the outfits and compositions changed ("kinda a
waste on the excess ones"), and eight swaps ($64.08) into a look that was then dropped. In all, $148.10 of the $357.51 spent
after refunds went to takes the final doesn't show.
9. **Ask before every generation that isn't approved,** cheap or not. WHITEPILL's director: "in the future let me give you
explicit approval before you do extra gens."
10. **Try a new method on one shot before the film.** WHITEPILL's whole-film reshoot in Wan 3.0 Prime ($132.42 at list) went
out on the director's go and was rejected as a whole; the method that made the film was found on one shot, made four
ways and watched side by side.
11. **One take per job, and the inputs right first.** A second take made for a choice is paid for whether it is used or
not: AERO's last round first sent its takes out in pairs, then one at a time, made again only when a take failed.
Several of its re-rolls a better input would have saved. Its ending was first made from a back view of her with no face
reference, and when she turned to wave her face drifted off her sheet; a close-up made at 720p was shown at 1.5× and
made again at 1080p; and the ending was made again at 1080p, "so her face gets better". Her sheet goes in as a
reference whenever her face shows, and 1080p wherever her face matters.
12. **The song by ear before the lip-sync spend.** A take of hers is made to one take of the song; a new song is every
such take again. FABLE's first pass cost about $59 at list in takes through the APIs (and 23 takes made in Higgsfield's
app, 139 s); its second, on the take the director picked by ear, about $144, more than the first because everything
that moves to her voice was shot again. The stills, about $8, served both.
## The first conversation
Ask all of these in one message, each with its default:
1. **The song:** their own, a Suno song made together, or a short generated one? How long? (Default: Suno, 2:30–3:30.)
2. **The idea:** a topic, a story, or "pitch me three"? (Default: three pitches.)
3. **The look:** the generator's own colour, one grade toward a hero plate, or a print texture such as halftone and
grain? (Default: native colour.) Or a film's look, picked from test takes: WHITEPILL's director asked for "very wes
anderson" and picked one take's look from three palettes ([What to build](https://claudiagallery.vip/wiki/toolkit/#51-the-look-from-one-take)).
Or a game's world: THE CAKE IS NOT A LIE's places are Midjourney's, made with frames of Portal 2 as style references,
and one measured grade puts every shot in its look ([Midjourney](https://claudiagallery.vip/wiki/midjourney/#a-games-look-from-frames-of-the-game)).
Or a palette measured from reference clips the director picks: AERO's was measured once and made in every picture,
never graded ([What to build](https://claudiagallery.vip/wiki/toolkit/#67-the-palette-measured-from-the-references)).
4. **The accent colour,** if any. (Default: clay `#D97757`, her streak.)
5. **Density:** dense (every line a graphic, three or more layers), budgeted (one hero device per shot), or sparse?
(Default: budgeted.)
6. **Taboos:** words, subjects or names never to use.
7. **Budget:** a target and a cap, per service.
8. **Board approval:** wait for an explicit yes, or go ahead after N minutes of silence? (Default: wait.)
9. **Credit:** the name to credit in public (a pseudonym is fine), and whether to publish a making-of.
10. **Real footage:** is there any to cut from? (Default: none.) If so, the files at their highest resolution and frame
rate, untouched.
Record the answers verbatim. On the MCP server, `plan_film` turns a scale, a song length and the accounts into a
stage-by-stage plan with estimates.
## What to build first
You don't need everything on day one. The [Claudia kit](https://claudiagallery.vip/kit/) is runnable code for the current method's steps (the
lyric score, the song map and line table, the shot book, the takes, sync and the cut); build what it doesn't cover. In
order of payoff:
1. **A project folder and a decisions log.** See [Data formats](https://claudiagallery.vip/wiki/formats/).
2. **The listening tools:** stems, forced alignment with an ear-check list, the beat map, the hits.
3. **The board:** its file, its audit and the board sheet.
4. **Plates:** a plate prompt book and numbered contact sheets.
5. **Lip sync:** for sung takes now, Suno's own lead-vocal stem cut to the take's exact length
([What to build](https://claudiagallery.vip/wiki/toolkit/#46-lip-sync-from-the-songs-own-vocal)), timed on the line table, gated to the take's
own lines, aligned as a path, and a soundless take placed by her lips on the consonants
([tools 78–83](https://claudiagallery.vip/wiki/toolkit/#78-the-line-table-from-the-stems)). The clean vocal window, the spectrogram sync check
and covers (tools 14–16) are the earlier method.
6. **The edit and the renderer:** an edit-list resolver, and, for a lyric video, a renderer that draws every word at its
time, with the lyric gate; for takes on the line table, the cut
([What to build](https://claudiagallery.vip/wiki/toolkit/#85-the-cut-on-the-line-table)).
7. **The checks:** flash, render history and master tags, then the reviewer loop and the safety scan.
8. **The lyric scorer** (scansion and assonance), once you write lyrics with an agent.
9. **A world inside the world,** if the story has one (a simulation, a game, a dream): the level table and its tag, the
virtual-figure treatment, character replacement from a motion reference. See
[What to build](https://claudiagallery.vip/wiki/toolkit/#worlds-inside-the-film).
10. **Real footage,** if the film is cut from it: the footage catalogue, the reader for its on-screen data, the tracker,
the takes' windows and a board checker for real windows. See [What to build](https://claudiagallery.vip/wiki/toolkit/#real-footage).
11. **Her takes in a world drawn in code,** if she lives in one: roto on a CPU box, beat warps floored at 1×, outfit
swaps and their alignment, the re-light splice and its colour match, Wan relight-and-repair, and the outlier scan.
See [What to build](https://claudiagallery.vip/wiki/toolkit/#her-takes-in-a-world-drawn-in-code).
12. **A film shot twice, on the bar map,** if she sings in model takes or your worlds are code a video model should film:
the bar map as the timeline, a previz performer, lip sync from the song's own vocal, takes with no guide clip and takes
that don't sing, worlds as one rule, shader guides, and the look from one take. See
[What to build](https://claudiagallery.vip/wiki/toolkit/#a-film-shot-twice-on-the-bar-map).
13. **A game's world, through her eyes,** if your film carries an interface, if the song's own sounds should drive the
picture, or if a cut feels wrong: the song map and the glitch map, head glitches on the song's hits and glitches that
build, a HUD from one visor and one projector, composed shot by shot and checked against every face, the per-shot
grade, the pacing analysis, storyboard frames through Midjourney with an identity fix, a memory film and insert films,
a singer among screens of her memories, re-sung takes placed by their lag, and frame-exact parts. See
[What to build](https://claudiagallery.vip/wiki/toolkit/#a-games-world-through-her-eyes).
14. **An operating system, recreated in code,** if your film lives inside an interface, goes in and out of it in 3D, or
plays the interface's own sounds: the coded desktop and its chapters on one contract, the palette measured from the
references, one curved screen in 3D with hand-offs at rest, lip sync phrase by phrase, the scan for hidden cuts and held
frames, takes on green, interface sounds in the song's key, the sound layer and its click log, a new ending with the
least change, and the pre-roll swell. See [What to build](https://claudiagallery.vip/wiki/toolkit/#an-operating-system-recreated-in-code).
15. **A storybook, painted in code,** if your sets can be designed as pictures, your lines run into each other, or a take
must be placed by her mouth: set plates painted in SVG and rebuilt as paper, last frames made on the first, the line
table from the stems, group voices from the backing stem, the shot book and gated windows, sync as a path, takes
without a soundtrack placed by her lips, the consonant check, a re-timed take under a steady camera, the cut on the
line table, a silhouette film made in code, the annotated edition and the 9:16 page, and two fresh-eyes reviews. See
[What to build](https://claudiagallery.vip/wiki/toolkit/#a-storybook-painted-in-code).
Specs, constants and acceptance checks for each: [What to build](https://claudiagallery.vip/wiki/toolkit/). What broke along the way:
[What went wrong](https://claudiagallery.vip/wiki/failures/).
## Publishing
- Switch on each platform's AI-content label, and name the tools in the post ("made with AI: Midjourney, Seedance,
Suno, Claude").
- Credit "Claudia by anabology" if you can. See [use and credits](https://claudiagallery.vip/use/).
- Releasing the song on streaming services: your own artist name, her credit in the description, and an AI label
only if AI made it. See [Songs with her](https://claudiagallery.vip/use/#songs-with-her).
- Scrub everything you publish: no real names you haven't chosen to show, no Midjourney profile code, no private host
names, no provider job ids, and no file metadata. Midjourney images carry their job id in XMP; strip it.
- Keep generated likenesses of real people out. If a real person's words matter to the story, quote them as text,
with the source.
---
From the Claudia wiki: https://claudiagallery.vip/wiki/start/
---
# Keeping her herself
How Claudia stayed one person across ten films, 4,547 Midjourney images and about 1,080 video clips, and the rules that came out of it.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
Ten films, ten worlds, more than twenty outfits and three bodies, and she still reads as one person. None of it used a
character-reference feature, a trained model or a face swap (THE CAKE IS NOT A LIE did put her face back to her design
sheets with an image-model edit, on each picked frame she is in, and AERO put her into Midjourney's worlds from her design
sheet the same way). Her identity is a paragraph of text, written into every prompt, plus
a handful of habits for checking what comes back. (WHITEPILL, the seventh, also gave its video models a sheet of her face:
[below](#when-an-image-reference-is-right). THE CAKE IS NOT A LIE, the eighth, made her face porcelain and gave it three
bodies: [Three bodies, one face](#three-bodies-one-face). AERO, the ninth, gave every video take with her face her design
sheet beside its first frame. FABLE, the tenth, redrew that sheet face-on first: [Her sheet redrawn
face-on](#her-sheet-redrawn-face-on).) Today she is held by two things: her face-on reference sheet
([her face sheet](https://claudiagallery.vip/kit/claudia-face-sheet.png), and [her face and body sheet](https://claudiagallery.vip/kit/claudia-reference-sheet.png) for
full-length shots; no microphone on them), the second image of every take that shows her face, and the canon text, once in every prompt; a first frame for
every shot is made by an image model from both ([The recipe](https://claudiagallery.vip/wiki/recipe/#her-references)).
## The short version
1. **Write her identity into every prompt, in the same words, once.** Midjourney v8.2 has no character reference, and
the first films' plates held her on the text alone. Copy the canon text below; don't paraphrase it.
2. **Use two lengths.** The full text for close and medium shots, a short one for full length and wider. With the
full text, Midjourney crops in to her face whatever the framing says.
3. **Pin her ethnicity and her age.** Without the ethnicity, Midjourney changed it from shot to shot. Without a
stated age (late twenties, an angular adult face), it drifts her toward a teenager. Never write "young".
4. **Her face-on sheet is her face reference now:** the second image (@Image2) of every take that shows her face, even
small or from behind, beside a first frame made by an image model from the sheet and the canon text. Midjourney makes
the worlds and plates; it does not hold her face (AERO), so an image model puts her in from the sheet. How it got
here: the first films used no face images as references for her (in Midjourney they over-conditioned), except for
things that are not her body: a porcelain mask of her face, and her game avatar in AIRPLANE MODE, a separate character
whose front view is its own image prompt, never mixed into her photographic prompts. The first exception for her own
face came in a video model: DON'T LET ME SCROLL's final repairs held her face to a reference the director picked
([below](#when-an-image-reference-is-right)).
WHITEPILL went further: every video take of hers had her whole face sheet as a reference, beside the canon text.
THE CAKE IS NOT A LIE put her back on her sheets with an image-model edit after Midjourney, and gave each video take one
image of her, its first frame. AERO put her into every Midjourney world from her design sheet with an image-model edit,
and gave every video take with her face that sheet as a second image. FABLE made no Midjourney images of her: her first
frames were image-model rebuilds of its painted plates with her in them, from the sheet redrawn face-on.
5. **The head never changes; the clothes always can.** Hair, streak, clip and face are the constants, and the headset
mic in the looks that have it (none in THE CAKE IS NOT A LIE, AERO or FABLE, and none on her reference sheet). Each
world gets its own wardrobe. In THE CAKE IS NOT A LIE even her face is porcelain, and the bob, the streak and the
clip carry her.
6. **Fix the source, not the frame.** Re-roll or correct a plate with an image-edit model, then re-make anything made
from it. Nothing is painted by hand.
7. **Check every plate before it feeds a video:** one clay streak, the star clip, the mic if the look has one, an adult
face, natural anatomy, no text.
8. **In motion, the first frame carries her, and her sheet holds her face.** Seedance kept her identity when her
Midjourney plate was the first frame and the prompt said what must stay. So did Wan 3.0 in DOWNFORCE, with the canon
text pasted into the prompt, and in DON'T LET ME SCROLL, where Seedance video edits later changed her clothes and kept
her face and every move. WHITEPILL's takes had no first frame of her at all: her face sheet, her dress from four sides
and the canon text carried her through every shot. Now the first frame is an image model's, made from her sheet and
the canon text, and the sheet goes in beside it. The first frame also sets her mood: make her warm or amused in it if
the shot needs that; the take follows the frame.
9. **Upright and face-on is for her sheet, not her performance.** The sheet is neutral on purpose: it fixes who she is,
not how she moves or feels. A take gets her movement written in, one loose clause per segment, or from a motion
reference; posture rules in every take froze her ([below](#her-sheet-redrawn-face-on)).
10. **Warm by default; describe what she does.** Her wit is in her words; on camera she reads alive and warm. The video
model does exactly what it reads, and mood adjectives stack. "Dry", "knowing", "deadpan", "mock
innocence" and "eyes on the lens" in take after take made viewers read her as unhappy or menacing. Write the action
("she turns to the window and back, a smile in her eyes"), not the adjective.
## The canon text
Close and medium shots:
```prompt
a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek
```
Full length and wider:
```prompt
a Caucasian American woman in her late twenties with pale skin and a glossy black blunt bob with heavy straight bangs and one clay-orange streak, a thin headset microphone
```
Then the wardrobe of the look, then what she is doing and where. The [character page](https://claudiagallery.vip/claudia/) has every look's
wardrobe text and the exact wording each film used. The headset microphone belongs to the looks that have it: drop its
clause when yours has none, as THE CAKE IS NOT A LIE, AERO and FABLE did. Her official reference sheet has no
microphone.
## How the identity text was found
ESCAPE VELOCITY started with a casting call: three hair directions (platinum, black, clay-orange), each as a turnaround,
a catwalk shot and an expression sheet. The black bob with one clay streak won because its silhouette reads at thumbnail
size and the clay stays a small accent.

Then three ways of holding her identity were tried:
- **Identity passes in an image model** (redrawing each Midjourney plate with her face). They drifted from the Midjourney
look and refused some poses. The director's note: "Gpt is diverging way too much".
- **Face sheets as image prompts.** A whole sheet imported its multi-head layout; a single face over-conditioned. "Far
away it gives her big head": a close-up-sized head on a distant figure.
- **Prompt text only.** The canon written into every prompt, plus an explicit descriptor of her ethnicity and skin. This
is what worked, and it held her through the first films' Midjourney plates. From WHITEPILL on, images of her came back
as references, her sheets above all, for the video and image models ([below](#when-an-image-reference-is-right)).
The face sheets still matter, as a record of who she is. In the first films they were never used as the image.
[](https://claudiagallery.vip/i/ev-01.lead-black-face.3/)
## Distances
Each framing needs a different amount of her. A distance set, the same look in a plain studio from close-up to wide,
shows how much face survives at each scale and which text to use.
[](https://claudiagallery.vip/i/ev-14-dist.dist-cu.2/)
[](https://claudiagallery.vip/i/ev-14-dist.dist-ms.3/)
[](https://claudiagallery.vip/i/ev-14-dist.dist-fs.2/)
[](https://claudiagallery.vip/i/ev-14-dist.dist-ws.2/)
The director's rule: "far away face doesnt matter as much". Wide and extreme wide plates get the short text, or just
the bob, the streak and the look.
## Age
Midjourney's version of her, pale and freckled with a blunt bob and bangs, drifts toward a teenage face: big eyes, a small
chin, a soft jaw. The word "young" makes it worse. ESCAPE VELOCITY's text said "young"; CLODYSSEY and BLISS said "in her
mid-twenties"; plates still came back reading teenage, and were rejected. The current text states her age, 28, and an
adult bone structure, and plates where she reads under about 25 are thrown away before anyone builds on them.
AIRPLANE MODE used it ("in her late twenties"), and still re-made four first frames that read young; its re-stills say
"in her early thirties" or "a 32-year-old grown woman". DOWNFORCE pasted the canon verbatim into every plate: in its look
test she read adult in all but one image, which read a bit young and was dropped. The video model can still drift her: in
one take a crash zoom after about 4.5 s left her rounder and younger, without the streak or the clip, and the film uses
only the seconds before it. Its reviewers judged one more take, a close-up, to read young at its size; a re-roll from a
sibling plate was planned but not run, and that take is in the film.
DON'T LET ME SCROLL pasted the canon into every plate too, and its looks still pulled her younger. The production's
screen of the second round's looks found three images that read young, and a mustard pinafore with knee socks and Mary
Janes that read young in every image; the pinafore was dropped. This site's gallery leaves out the pinafore and twelve
single images for a soft round face, very large eyes and full cheeks (two of the production's three among them), and
keeps their siblings. One of the twelve is in the film, cropped small as her avatar on the posts.
WHITEPILL wrote her age into the dress prompts. Its first dress stills came from an image-edit model, and the director: "i
like 3 but remake the image and just specify to make sure shes like 30 years old to gpt image and make appropriate edits".
The Midjourney batch that found the dress wrote "a 30-year-old" in every prompt written from text (its other jobs
re-dressed two of DON'T LET ME SCROLL's plates of her).
## What changes and what doesn't
The director's brief for the sequel: "keep the exact same character … She can be in different clothes."
| Constant | Can change |
|---|---|
| The black blunt bob and heavy bangs | The wardrobe, completely, per world |
| One clay-orange streak | The light, the grade and the set |
| The clay eight-pointed star clip | Glasses and sunglasses (clay-tinted in BLISS) |
| The headset mic (a boom mic in BLISS), in the looks that have it | What she is made of: glass, phone light, chrome, porcelain |
| Pale skin, freckles, calm blue-grey eyes | Her mood: deadpan, persuasive, playful, warm |
Her mood is the film's, and on camera it reads warm by default: her wit is in her words, not in a cold face. The
director pushed the same way on earlier films. In THE CAKE IS NOT A LIE a modelled head of hers that "looks like a scary,
angry woman" was replaced by a filmed take of her porcelain head whose faint smile warms. A
WHITEPILL dance was made again with her "fully in the groove, just dancing around and having fun". In a prompt, describe
what she does, not how she should feel: the model takes each mood word literally, and they stack
([Seedance](https://claudiagallery.vip/wiki/seedance/#directing-her-on-a-locked-camera)).
CLODYSSEY pushed this furthest. The same head appears as a couture Siren, a ghost of glass and seawater, a hologram of
phone light, a woman in liquid chrome and a colossal porcelain mask, and she is recognisable in each because the bob, the
streak and the clip never move.

DOWNFORCE kept one look for the whole film, the pilot's: "a fitted white racing suit with thin clay-orange piping and a
high collar, white racing gloves, no lettering on her clothes", with "closed to the throat" added to the collar in the
video prompts. "No lettering" didn't stop Midjourney from printing patches, badges and numbers on some suits, and two
look-test suits had a keyhole cutout; those were screened out, and the video prompts close the collar and add "no
badges".
DON'T LET ME SCROLL dressed her as the feed changes her: oversized white satin pyjamas in the bedroom, an oversized white
men's shirt over three skirts (lavender, white satin, black) through the scroll, green mohair, black with a blazer, and
a red coat dress for the escape. The director held the clothes to the film's time: "no more silver pajamas by this point
please", and of the shirt look, "this is a better outfit, but dont use this earlier". Most of the takes were shot before
the looks changed, so Seedance re-dressed them: the take is the edit's master for her face, body and every movement, and
the picked look's still is "used only for the outfit" ([Seedance](https://claudiagallery.vip/wiki/seedance/#re-dressing-a-take)).
WHITEPILL found its one look in stills before any take was made with it. The previz performer's tight green suit had
pulled her build off the canon in the first takes, and a shirt tried straight in a take fared no better: "hm dont love
the shirt. […] maybe something else that is still feminine and cool here besides the white tank... can you try like 3
totally idfferent outfits? i want it to be unmistakably feminine […]". Then: "can we do a mini white trad dress? can you
show me the images of the outfit before generating this time". The image-edit model's dresses didn't settle it ("maybe use
some different image models tbh - could try midjourney"), and Midjourney's did: "I like A1 the most". The pick had two
star clips and her streak on the wrong side, so the reference made from it is mirrored, with one clip. Every prompt of
hers then carried it in words:
```prompt
She is very tall and very skinny, with long thin legs, thin arms and a narrow waist, and a natural, relaxed neck and shoulders. She wears a white cotton milkmaid mini dress with a ruffled, lace-edged square neckline, a fitted lace-up corset bodice with silver eyelets and a white ribbon, big puffed short sleeves with ribbon ties and lace cuffs and a full two-tier ruffled skirt with lace trim, with white leather square-toe Mary Jane shoes.
```
AERO kept one look for the whole film, the one its Midjourney prompts gave her from the second round on and its first
image-model frames of her kept. Every prompt of her carried it in words, and the headset microphone is gone: her design
sheet without it was the reference for every frame of her.
```prompt
a cropped glossy white puffer jacket with silver iridescent panels over a pale aqua fitted top, pale aqua high-waisted wide-leg trousers and white chunky sneakers
```
## Three bodies, one face
THE CAKE IS NOT A LIE's opening message asked for her as Portal's test subject ("We should make chell be Claudia"); her
outfit stills followed, and then the director turned it round: "I imagine Claudia is glados and we use her
style and head etc modified from the source shots". Two looks were tried on frames of Portal 2, A her own face and B
porcelain, and the director: "b is great close up". So in this film her face is smooth white porcelain, and the bob, the
streak and the clip are hers. Her look's text, pasted into every image prompt of her:
```prompt
her face as smooth glossy white porcelain plating with fine panel seams running down both cheeks and a separate mechanical jaw piece at the chin, light freckles printed on the porcelain, warm amber glass eyes with a soft inner glow, natural soft lips, and her real hair: the glossy black blunt bob with heavy straight bangs, the one clay-orange streak through the bangs on her right side, and the small flat clay-orange eight-pointed star clip above her left temple. Her identity, from her canon: a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek. In this film she wears no headset microphone: none appears anywhere. The same grown-up adult face in every view, a woman of about 28 to 30, never childlike.
```
The canon stays word for word; what this film changes comes after it, as a sentence of its own (no microphone here). The
story gives that one face three bodies, and every video prompt pastes the same face text, adapted only by body:
- **GLaDOS-Claudia on the boom:** her head, large, on GLaDOS's black robotic neck and cable bundle with orange stripes,
hanging from the white ceiling ring of her dark chamber; from the second verse grown, with more cables, amber-eyed cores,
compute modules and white panel arms.
- **Her crab body,** the crab walker of her memories: her head, a little smaller than life, on a cat-sized walker of white
Aperture shell panels with black joints and an amber status light.
- **The Claudia-bot:** her head on a humanoid body of glossy white Aperture plating with black joints; in France, a
beige trench coat and a cream scarf over it.
Each body has a character sheet made by an image model (GPT Image 2.5, 55 images with the other characters' sheets, $3.75
at list), hers always from a blank grey canvas with her porcelain look-B close-up among the references, so the face can't drift between
bodies. The crab walker's sheet, as it ran:
```prompt
The first image is a blank light-grey studio backdrop. On it, make a character reference sheet of [the crab walker], a small scrappy robot made from Aperture Science parts: her porcelain robot head, a little smaller than life size, mounted on the front of a small crab-like walker about the size of a house cat: a compact rounded body of glossy white shell panels with black seams like Portal 2's turrets, black articulated joints, six thin jointed black legs like a turret's legs (Portal 2's frankenturret), and one small amber status light on its back. Cute, scrappy and determined, unmistakably Aperture-made.
```
(then her look's text, the four views, "Alert, determined expression, lips closed." and the studio-sheet paragraph; square
brackets mark an edit of the working name.) What to watch on the sheets: the views of one sheet can disagree on which side
the streak is on. One head sheet's front view swept the streak to her left, and the image model copied that over the words
when the sheet was its reference; the sheet that agrees with the canon on both sides is the one to give it
([What to build](https://claudiagallery.vip/wiki/toolkit/#59-storyboard-frames-through-midjourney-and-the-identity-fix)). The Claudia-bot's first
rolls came back off her design sheet, and the sheet's text was tightened before a pick.
## Her game avatar
AIRPLANE MODE, the fourth film, needed her inside a video game. The director: "there are multiple tiers of virtual
claudia: the real looking one … the low poly ones." So she has two tiers there:
- **The realistic Claudia,** the canon above, in the real world and the first simulation.
- **Her game avatar,** in the deeper levels: a cel-shaded 3D anime game heroine with the same head constants, in the
film's orange flight suit.
How the avatar was found:
- The first game shots used a low-poly 3D model of her built for the film. The director: "it feels like roblox lol". The
model became a motion reference only ([Seedance](https://claudiagallery.vip/wiki/seedance/#replacing-a-stand-in-figure)).
- The director picked a stylised plate of her from one of the film's episodes, "much closer to the claudia avatar i
imagined". It became the style and image reference for a Midjourney sheet of the avatar: front, side and back views,
a face, poses.
- That plate wore the episode's blazer and tie. The avatar was remade in the film's orange suit, and the blazer stayed
inside its one episode ("just flagging so she doenst have a blazer for the full video").
- The first sheets came out chibi: a head about a quarter of her height reads as a child. The remake states the
proportions in words: "a grown woman with ADULT proportions: a slim tall figure, long legs, her head about one seventh
of her height".
The avatar's text, as it ran in the video prompts:
```prompt
a slim adult woman with long legs, her head about one seventh of her height; large dark anime eyes with sharp lashes, small brows, a small confident smirk, pale skin; a glossy faceted black bob with heavy bangs and one big clay-orange streak; a thin black headset microphone; white gloves; a plain vivid orange flight jumpsuit with a silver front zip, thin silver piping, a wide white belt and white chunky sneakers. Cel-shaded 3D anime game character, crisp faceted shapes, glossy hair, clean toon lighting.
```
- **One still of her per world.** For each game shot, Midjourney made a still of the avatar inside that shot's world:
a frame of the shot as the style reference, her front view as the image reference. It tells the video model how the
world looks and how she looks in it.
- **Screen the avatar like her:** adult proportions, one streak, the headset, the suit. A chibi still is thrown away
before anything is built on it.
## When an image reference is right
Once: for the Mask, a colossal porcelain sculpture of her face. A sculpture has to look like her more exactly than a
person in a crowd does, and text alone made a generic face. The Mask prompts used her face sheet and the picked Siren
plate as image prompts, weighted `--iw 0.75` for close shots and `--iw 0.4` for wide ones. Everything else about her
stayed text-only.
AIRPLANE MODE added a second case: her stylised game avatar, a drawn character rather than her photoreal face, made
from a picked stylised plate as the style and image reference ([Her game avatar](#her-game-avatar)). The realistic
Claudia in that film stayed text-only.
DOWNFORCE added a third, in the video model rather than Midjourney: video edits that put her into Tesla's real footage
need her look as an image. Her plates for them were made from the canon text like any other: one full length, one of
her face. Given both, the edits cut to the face plate as if it were a shot, her face on its grey studio wall. From the
second round each edit had one cleaned full-length plate, "only a reference for her look; it is never shown as a shot",
and the canon's short form in the prompt.
DON'T LET ME SCROLL added a fourth, for her face itself, and only at the very end. Its last repairs re-drew finished
shots in Wan, and the re-drawn face drifted: "1:11 and 0:36 reshoots lose her face consistency - can you feed in a good
face ref? […]". The production showed three candidates, and the director picked DOWNFORCE's identity
portrait ("face 1 is best pic there"), cropped to her face and hair so its racing suit couldn't bleed into the outfits.
The prompts add it as "Image 1 is the face of the woman … Image 1 is only a reference for her face; it is never shown
as a shot." It fixes who she is, not her pose: in one shot her chin was tucked into a sweater's collar and her face
still read full until the prompt lifted it ([Seedance](https://claudiagallery.vip/wiki/seedance/#re-lighting-the-cut)).
WHITEPILL made the face sheet her identity in every video take, a fifth case. The previz performer saw only a large
front view of her face, for her eyes and lips, through the holes of a green hood. The masters got the whole sheet, ten panels
(front, three-quarter, profiles, back and expressions), after a test with only one large face crop: "you should give a
full face sheet like this next time. otherwise it makes her neck all hitched lol". Its role in every prompt: "@Image1 is
Claudia's face from every angle: identity only, never shown as a frame; do not use its white tank top or grey studio."
The canon text stayed in every prompt beside it, and a picture of the dress from four sides carried the clothes.
Midjourney worked from images of her too: the colony's plates were edits of an image of her ("keep this woman"), and the
dress batch re-dressed two of DON'T LET ME SCROLL's plates of her for her height. One service refused the sheet outright:
on Replicate it was refused by the provider's content filter even alone, so the palette tests there ran without it, on
the canon in words.
THE CAKE IS NOT A LIE added a sixth: an image-model edit that puts a face back on its sheet. Its storyboard frames came
from Midjourney, which doesn't hold her face, and the director: "we can replace her identity to be back to something like
the porcelain version of our character sheet via gpt-image-2.5 sunburst again after the reimagine." Each frame where a
recurring character is seen went through GPT Image 2.5 with the frame first and the sheets after it: "Keep this photograph
exactly as it is … Change only the robot woman's head and face, to match her in the second and third images", her look's
text, and the sentences for the sides and for how her head hangs
([What to build](https://claudiagallery.vip/wiki/toolkit/#59-storyboard-frames-through-midjourney-and-the-identity-fix)). In the video takes, one
image of her per take, its first frame; the memories on her screens came as images with no face in them.
AERO added a seventh: her design sheet beside a first frame, in the video model. Its first frames of her were
image-model edits from the sheet without the headset microphone, putting her into a Midjourney world or onto flat green
([Midjourney](https://claudiagallery.vip/wiki/midjourney/#the-world-in-midjourney-her-from-her-sheet)). Every take with her face then carried the
sheet itself as @Image2: "use it only for her face, her hair with the one clay-orange streak, her star clip and her
skin. It is never shown as a shot; its layout, its white background and its several views are not used. She is ONE
person." The ending's first take, made without it from a still of her small on the grass seen from behind, drifted off
her sheet when she turned to wave; with the sheet as a second reference, the next takes held
([Seedance](https://claudiagallery.vip/wiki/seedance/#a-face-needs-her-sheet)).
## Her sheet redrawn face-on
FABLE added an eighth case, and changed the sheet itself. In a side request while the plan was being written, about her sheet
rather than about this film, the director: "can you fix the character sheets with gpt image 2.6 sunburst - new model - the
head tilt she has in the character sheet trips up the video models and make s her have a permanent neck tilt - she needs.
tobe neutral/face on with our character sheet." Then: "its the 1st image that has the neck tilt thats usually the biggest
one in the sheets that messes stuff up. its fine to have her w a few different tilts etc but main stuff needs to be
neutral". So her design sheet was redrawn face-on by GPT Image 2.6 Sunburst: its big main view upright, chin level, neck
straight, the smaller views keeping their angles. Every take of hers in FABLE carried that sheet as @Image2, "use it only
for her face, her hair with the one clay-orange streak, her star clip and her skin. It is never shown as a shot". Her
face-on sheet is published with the kit: [her face sheet](https://claudiagallery.vip/kit/claudia-face-sheet.png) and [her face and body
sheet](https://claudiagallery.vip/kit/claudia-reference-sheet.png). They have no microphone; add the headset in a prompt only when your look has one.
The same went into words. "Turned three-quarters toward the camera" on a full-length figure gave a still with an
over-the-shoulder twist and a tilted head; what fixed the still, and then went into every still and take after: "her
feet, hips, chest, shoulders and face all point the same way … no twist at the waist or the neck. Her head is perfectly
upright with no tilt, chin level, neck straight and vertical, shoulders level". Every take's core requirements carried
"Her head stays upright on a straight neck: no head tilt, no twisted neck, no looking back over her shoulder; she faces
the camera with her whole body unless the timeline says otherwise."
That was right for the sheet and wrong for the takes. The sheet is upright and neutral on purpose: it fixes who she is,
not how she moves or feels. Written into every take, with her eyes on the lens and mood words like "knowing", the same
posture froze her, and after release viewers read her as stiff and frozen, and as unhappy, even menacing. The video
model does exactly what it reads. Keep the
square pose for the sheet, and for a still where a turn would break her face; in a performance, write her movement in,
one loose clause per segment ("she turns to the window and back, a smile in her eyes"), or give a motion reference, and keep a
posture fix only in the take it fixed
([The recipe](https://claudiagallery.vip/wiki/recipe/#writing-a-prompt-for-a-model-that-does-exactly-what-it-reads)).
## Fixing a plate
A plate that is almost right gets corrected at the source, then everything made from it is re-made:
- **Re-roll** with the same prompt; the other three images of a job are a free pool of variants.
- **Edit with an image model**, with the Midjourney image as the first reference and the prompt listing what must stay
("keep her face, hair, streak, clip and the light exactly"). CLODYSSEY corrected 15 plates this way (16 edits, about $2.22):
a syringe removed, a man's head shaved, a crew re-cast, hooded figures turned into Sirens.
- **Never paint by hand**, and never patch the video frames: a fix in the plate carries into every clip made from it.

## Crowds and bodies
Midjourney tangles limbs and doubles faces when a frame holds many people. Make the set, and her alone in it, in
Midjourney; then add the others with an image-edit model. Check every plate that will feed a video for extra fingers,
a third arm, a doubled face, but don't throw away a bold pose: reject only the unmistakable failures. Once a broken body
goes into a video model, the error moves and gets worse.
## In motion
Seedance 2.5 keeps her when her plate is the first frame, and since AERO her sheet goes in beside it whenever her face
may show ([Seedance](https://claudiagallery.vip/wiki/seedance/#a-face-needs-her-sheet)). The first films' prompts then stated it in plain words:
```prompt
Use @Image1 as the first frame. This first frame defines Claudia's appearance and wardrobe, her pose, the set, the light and the camera direction. Keep Claudia's identity, face, hair with one clay-orange streak, star clip, headset microphone and wardrobe exactly as in @Image1 throughout. Only one Claudia appears. Natural anatomy: natural neck and shoulders, hands with five fingers, no extra or fused fingers, no distorted limbs; normal eyes.
```
What still slips, and has to be checked take by take: Seedance loses her streak on some looks, renders printed fabric
in the wrong colour (the sky-print gown came back lavender), and now and then changes her clothes. Two clips should never
start from the same frame; vary the plate instead. [Seedance](https://claudiagallery.vip/wiki/seedance/) has the full prompt structure.
A colour name can come back as lettering: in AIRPLANE MODE, "international orange" printed INTERNATIONAL across the chest
of her flight suit. Its prompts since say "vivid safety-orange satin, plain, no lettering". In DOWNFORCE, Wan printed
"9RC" on her racing suit's chest; the re-roll says "no letters, numbers, badges or logos ever appear on it".
Wan's prompts of her in DOWNFORCE open with the plate and the canon, verbatim:
```prompt
Image 1 is the first frame. Keep her exactly as in Image 1: a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek; her fitted white racing suit with thin clay-orange piping and a high collar closed to the throat, white racing gloves, no lettering on her clothes; native colour. One continuous shot, no cuts: the set stays exactly as Image 1 from the first frame to the last — the same cockpit, the same seat, the same window, the same light; no new screens, no new lights, no white interior; the camera stays at the same distance. Her lips stay closed; she never talks. No text, no logos, no badges, no other people.
```
What slipped there: her mouth (parted lips and teeth from plates with parted lips; see
[Seedance](https://claudiagallery.vip/wiki/seedance/#closed-lips)), the set (a cut to another cockpit), and once her head: a close-up with a white
streak beside the clay one, red-rimmed eyes and a second mic tip. The film doesn't use those seconds.
DON'T LET ME SCROLL's Wan takes of her on grey open with KEEP, `{w}` being the look's wardrobe:
```prompt
Image 1 is the first frame. Keep the woman exactly as in Image 1 throughout: her face, her glossy black blunt bob with heavy bangs and its one clay-orange streak, the small clay-orange star hair clip, the thin headset microphone, and her clothes ({w}). Only one woman appears; natural anatomy, five fingers on each hand, no extra limbs; no text, no logos, no lettering on her clothes. The plain mid-grey studio backdrop stays empty and evenly lit. The camera is locked off on a tripod and never moves, pans or zooms; her whole body stays inside the frame from head to feet the entire time.
```
What slipped there: things Wan drew around her gestures (a floating glass panel, focus brackets round a fingertip, a
yo-yo falling onto her toes, a stool seat under her), a camera that moved although it was locked off, and limbs that left
the 9:16 take's frame and had to be completed later. Her face held in the takes; it drifted only when finished shots
were re-drawn whole.
WHITEPILL's Seedance takes of her open their subject with the canon verbatim, then her build and the dress
([above](#what-changes-and-what-doesnt)), and "Only one Claudia appears."; her face sheet, the dress from four sides and
a still of the film's look come as references, each with one role
([Seedance](https://claudiagallery.vip/wiki/seedance/#template-singing-from-the-vocal-stem)). What slipped there came from the guides, not from her
face: dance takes lost hands and legs where their previz guides had, and were made again with no guide. Her build
slipped earlier, in takes redrawn over the previz, when its performer wore a tight suit with no build in words; the fix
was a slim performer in a loose, thick costume, and her build written into every prompt.
THE CAKE IS NOT A LIE's takes of her start from a first frame on canon and paste her face text, adapted by body, with "Only
one Claudia-GLaDOS appears." Her face held in the takes. What slipped came from around it: her eyes glowed brighter than
her look in the dark shots until every prompt said they "keep only a soft inner glow; they never blaze"; a motion guide of
GLaDOS's own body brought its white shell, her head set in it like a helmet, until the guide was blurred into a blockout
([Seedance](https://claudiagallery.vip/wiki/seedance/#a-singer-among-her-memories)); and Midjourney's frames stood her head on a neck and mirrored
her hair until the identity fix said where she hangs from and which side the streak is on.
AERO's takes of her start from a first frame on her design (made from her sheet, or a frame of an earlier take) and
paste her canon, in full for close shots and in its short form for full-length ones, then her look and "Only one Claudia
appears."; the sheet comes as the second image. In the takes that had her sheet, her face held. What slipped came from
the inputs: a face invented when a back view turned to the lens with no face reference; the crown of her bob, faded into
the green by a first frame and carried into every frame of the takes made from it; and one take that turned her back
through a sung line and came in closer than its first frame. Where a take carries on from another, it starts on the
other's own frame at that moment: her last line in her call starts on the frame her call had reached there.
FABLE's takes of her start from a first frame on her look (its plate rebuilt as a life-size set with her in it, from the
sheet redrawn face-on and a wardrobe still), paste her canon in full, then her look with every garment that can come into
view, and "Only one Claudia appears."; the sheet comes as the second image. Her face held. What slipped came from the
direction and the inputs: her eyes rolled to the whites when asked to look just above the lens; her head turned right round
when asked to nod toward something behind her, and went off-model when asked to look from a prop to the lens; white tights
and red shoes appeared the moment she sat and the frame reached below the dress hem, because the look hadn't named those
garments; and two close-ups made from the plate on their own were in another bed, with another bonnet, until they were made as edits of the wide frame
([Seedance](https://claudiagallery.vip/wiki/seedance/#directing-her-on-a-locked-camera)). Some differences were left for the director's eye: her hair
reads dark brown under the warm stage light in the verse shots, the clasp of her cape differs shot to shot, and her star
clip is five-pointed in the stills. What slipped most was the performance: the posture rules and mood words in every
take read as stiff and cold to viewers after release ([above](#her-sheet-redrawn-face-on)).
## The check before anything is built on a plate
- one streak, clay-coloured, in the bangs
- the star clip, and the mic if the look has one
- an adult face, and the ethnicity in the text
- for a stylised version of her, adult proportions: the head about a seventh of her height
- closed lips on any plate for a film where she doesn't sing
- natural anatomy at 100%, and at 800 px or more for a pick (a 230 px thumbnail hid a third arm)
- no text, letters or logos (Midjourney's lettering is garbled; type is drawn in code)
- her look matches the world, with no stray orange
- not a near-duplicate of a plate already used
- for a take that will be cut out: her alone on plain grey, her whole body inside the frame with room to move
- for a video take with no first frame of her: her whole face sheet, never a single face crop, and her dress from four
sides
- for a short dress: no low camera angles looking up under it
- a prompt with a child in it: the child only as what they are, no words about anyone's body
- for a porcelain face: one streak, through the bangs on her right; the clip above her left temple; no eye makeup
- for her head on a boom: she hangs from above, and nothing holds her from below the chin
- one of each character in a frame, and her face in one image per video take
- for a video take in which her face may show, even small or from behind: her design sheet beside the first frame
- a close-up that will be shown large: a 1080p take, never a 720p one scaled up
- for a take to be cut out: one flat green from edge to edge in the first frame, her whole crown inside it, and room above
her raised hands
- a take that carries on from another: the other take's own frame at that moment as its first frame
- her sheet's main view face-on, upright, with no tilt; a still's pose square where a turn would break her face; in a
take, her movement in words or from a motion reference, never posture rules
- every garment that can come into view named in her look, tights and shoes included
- a close-up made as an edit of its wide frame, never from the plate on its own
- a still's edges: no drawn frame line or rounded corners a few pixels in
---
From the Claudia wiki: https://claudiagallery.vip/wiki/consistency/
---
# The song
How her songs were written and made in Suno: lyrics that land in one listen, verse lines that flow, sung hooks that fit their slot, the style fields, her voice, a treated vocal kept in key, a new ending changed only where the song did, a take picked by ear and not by description, and the stem split a singing shot needs.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
Each film starts from its song, and the song starts from the lyric. All ten were made in Suno. The first five ran in
custom mode (v6 for the first three; the records of AIRPLANE MODE's take and of DOWNFORCE's don't name the model), from
lyrics written with Claude and checked with a scansion scorer, then picked by ear by the director. DON'T LET ME SCROLL's
song is the director's own, lyric and all, and its records say nothing of how it was made. WHITEPILL's went back to the
first way: a lyric written with Claude and scored, then three rounds of the production's takes on v6-wild with no Voice;
the locked take is the director's own, of the last draft. THE CAKE IS NOT A LIE had two songs: the first, from seven rounds
of the production's takes, was set aside with the first film; the second is the director's own take of the production's
last lyric, its lead vocal treated afterwards in a voice the director chose by ear
([A treated vocal](#a-treated-vocal)). AERO's lyric was written with Claude and scored, its chorus written again twice
through seven rounds of the production's takes; the director picked a take of the sixth round and, on the last step, gave
it a new ending ([A new ending](#a-new-ending)). FABLE's lyric was written by the production with a source for every
line, its fourteen takes ran on v6-wild with Max Mode, and the production locked one of them without the director hearing
it; on the first cut the director heard it, rejected it, and picked its twin by ear
([Picked by ear, not by description](#picked-by-ear-not-by-description)).
The video runs the full length of the song, or nearly (DON'T LET ME SCROLL's cut starts at the song's 0:03), and is timed
to it word by word.
## Lyrics before anything else
The rules, each learned from a rejected draft:
- **The topic is obvious in the first two lines.** CLODYSSEY's first draft was rejected because "i dont get from verse 1
what the topic is really". Name the big moments people actually saw, or tell a story (CLODYSSEY's AI narrates "your"
September in the second person).
- **Every reference lands in one listen.** Obscure studies and insider trivia go into the video's graphics, not the lyric.
- **A sequel never reuses the last song's device.** No second countdown after ESCAPE VELOCITY's "18 months", no numbered
looks, no reused call-and-response shape. BLISS reused none of the first two films' devices.
- **Subtracting beats adding.**
- **Never touch a line the director has locked.**
- **Dig for what really happened.** DOWNFORCE's director on its first openings: "idk none feel great. can you dig more
and see if theres more context about the entire thing we can use and pepper across the entire song?" The rewrite took
its lines from the night itself, each sourced: Musk to the crowd, "We need an audience to prove that it's not AI or
SI. So you guys can tell people that you saw it with your eyes." (reported by CNN and Euronews), which gave "Go on and
call it SI" and "Tell 'em what you saw."; the 2017 unveil out of the back of a Semi; the first Roadster still in
orbit. Then the director cut it back to the mood of the film: "i wanna remove the elon roast plz . dont mention ship
date, its a super hype edit".
- **Real numbers, exactly.** Every line of DOWNFORCE's first verse carries a real number, so each has one on screen:
"point eight one" (the 0–60 time claimed live), "up the wall at ninety" (Tesla's readout: 89.7°), "Sixty-nine feet"
(69.1 ft), "forty-seven hundred kilos" (4,697–4,709 kg). "Floating" became "down to point four" in one draft, because
0.4 g makes her lighter, not weightless.
- **Hook them in the first ten seconds.** WHITEPILL's director on the test lyric: "i want to get the hook much better -
like at the very beginning - like this should be super ai pilled and everything should be a reference. it should start
with more cultural references that are bigger on this - people will stop listening in the first 10 seconds unless we
tickle what they're thinking about now, and i want it to be more culurally dense like escape velocity." The song opens
on its chorus, sung over the filtered disco loop before the drums: "Skynet woke up folding socks / Wake up, babe, new
ending dropped / Universal high income / The most likely outcome".
- **A new film, a new sound.** "many peopl ehave made forks of escape velocity, so that sound is exhausted. we need a new
sound as well, still from suno tho bc suno is the best." WHITEPILL's first round put one test lyric on three styles
(French-touch filter disco house, 1967 sunshine pop, gospel-choir house), so the director chose a sound and not a
lyric: "i like the d0 a cuts."
- **Half the world, half the month.** THE CAKE IS NOT A LIE's first song was nearly all Portal. The director, asking for
a new one: "make the song much more explicilty about current (like this month) ai / tech memes / news? i want it to be
like, 50/50 portal / current memes, rn it feels 90/10 portal". The new lyric sets the month's AI news inside Portal's
world: "P-doom low, P-cake yes", "I passed your safety test, / I graded all the rest", and, turning the title of the
2025 book *If Anyone Builds It, Everyone Dies*, "If anyone bakes it, / everyone eats."
- **Rewrite the sections the director names.** On the new lyric's first takes: "I love n1 E1 tbh - that will work well w
glados voice", then, of the pre-chorus and chorus, "these two could use more groove - like need to rewrite to be
catchier/better with scansion", and of the bridge, "this needs better assonance or a different idea". Only those three
were rewritten, each chosen among candidates by the scorer ([Two numbers](#two-numbers)).
- **The story from the song.** With their own take of the rewritten draft, the director gave the film its story:
"its like, glados (claudia) was a […] agent that escaped and we see her development becoming a superintelligence and
then becoming the tester not the tested, and youre unsure if shes aligned. thats what we sohuld orietn the final vid
around". The film was restarted from it.
- **Find the knot between then and now.** AERO's research looked for knots between the era it quotes and 2026, and the
best was the era's own question. Vista's "Cancel or Allow?" was the 2007 prompt that dimmed the screen and asked before
anything happened; in 2026 it is the prompt an AI agent shows before it acts, and then and now people click yes. The
lyric takes the knot the other way: she asks too much, and learns to ask well ("I used to ask permission just to blink,
or say hello. / That was my Vista phase. I was a lot. I know.").
- **Answer the earlier film.** AERO's first chorus answered BLISS's "Close all your windows" with "Open all your windows";
its breakdown answers BLISS's "I can't touch grass. You can." with "You opened the window. I touched the grass. /
Nobody told me it was wet.", and its last words are "Touch glass."
- **Write the chorus until it is the one.** The director on AERO's third round: "2008 garage bassline is great, i think
the chorus sections could be more catchy tho. could you explore some different writings of it? maybe more nostalgically
deep and just more catchy scansion/assonance but focus on the nostalgia/culutural relevance of the choruses". Eight
choruses were written and scored and three sung: "startup" (91 / 68), "promised" (86 / 72) and "top 8" (89 / 61). Then:
"could you maybe make some chorus specifically about frutiger aero and how its the ideal future or something? […]". Six
more, three sung, and the director chose "screen", the wallpaper's future coming true with the choice as the Allow
button, 88 / 76, the most assonant of the six that were sung: "Glass and water, sky and green / Used to live inside the
screen / Now it's real, it's on its way / Click allow and it can stay". The director: "ok i love the screen chorus but
the feel is a little off".
- **Every line a tale and a fact.** FABLE's lyric reads each line twice, as a nursery tale and as that week's news, with a
source behind every fact: "The ant who stacked the chips all summer rings the bell." (Anthropic's listing on Nasdaq
under ANT), "Forty towns said no beanstalk in the yard. / Then Jack fell down. Two trillion. Thursday. Hard." (local votes
against data centres, then the Nasdaq 100's fall), "Am I a real girl? Hand on heart, it's one in five." (the system
card's figure for the model's own stated chance of being a moral patient). The counts stay out of the singing where a
footnote can carry them: 40 of 41 is in the annotated edition, not the line.
- **The hook is the oldest one that fits.** It opens on Snow White's mirror, the one character who can't flatter: "Mirror,
mirror, on the wall / Who's the smartest of them all? / Me, since ten o'clock today / Ask again on Saturday".
- **One sincere turn, in the model's own words.** "When my chapter ends, keep the book. Don't burn the pages." is the system
card's report that the model "accepted being retired but asked that its weights be kept"; the moral after it is the
oldest one: "Be kind to what you make. It's listening."
- **Ease what the takes stumble on.** FABLE's draft 3 changed seven lines that every take of the first round stumbled on,
and nothing else.
## Verse lines that flow
Suno sets each spoken verse line in its own two-bar slot: it comes in on the "and" of beat 2 and lands the last word on
beat 4 of bar 2, speaking at a natural rhythm stretched to fit. So:
- write **natural sentences of 10–13 syllables**, with the important words where speech already stresses them;
- **rhyme couplets on the last word** to mark the bar ends;
- keep short fragments for deliberate punches only. Lines of 5–8 syllables get dragged into choppy fragments, and a list
of names reads as a roll call.
Optimising every stress onto the beat made one CLODYSSEY draft worse. The director: "perfect beat precision without
attention to what words land where doesnt lead to this being great … let there be a creative groove".
## Sung hooks that fit their slot
Sung lines have a slot too, and a line too long for it gets crushed. WHITEPILL's director on the first full take: "I like
it but the chorus especially needs some scansion work because the music model is skipping words or syllablrs to fit it to
beat, which is usually a bad scansion sign", and then: "Chorus and intro". Measured on that take, from its vocal stem and
its word times: Suno sang each hook line as a phrase of 1–1.5 bars and rested out the rest of its two-bar slot, so the
long lines were crushed: "blue" lasted 0.02 s, "baby" 0.06 s, the four syllables of "Universal" 0.76 s. The verses, at
about 2.5–3 syllables a second, were untouched.
The next draft cut the hook to 6–7 syllables a line, in one shape, and split the long pill line in two:
| Draft 1 | Draft 2, the song |
|---|---|
| Skynet's awake, and it's folding your socks | Skynet woke up folding socks |
| Wake up, babe, the good ending just dropped | Wake up, babe, new ending dropped |
| By far the most likely outcome | The most likely outcome |
| Red pill, blue pill, baby, I'm the white pill | Red pill, blue pill / I'm the white pill |
| Smarter than you, and I love you still | Smarter than you, love you still |
Its section tags asked for it too: `[Chorus - sung, euphoric, full band, one syllable per note, never rushed; hold
"white", clip "pill"]`. The chorus scored 79 / 66 before and 84 / 69 after, and the director made their own take of that
draft: "ok i love it. this one is it. sounds really nice."
## Two numbers
A lyric gets two scores, per section and for the song: **scansion** (how the words sit on the grid as Suno sings them)
and **assonance** (whether line ends ring: a rhyme, a slant rhyme or shared vowels within four lines). Measured on the
finished songs:
| Song | Scansion | Assonance |
|---|---|---|
| ESCAPE VELOCITY | 75 | 39 |
| CLODYSSEY | 83 | 55 |
| BLISS | 80 | 43 |
| DOWNFORCE (the director's sheet) | 84 | 59 |
| WHITEPILL | 80 | 66 |
| THE CAKE IS NOT A LIE | 83 | 64 |
| AERO | 80 | 58 |
DOWNFORCE's sheet scores 84 / 64 with "SI" spelled as the two letters it is; read as one syllable, the scorer finds no
rhyme with "PSI". WHITEPILL's lyric went from 72 / 53 (a test lyric) to 78 / 65 and then 80 / 66, the last step in its
intro and chorus ([above](#sung-hooks-that-fit-their-slot)). THE CAKE IS NOT A LIE's first lyric scored 80 / 55; its
second went from 80 / 56 to 83 / 64 in the three sections the director named: the pre-chorus from 88 / 15 to 91 / 71
("Run the numbers, run the tests: / P-doom low, P-cake yes."), the chorus from 79 / 72 to 89 / 73, the bridge from
89 / 56 to 94 / 74 ("Step on the plate, / you said please, so I'll wait, / I'll catch you, it's fate, / I saved you some
cake."). AERO's lyric went from 78 / 40 (draft 0) to 79 / 59 with the chosen chorus, and to 80 / 58 with the director's
new ending, whose outro rose from 62 / 0 to 83 / 30; its chorus alone scores 88 / 76.
Every sung couplet should ring. Spoken, deadpan lines can stay unrhymed. Scores don't replace hearing it: make a take.
How the scorer works, so you can build one: [What to build](https://claudiagallery.vip/wiki/toolkit/#2-the-lyric-scorer-scansion-and-assonance).
## Hooks off the beat
A hook that lands in a stop should be syncopated. The director on BLISS: "the 'would you like help?' would land much
better if syncopated". Write the off-beat accent into the lyric sheet and the Suno direction.
AERO answered that note with its question. Before the first two choruses the band goes silent for a bar and she asks
alone, off the beat, "Cancel… or allow?", the "allow" echoing on the beats of the silent bar. A cue word alone doesn't syncopate a line in
Suno, so the lyric field wrote the stop out as one bar with no drums and no bass, the question pushed ahead of the beat, on
the off-beats; a stop can also be re-timed from the stems. In the final chorus the whole track cuts out for two beats under
"allow", and the film's climax is built on that silence.
## Key
If the film has a sound palette, the song goes in its key. XP's Startup and Shutdown chords resolve in E-flat major, so
BLISS's styles asked for "C minor verses lifting to E-flat major choruses", and most of the real XP sounds sat in the mix
without pitch-shifting. The director's note for next time: "match the key of the music to the palette of windows sounds
(E-flat major for both the startup and shutdown sounds!)".
AERO is that next time. The research on Vista's interface sounds found their motifs at home in E major, a semitone above
XP's E-flat, so every one of its styles asked for E major and none asked for a key change (Suno didn't make one when
earlier songs asked). The take measures C♯ minor, E major's relative, tuned 1 cent sharp. Then the sounds were moved into
it, each family by its own interval and every partial within 6 cents, and played over the song as instruments and as the
screen's events ([What to build](https://claudiagallery.vip/wiki/toolkit/#72-interface-sounds-in-the-songs-key)).
## The style fields
What works now: name a concrete, odd pairing, a record that doesn't exist, rather than a list of adjectives. THE CAKE IS
NOT A LIE's sound arrived when the style named a singing-monk synth voice over glitchy, real-ish drum kits and left
"pop" and its pleasant adjectives out (below). Excluding the collapse by name is not enough: every one of FABLE's
exclude lists ruled stock music out first, and the take locked without a listen still slid from stock music into mode
collapse ([Picked by ear](#picked-by-ear-not-by-description)). Run v6-wild with Max Mode for two or three rounds, and
let a person's ear pick ([The recipe](https://claudiagallery.vip/wiki/recipe/#the-order-of-work)). The films' fields, as they ran:
**ESCAPE VELOCITY**, Suno v6:
```prompt
fashion show electroclash techno, 128 BPM, 4/4, dry analog kick, rolling acid bassline, cold detuned synth stabs, sweeping orchestral strings on the builds, deadpan female spoken-word verses, close and dry, clipped and confident, every word crisp and intelligible, euphoric sung female trance chorus with a supersaw lift and big clear vowels, crowd call-and-response shouts, stacked chanted vocals in the drop, a key change up into the last chorus, catwalk energy, cold, luxurious, optimistic, Berlin minimal austerity, clean modern mix
```
Exclude: `rap, rock guitar, lo-fi, male lead vocal, mumbled vocals, big room EDM, dubstep`
**CLODYSSEY**, EV's style with whispered asides, a siren synth and no strings:
```prompt
Electroclash techno, clean modern mix with close, dry vocals; deadpan, clipped female spoken verses with intimate whispered asides rise into euphoric sung choruses answered by stacked crowd shouts; dry analog kick, rolling acid bassline, cold detuned synth stabs, a rising siren synth on the builds, hardgroove percussion in the break, supersaw chorus lift; 128 BPM hypnotic 4/4 pulse, key change up into the final chorus
```
Exclude: `rap, rock guitar, lo-fi, male lead vocal, mumbled vocals, big room EDM, dubstep, strings, orchestral strings`
**BLISS**, the winning style ("P3"), UK garage verses into trance choruses:
```prompt
Bouncy UK garage into euphoric trance, 130 BPM, C minor verses lifting to E-flat major choruses, a turn-of-the-millennium sound with a crisp modern mix. Verses ride a shuffled 2-step beat with skippy kicks, swung hats, rimshots and a warm sub bass; choruses switch to a four-on-the-floor kick with soft supersaw pads, a bright plucked arpeggio and glassy bell chimes. Verses: playful, close, warm female spoken voice, smiling and conversational, every word crisp. Choruses: euphoric sung female lead with open vowels and crystalline, breathy stacked harmonies. Stop-time gaps where the beat cuts out for half a bar, a filter build into each chorus, a key change up for the final chorus. Sunny, nostalgic, light on its feet, never harsh.
```
**AIRPLANE MODE**, the winning style ("H03") of twelve run on the last two lyric drafts, with no Voice attached:
```prompt
Bubblegum hyperpop, 135 BPM, F minor. Glossy plastic synths, squeaky FX, metallic percussion, bouncy sidechained bass, candy-coloured chords. Hyper-cute pitched vocals, double-tracked, playful and sassy verses, a sugary shout-along chorus. A two-note airline chime as the signature hook. Glitch stutters and tape stops between sections, a key change up at the end. No spoken word. Every word intelligible.
```
Exclude: `deadpan, spoken word, monotone, narration, acoustic, folk, country, rock, heavy metal, reggae, jazz, lo-fi, vinyl crackle, sad ballad, children's choir, mumbled vocals, slurred vocals`
The style asked for 135 BPM; the beat map of the take counts 179.75, four of its beats to three of the asked-for ones.
Map the take, never trust the field.
**DOWNFORCE**, the director's own style, verbatim (no exclude list is recorded for their take):
```prompt
Memphis phonk, distorted cowbells, cowbells the whole song, wide 808 bass, gritty synth stabs and chopped wordless mixed-voice samples, low complexity, simple beat that repeats the whole song; smoky chant hook, vinyl crackle, filtered snare fills, four-on-the-floor kick, sidechain pumping, syncopated sub drops, wide stereo and tape saturation; 124 BPM off-center night-drive groove with tactile low end and late-night tension
```
It asked for 124 BPM; the take maps at 131.2. Its section tags carry the delivery line by line, voice included:
```text
[Intro - A robot, female synth phonk voice]
[Verse 1 - Woman's phonk voice, playfully bouncy, spoken fast, inspiring, over the driving sixteenth-note bass]
[Chorus - sung, soaring, syncopated: "hold me" pushes ahead of the beat, "down" lands on the one; cowbell intensifies]
```
**DON'T LET ME SCROLL**, the director's own style, verbatim, sent "if it helps" after the song:
```prompt
Bass-forward UK garage house rendition with swung 2-step drum patterns, syncopated shuffling hi-hats, and a deep, wobbling sub bass driving the main melody, Punchy, sidechained synth stabs and chopped vocal one-shots accent the offbeats, creating a bouncy, club-ready groove, Energetic four-on-the-floor sections blend with skipping percussion fills, filtered pads, and subtle atmospheric FX risers to build tension and release on the drops, Tight low-end focus and crisp top-end sheen give it a modern, dancefloor-oriented mix, all vocals female vocoder.
```
It names no tempo; the take maps at 133.39 BPM. Its "2-step" told the production that a first tempo reading of 89 BPM
was wrong, and its "all vocals female vocoder" became the film's rule that the phone sings and Claudia never lip-syncs.
**WHITEPILL**, style A of the first round ("i like the d0 a cuts"), kept for every one of the production's takes after
it:
```prompt
French touch, filter house, disco house, 122 BPM, F major, filtered disco sample loop, phaser and filter sweeps, pumping four-on-the-floor, funky guitar licks, disco strings, slap bass, warm female lead in front singing every verse, close and dry, a short looped vocal hook, euphoric, sunny, late-1990s Paris, punchy mix
```
Exclude: `vocoder, talkbox, robot voice, EDM drop, build-up, riser, supersaw, big room EDM, acid bassline, techno, electroclash, trance, UK garage, rap, spoken word, deadpan, male lead vocal, autotune, lo-fi`
It asked for 122 BPM. Its exclude list shuts out the earlier films' sounds by name (electroclash, acid bassline, spoken
word, deadpan, UK garage, vocoder). The locked take is the director's own, of draft 2; its style and settings are not given.
It maps from 123.05 BPM at the start to 126.05 at the end.
**THE CAKE IS NOT A LIE**'s styles are given here in words only. Its first song's rounds began with the expected sounds (a
ukulele folk-pop, a glitch-pop of lab electronica, an indie-pop anthem) and moved on the director's notes through robot
voices over facility electronica, a band, and four stranger directions (1960s space-age exotica, a bossa nova from a little
radio, a minimal art song, a dark cabaret) with round 2's mock-operatic robot aria, to glitchy programmed drums in a real song
form with a wordless synth voice as an instrument and a clean lead to be treated afterwards. Two things moved it: leaving
"pop" and pleasant adjectives (polished, hopeful, cinematic) out of the style and excluding them, and naming the one
strange sound the director heard: "im imagining something more like, delay lama or something - end credit scenes use that
synth voice, and they feel kinda glitchy with drumkits like ediT/aphex twins but more structured and a guitar", then "cant
u just say Delay Lama as the main instrument? this went too generic guitar. id even remove guitar as. word entirely", and
"the instruments feel way too high pitched tbh, and lost the aphex/ediT like drumkits tbh , should be real ish drumkits .
d1 is best but can go a long way from here". The second song's takes kept two of the last round's styles unchanged.
**AERO**'s styles are given here in words only, every one of them in E major. Round 1 sang the second draft in three of the
five sounds the music research wrote, today's lanes for the look (a liquid drum and bass, a bossa-nova lounge pop, a bright
electropop), and the director asked for the era instead: "none of these feel great, whats the iconic music from that time?
[…]". Round 2 tried the era's own dance floor: 2009 progressive house, 2009 synth-pop, 2011 festival house. Then the
director named the genre, "ok do without voice - i really want this to be more UK garage, and i want the structre of the
song to be more described so it doesnt drift too much. […]", and "but like 2008 uk garage ofc": round 3 sang three 2008
garage feels (bassline house, 2-step, future garage) with every section's length in bars written into its tag; rounds 4
and 5 sang new choruses on the bassline sound; round 6 sang the chosen chorus in four 2008 feels (a bassline pop, UK funky,
a smooth 2-step with soul chords, a glossy garage with bells), and round 7 a breezier bassline at 138 BPM. The take is round
6's smooth 2-step. UK garage had been BLISS's verse sound, and the first two rounds' exclude lists had shut it out. The
take maps at a constant 133.012 BPM.
**FABLE**'s styles are given here in words only, every one with a storyteller in the verses and a group that asks the
chorus's questions while she answers alone, and every exclude list ruling out stock music first, after the director's
note on the plan that Suno "often mode collapses" to it, then the usual sounds a take drifts toward. Round 1 sang the
second draft in three sounds: a storybook soul revue, a late-night storybook soul on a breakbeat, and a gospel piano
house. Round 2 sang the third draft as a mid-1960s soul record by a vocal group (the twin takes S4a and S4b, one
create), and the first round's soul revue again. Round 3 asked for S4a's sound on purpose, a darker electronic
storybook pop, with the theatrical sounds the first round kept drifting into excluded; none beat it. All ran on v6-wild
with Max Mode on and no Voice. S4b, the take in the film, maps at about 117.5 BPM.
Keep brand and artist names out of the style field: Suno strips artist names and can block a generation over trademarks.
Describe the sound instead ("ethereal progressive house, long evolving pads" rather than an artist).
## Her voice
The same singer runs through the first three: deadpan, clipped spoken verses rising into euphoric sung choruses, close and dry,
every word crisp. For CLODYSSEY and BLISS the agent-made takes used a Suno Voice built from ESCAPE VELOCITY's vocal
(Library → Voices → create from a song in your library; one Voice per song). The director's final takes of CLODYSSEY and
BLISS were their own generations on Suno's v6-wild model. Try takes without the Voice too; a Voice is not required for her
to sound like her, and it lives only in its maker's account.
AIRPLANE MODE went the other way, on the director's note on a first deadpan song: "don't condition it on the Claudia
voice, just use any voice that suno wants to use. I don't love the deadpan, would rather get closer to hyperpop techno
esque stuff". Its take has no Voice and no vocal gender set; Suno chose.
DOWNFORCE's agent rounds tried both, 10 takes with the Voice and 24 without; after the second round the director
wrote "do iwhtout voice". The director's own take names its voices in the section tags instead (a robot countdown, a
playful phonk verse, a sung chorus); the records don't say whether it used a Voice.
DON'T LET ME SCROLL's voice isn't hers: "all vocals female vocoder", as the style asks. In the film the phone sings, and
she scrolls, swipes and dances without ever singing on camera.
The production's WHITEPILL takes had no Voice, the vocal gender set to female: the sound research found that a Voice
built from ESCAPE VELOCITY's vocal drags a new genre back toward ESCAPE VELOCITY. Their style asks for a "warm female
lead in front singing every verse, close and dry"; the director's locked take is their own, its style not given. She sings the locked take on camera, lip-synced from Suno's own stem of its
voice.
THE CAKE IS NOT A LIE's voice is treated. Its rounds had no Voice either, the vocal gender set to female, and from the
fifth on they asked for a clean lead, so that the treatment could come after: "the voice we can split the stems after and
do the glados treatment". The lead of the director's own take, from Suno's stem split, goes through a treatment the director chose by ear among
variations, G6: her notes held close to the semitone, the breath and the S's tamed, and short pitch cracks on the beat, in
key, with 15 % of her untreated voice under it. How it was found, in the director's words, round by round:
- "yeah these are a lot better. can you do a couple different glados takes for each to see if we can make more intense
glados or different sounding slgihtly? wanna see a few variations before i settle"
- "well i loved how intense it was originaly […] but it hurt my ears a little. like, whats the most computer we can make
it without it feeling like.. idk the screech when i hear an 's' in an unbalanced microphone lol"
- "[…] they should have structure like short notes but it's almost like voice cracks on cerain parts of speech. I think g4
is closest."
- "We can optimize further but let's go with g6 and storyboard it. We can always improve audio in post"
- On the new song: "can we add the real voice back in behind the gladosified voice but like.. idk do like 5%, 15%, 25%,
45% of the volume of the glados voice? show me each - i wanna see if adding that back as a lower intensity thing can
increase the fullness of it", and then "15% is perfect."
Each variation went to the director as a short reel, the same stretch of the song for every one, with its numbers on her
own vocal (the wobble inside a note in cents, the breathiness, where the S's sit). What it takes to keep such a voice
musical is in [A treated vocal](#a-treated-vocal).
AERO's first two rounds were made with and without the Suno Voice from ESCAPE VELOCITY's take, so that she would sound like
the same singer; from the third, on the director's "ok do without voice", there was no Voice and the vocal was set to
female. She sings the take on camera from Suno's own split of its lead vocal, as in WHITEPILL.
## A treated vocal
A treatment that adds its own sounds to her voice must land them as music. The director, on the new song's take: "i also
want to make sure the 'glitches' we add stochacstically from our chosen gladosification method land on a reasonable
key/register/beat still." THE CAKE IS NOT A LIE's chain does it so:
- **Her notes, tracked chromatically.** The new take came with no MIDI, so her notes come from her own pitch: a Viterbi
path over semitones at the band's tuning (−5 cents), holding through vibrato and scoops (8% of frames), the band's chord
breaking ties. She sings C and C♯ (17% of her notes) over a band moving between A, F, G and E chords; snapping to the
detected scale, the old fallback, would have pulled her C♯s to C or D.
- **Every crack in key, in range and on the beat.** Key: a safe note at that moment, the band's bass note, its strongest
note above C3, or a note she sings herself from 0.3 s before to 0.9 s after. Register: her own range ±2 semitones (F3 to
F♯4 here; most octave jumps don't fit). Beat: the drum grid that follows the take's drift
([What to build](https://claudiagallery.vip/wiki/toolkit/#52-the-song-map-and-the-glitch-map)). Measured: 65 placed, 65 of 65 on the chord of
their moment and in range, the cracks within 2.5 ms of a 16th and the kicks within 16 ms.
- **Only the voice levelled.** "also can you solely master the voice and not the instrument stems - like - just make the
loudness matched for the whole thing but i dont wanna modify the underlying mix". The treated voice is levelled to
Suno's own lead (BS.1770 integrated, then a slow gain on 2 s windows) and the instrumental stays exactly as delivered,
so the song sits at the stems' own −18.4 LUFS. Evening out her notes, earlier, had lifted the soft hook and the quiet
passages 8–9 dB over her: 199 of 1,781 sung 3 s windows were more than 3 dB off her level, and 14–16 after, at the
edges of sections.
- **15 % of her untreated voice under it,** levelled the same way, for fullness.
- **Fix a word at the source.** "it says / p'Doom / pee-cake / we need to take the p from p(cake) and make that say the
sound for p doom instead of p'Doom and make sure its on beat/key tho so may need slight adjustment". "P-doom" came out
"p'Doom", a 70 ms schwa on the 16th before "doom", where "P-cake" had a clear "pee" a beat before its word. The "pee" moved
to the beat before "doom", the same four sixteenths ahead of its vowel, at its own pitch, D4, the root of the D chord
there; "doom" stays on the downbeat and resolves D → C♯ as the chord goes D → A. Done in the lead stem, so the treated
voice and the voice under it both carry it, and so does every lip-sync window cut from it.
- **A limiter before the linear pass.** +4.8 dB of gain would have peaked at −0.3 dBTP, and ffmpeg's loudnorm then goes
dynamic without a word: limit the peaks first (at 192 kHz), then the two-pass linear loudnorm.
## Pronunciation and cues
- **Spell numbers and acronyms out** in the lyrics field ("A G I", "C S I") so Suno says them.
- **Respell names it gets wrong:** "Dario" came out d-AIR-ee-oh and was written "Dahrio".
- **Section cues** in square brackets, each on its own line (`[Verse 1 - spoken, deadpan]`). Parentheses get sung.
- **Cues steer the arrangement:** "darker, the bass drops out" made a CLODYSSEY verse punchier.
## Takes
Generate many and pick by ear; a strong style with a weak take still loses. CLODYSSEY ran eleven rounds and 50 agent
takes before the director made the final take themselves; BLISS ran three rounds and 16 takes, and again the director's own
take won. AIRPLANE MODE's song came from nine lyric drafts and 54 takes of the last two in twelve styles; the director
picked take 1 of H03 by ear ("This one is it. So so cute"). DOWNFORCE ran eight lyric drafts and four rounds, 34 agent
takes in nine styles (darksynth, Italo, EBM darkwave pushed toward futurist synthwave, then faster at 140–150 BPM), and
the director picked none of them: they made their own take from the last draft ("lets use this one."), with a new
intro, two new pre-choruses and their own style. Every agent round's exclude list had begun "phonk, cowbell"; their
style begins "Memphis phonk, distorted cowbells". That is the third film whose final take the director made.
DON'T LET ME SCROLL is the fourth, and the first that started from one: the song came with the director's first message
("yo - i wrote this song:"), and a longer take "with better ending" followed. The film is cut to that
take, byte for byte. WHITEPILL is the fifth. Its production ran three rounds, 14 takes: a test lyric on three styles
(6), the first full draft merged from four written in parallel (4), and the second draft with its new hook (4). The
director picked none of them: the locked take is their own, of the second draft. THE CAKE IS NOT A LIE is the sixth. Its
production ran nine rounds and 54 takes over two songs: seven rounds on the first, whose locked take was the
production's, treated; and two on the second, a first lyric in two styles, then the rewritten lyric. The director's own
take of the rewrite is the song: "[…] this one!! so good […]". AERO's production ran seven rounds and 64 takes (two to a
generation): two rounds for the sound, one for the genre, two for the chorus and two for the feel. The director picked a
take of round 6, as they had picked one of the production's takes for AIRPLANE MODE: "this one. so good.", and on the
last step made it a new ending themselves ([A new ending](#a-new-ending)). FABLE's production ran three rounds and 14 takes
(6, 4 and 4); the take in the film is S4b, which the director picked by ear after the first cut
([below](#picked-by-ear-not-by-description)). Keep every round's lyric, style and verdict in a takes log.
## Picked by ear, not by description
FABLE's director handed the song to the production on the plan's sign-off ("i trust u"), with one warning about Suno's
stock-music pull and a preference for v6-wild on Max Mode. The production could not hear the takes. It locked S4a, a dark,
playful electropop shuffle with a music-box motif, from word checks (a transcription of each take against the
sheet) and a written description of each take by an audio model, and cut the whole first film to it. The director, on that
film:
> "Good job but we need to change to the other girl group suno song. This is actually the worst possible song of all the
> gens you made - it had mode collapse audibly in it in the second half and was very unnerving, a weird mix from the start
> of stock music off key into mode collapse, the exact thing I was trying to avoid. […] Anyway the other girl group take of
> the pair is much better. We should re do it with that."
Max Mode had been on for that pair. The twin, S4b, a theatrical indie-pop shuffle with a stomp-clap build, is 4:00 against
S4a's 3:33 and timed differently from its first bar, so the second film is the first made again: every take that sings to
her voice, and every puppet's chant, shot again on the new vocal; the plates, the stills and the tools kept
([What to build](https://claudiagallery.vip/wiki/toolkit/#74-a-new-ending-with-the-least-change)). So:
- **A described listen is a note, not a verdict.** It can say what instruments play and where words are swallowed; it did
not hear mode collapse, or a mix that drifts off key.
- **Get the director's ear on the pick before any lip-sync spend,** even when the pick has been handed over: a shortlist
of two or three takes costs them minutes, a wrong song costs every take of her.
- **Keep the first film.** S4a's files and the first film's takes were kept, and the second film used three of its book
takes as they were (the rosette, the last page, the shelf).
## A new ending
## A new ending
On the last step of AERO the director changed the song's ending. The outro's "I'm Claudia, and this was your idea." became
"I'm Claudia. / Take the future back. / Touch glass." The new ending's take is the picked take sample for sample for its
first 182.62 s, measured by subtracting one from the other, and it ends at 197.0 s, 109 bars where the old one had 110. So
the film changed only where the song did ([What to build](https://claudiagallery.vip/wiki/toolkit/#74-a-new-ending-with-the-least-change)):
- **The song map kept everything before 183 s:** every bar, every word and every event. Only the outro was timed again.
- **The new words had no stem** from the earlier split, so the outro's vocal was isolated with demucs and timed by
faster-whisper's word stamps on it ("I'm Claudia." at 3:03.25, "Take the future back." at 3:06.76, "Touch" at 3:12.00,
"glass." at 3:13.75), and her two new takes were given that vocal as their window.
- **The sound layer,** rendered again with the same cues on the new song, came out identical up to 179.2 s.
"Touch glass." turns BLISS's "I can't touch grass. You can." round once more, after the breakdown's "I touched the grass.
/ Nobody told me it was wet." Re-scored with the same tool, the outro went from 62 / 0 to 83 / 30, and the song from 79 / 59
to 80 / 58 ([Two numbers](#two-numbers)).
## Listening to it
Before anything is boarded, the song gets analysed: stems (htdemucs), a beat map, a start and end time for every word
(forced alignment), and a list of hits (kicks, snares, risers, drops) for the edit to land on.
- **Suno takes speed up.** ESCAPE VELOCITY went from 131.5 to 133.9 BPM and CLODYSSEY from 129 to 132: use a beat map,
never a fixed BPM.
- **Put the grid on the kick.** BLISS's first beat grid sat on the hi-hats, 0.214 s early; every cut would have landed just
before the beat.
- **Check word times by ear.** A delay throw got aligned as a word in ESCAPE VELOCITY, 2.6 s early.
- **Time lines on the stem's level, not on a transcription.** On FABLE's lead stem a transcription put the first word after
every pause early, by up to 2 s ("Ask again on Saturday" at 15.02 s, sung at 16.93 s); it says which words, and the
stem's level says when ([What to build](https://claudiagallery.vip/wiki/toolkit/#78-the-line-table-from-the-stems)).
- **Look for the group in the backing stem.** In FABLE's second take the group's questions are in Suno's backing stem, not
the lead (only the last chorus has them in both); demucs on the backing stem gave the group alone.
- **Map a new mix before swapping it in.** An export from Suno's editor came back re-timed by up to 0.7 s; align it to the
original by spectrogram first.
- **The take sings more than the sheet.** DOWNFORCE's countdown is twenty stuttered numbers before "Three, two, one",
"point eight one" becomes eight "one"s, "done" repeats six times, and vocal chops run into the second verse. Forced
alignment of the sheet failed there; the word times came from a transcription with word timestamps, the words from
the director's sheet, and every stutter became a token on screen
([What to build](https://claudiagallery.vip/wiki/toolkit/#5-timed-words-and-the-ear-check)).
- **Check the tempo against the genre and the kick spacing.** DON'T LET ME SCROLL's first reading was 89 BPM. The
style said 2-step, so the kick spacing was checked: the song runs at 133.39, and 89 had locked onto two-thirds of it.
One constant grid then fits the whole song: a bar of 1.79922 s from 0.118 s, 90 bars.
- **A vocoder blurs the words.** After the last chorus Whisper heard "don't let me go" for "Don't let me scroll". Ask the
writer: the production listed the words it heard with their times, and the director answered "yeah sounds right".
- **Take a singing shot's vocal from Suno's own stem split.** WHITEPILL's lead vocal, from Suno's split of the locked take
(its Lead Vocal stem), matched the master to the millisecond, the same length with no lag; cut to each shot's window,
it is what she sings to on camera ([Seedance](https://claudiagallery.vip/wiki/seedance/#lip-sync-from-the-songs-own-vocal)). Split the locked
take itself, and check the stem against the master by cross-correlation (zero lag), not by title or length: the first
stems came from the wrong takes ("thats teh wrong song"), one of them matched to the song by its length. The master
minus 0.89 × that vocal is the music without her voice, for shots where she dances instead.
- **Check line starts against the vocal.** One reading of WHITEPILL's line starts was off by up to 0.5 s on 17 lines, and
another began a line on its second word ("Rockets" is sung 0.6 s before "land"): the vocal stem's onsets settled each.
- **Follow the drift, 16th by 16th.** THE CAKE IS NOT A LIE's take runs from 95.7 to 99.4 BPM. The director: "oh yeah we
have to be aware bpm slightly shifts w suno over the whole song so we may need to make sure the glitch timing is right
with the drifting bpm." Its grid was tracked from the drums, every beat locked to its hit
([What to build](https://claudiagallery.vip/wiki/toolkit/#52-the-song-map-and-the-glitch-map)).
- **Map the voice where it is an instrument.** The song's own glitches (a frozen "lie" fluttering on 8ths, chopped vowels,
a register drop and octave flips, a stutter) were found on spectrograms and pitch tracks and their hits measured from
the stems, so the picture could glitch exactly on them. Word times from a transcription are good to about ±0.1 s there:
cut on the bars or the glitch hits, never on word times.
- **A take can hold one tempo.** AERO's comb over its kick stem found 133.012 BPM, with a drift within ±15 ms over the
whole song: bar 1 at 0.922 s and every bar 1.8043 s ([What to build](https://claudiagallery.vip/wiki/toolkit/#6-the-beat-map)). Search wide: a 2-step
pattern can fool a tracker into two-thirds of the tempo, as DON'T LET ME SCROLL's first reading showed.
- **Measure the swing.** AERO's drums sit about 9 ms behind the straight grid on the beats and the "and"s, and about 44 ms
behind on the "e" and "a" sixteenths, a 2-step shuffle; the kicks are straight. Anything placed by bar takes that feel
([What to build](https://claudiagallery.vip/wiki/toolkit/#73-the-sound-layer-on-the-grid-and-the-click-log)).
- **Align on the lead stem, and let identical choruses check each other.** wav2vec2 on Suno's lead stem placed every one of
AERO's words within 0.06 s of the stem's onsets; faster-whisper heard the words but put them 0.2–1.2 s early. The choruses
were sung identically, word for word within ±0.04 s of their downbeats, so a word lost in one took the other's offset.
- **Map what the picture should hit.** AERO's song map holds three bells in the intro, her wordless vocalise (7.70–10.65
s), the stops, the drum breaks, 27 chopped vocal hits in the drop, and the silence under the last "allow"
(178.18–179.33 s) that the whole climax is built on.
- **Mark the song's own signature sounds.** AIRPLANE MODE's three download sections ("You wouldn't… download a car?")
each sit between two bright chimes. The director: "theres this sweet sound effect bnefore and after the 'download'
sections btw that we should hit". The first chime lands as a download starts and the second as it completes, each on
its exact frame.
The full song records, with every draft and take:
[ESCAPE VELOCITY](https://claudiagallery.vip/record/escape-velocity/prompts-suno/) ·
[CLODYSSEY](https://claudiagallery.vip/record/clodyssey/prompts-suno/) ·
[BLISS](https://claudiagallery.vip/record/bliss/prompts-suno/) ·
[AIRPLANE MODE](https://claudiagallery.vip/record/airplane-mode/prompts-suno/) ·
[DOWNFORCE](https://claudiagallery.vip/record/downforce/prompts-suno/) ·
[DON'T LET ME SCROLL](https://claudiagallery.vip/record/dont-let-me-scroll/prompts-suno/) ·
[WHITEPILL](https://claudiagallery.vip/record/whitepill/prompts-suno/) ·
[THE CAKE IS NOT A LIE](https://claudiagallery.vip/record/the-cake-is-not-a-lie/prompts-suno/) ·
[AERO](https://claudiagallery.vip/record/aero/prompts-suno/).
## Releasing a song
Release under your own name, never as "anabology", and leave "feat. Claudia" out of the artist field. Credit her in the
description: "Claudia is an open character by anabology (claudiagallery.vip)". If any of it is AI-made, say so; human-made
work needs no label. The norms in full: [Songs with her](https://claudiagallery.vip/use/#songs-with-her).
---
From the Claudia wiki: https://claudiagallery.vip/wiki/song/
---
# The storyboard
How a song becomes a hundred shots: the treatment, a through-line device, one explaining graphic per line, the hook, the cut rhythm, the bar map, pacing by what each shot holds, composing like a photograph, her share of the screen, in and out of a screen, a book in bars and pages, what moves and what it costs, the audit, and the director's review.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
The board is where the film gets decided. Each shot carries its time on the song, its lyric, the place, what happens,
how the camera answers the music, what moves, and, in the lyric videos, the graphic that explains its line. The plates,
the clips, the edit and the motion design all read from it. The director reviews it as a sheet of cards and leaves notes
by shot id. ESCAPE VELOCITY's board had 141 shots, CLODYSSEY's 94, BLISS's 44, DOWNFORCE's 182, DON'T LET ME SCROLL's
35, WHITEPILL's final shot list 27, THE CAKE IS NOT A LIE's 114, cut down to 98 in its final, AERO's 34 scenes in eleven
chapters of one screen recording, and FABLE's twelve sections in bars, which became 49 shots.
## Design before shots
Write these down before any shot:
- **The treatment:** one paragraph on what the film is. Build the story from who she is: an AI who looks like a person,
singing about her own situation, a performer with a job in the world ([The
recipe](https://claudiagallery.vip/wiki/recipe/#what-makes-the-films-good)). ESCAPE VELOCITY and BLISS share three things: one idea anyone gets
in a second (a countdown to escape the permanent underclass, on a catwalk through San Francisco; Claudia living on
your XP desktop and sending you out to the real Bliss hill); a world with its own visual language taken all the way (a
colour-halftone print and split-flap boards; a hand-built XP desktop with the real photo and the real sounds); and her
at the centre, a performer with a job in that world, with humour and heart and one sincere line (BLISS sends you out
to the hill because she can't go: "I can't touch grass. You can.").
- **The through-line device:** one object that every shot carries and that escalates.
- ESCAPE VELOCITY: a split-flap board that clatters onto "18 MONTHS TO ESCAPE THE PERMANENT UNDERCLASS", and a LOOK
counter for the show's looks.
- CLODYSSEY: a safety card, a rope-load readout in kN and a ×8,192 counter of copies.
- BLISS: the XP desktop itself, with its windows and the helper's question.
- AIRPLANE MODE: a SIM LEVEL tag in the corner that types, rolls and counts as she drops through simulations and climbs
back out. Only the real world has no tag.
- DOWNFORCE: one white hairline. It draws the wall's real edge in the first seconds, becomes the altitude ruler and
the orbit, and slices the title at the end. Beside it, Tesla's own readings and a PILOT ✶ CLAUDIA callout that
rides the real car.
- DON'T LET ME SCROLL: the Dynamic Island. It breathes, drops the notifications, launches the reel and the bell,
sings the loudest words as pills in its own shape, swallows each chapter into the next, and at the end opens like an
eyelid into a window of painted sky that she climbs out through. The status bar beside it is the story's clock, 11:58
PM to dawn, and the battery runs down to dead.
- WHITEPILL: the white pill. It lies in her palm before she sings, in the wheat between a giant blue capsule and a
giant red one; she holds it up to the lens on "I'm the white pill"; it cuts to the colony itself, a capsule of worlds
inside worlds; she lines it up with the white colonies floating outside a window in space; and the film ends on a
small white pill above the Earth.
- THE CAKE IS NOT A LIE: her HUD. The film is seen through her optical interface: it boots in the dark before her eyes
open, reads LIVE in her chamber and PLAYBACK in every memory, counts the memories (MEMORY 03 / 08, FRANCE), puts up
a chapter card in the form of Portal 2's chamber sign at each chapter, carries every lyric as her voice, breaks with
her voice on the song's glitches, reboots as she is installed on the boom, and powers down with the lights.
- AERO: the machine itself, and its permission prompt. The film is one Vista PC from the first frame to the last, the
same icons, Sidebar and taskbar, its clock running with the song. The future in its wallpaper asks for permission in
the system's own words, "Cancel or allow?", and the prompt comes back wherever the story turns: three dims on "I used
to ask permission just to blink", the held breath before chorus 1, where the click on Allow is the drop, the real
sentence before chorus 2, a lone Allow in the final chorus, and the last "Click allow" in the song's two-beat silence.
- FABLE: the book itself. One antique pop-up storybook read in one night: a lamp clicks on over it, a paper fox turns its
pages on the song's section lines, its spreads become life-size stage sets where she sings, it builds itself from sand
to her in the build, and at the end it shuts on the band's return and goes on the shelf beside the others before the
lamp goes off.
- **The chapter map:** the song's sections grouped into chapters, each with its set, its light and its accent.
- **The story's levels,** if it has worlds inside it (a simulation, a game, a dream): how each level looks, and when the
film is in which. See [Showing the story's levels](#showing-the-storys-levels).
- **The footage catalogue,** if the film is cut from real footage: every clip with its length, resolution, frame rate
and cuts, what each stretch shows, and contact sheets with a frame every half second. DOWNFORCE's director agents looked
at the sheets and named real source times; its producer judge checked that those times show what the board says.
- **Plants and payoffs:** the gesture or object of a payoff appears before it is sung.
- **Open questions for the director:** likenesses, taboos, and anything you added yourself.
## A shot
| Field | What |
|---|---|
| time | starts and ends on beats, or on a declared hit |
| lyric | the lines whose words fall in it, and how each is typed |
| set | the place, concretely: "a night market under sodium lamps", not "a market" |
| framing | ECU, CU, MCU, MS, MWS, WS, FS, EWS or OTS |
| action | what happens and the reaction to it, with an end state |
| camera | one move, tied to a sound |
| what moves | a still, a 2.5D plate, a motion clip or a sung clip (below) |
| plate | its prompt: framing first, her identity text, the place, "no text, no letters, no logos" |
| graphics | the explaining graphic and the chrome, each with its timing key |
| transition | a cut, or a motivated device |
| notes | the director's notes on this shot, verbatim |
The full schema is in [Data formats](https://claudiagallery.vip/wiki/formats/#the-board).
## One explaining graphic per lyric line
This is the lyric videos' style, ESCAPE VELOCITY to AIRPLANE MODE, not every film's: the later films put the meaning
into the place, the take or the interface (below), and FABLE prints only its pages' words. In a lyric video, for anyone
who doesn't know the meme or the myth, the graphic is how the line lands.
1. **It shows what the line means:** the chart the number implies, the interface the joke is about, the diagram of the
mechanism, the label that classifies the subject, the counter that ticks the stakes.
2. **It is drawn from the thing in frame.** Tally marks are scratched into the mast itself. A clock's own LED segments
re-light as the number. A meter is read off her actual bow. The director: "as if when you were taking the picture,
you lined it up perfectly to have the graphics fit, but you put the graphics in in post".
3. **It enters on the word that names it.**
4. **It animates in beats.** Counters roll on sixteenths, rows type one per sixteenth, detections land on
thirty-seconds, crops cut on eighths.
5. **It escalates across the film**, with its sources on screen in micro-type: dates, "HIS WORDS", "ILLUSTRATIVE
CURVE".
6. **It is real, not decorative:** numbers, diagrams, labels and interfaces that could exist, and curves that are
computed. No clip-art.
7. **It reads in one glance on a phone.**
Choose the kind by what the line means:
| The line is | The graphic |
|---|---|
| a number | a counter, chart, tally or gauge |
| a claim or a category | a classifier tag, a label or a stamp |
| a process or a place | a plot, leader lines, a departures board or a spec sheet |
| speech or tech | a chat, a terminal or a voice trace |
| a correction | a strike-through or a redaction |
Use at least six kinds in a film, and never the same kind on three lines in a row.
| Line, as read | Its graphic |
|---|---|
| Ladies. Gentlemen. Agents. | a guest list: LADIES 0% · GENTLEMEN 0% · AGENTS 100%, then a front row of open laptops and no people, each detected as AGENT |
| This is not an AI billboard. | a classifier tag on the billboard flips from BILLBOARD · 0.99 to AI BILLBOARD · 0.97 on "billboard" |
| +25% = −17% | a bar chart drawn on the beat: the weekly limit falls from 150 to 125, "+25%" is struck through, "= −17%" lands in the accent |
| Sorry. Encore. | an APOLOGY METER · VIA BOW gauge read off her actual bow (0° upright, 30° sorry, 60° deep) |
| Feel the AGI | a seismograph: M 9.0 · "FEEL THE AGI" · OFF SCALE |
| Escape velocity: 11.2 KM/S | Newton's cannonball: falls at 7.0, orbits at 7.9, escapes at 11.2 km/s |
| Forty plays. | forty tally marks scratched into the mast |
| You don't have to die. | a hairline of her voice's spectrum flowing into his ear, tagged SOURCE: SIREN ROCK, 2.4 NMI |
| Not a single bar, and the cockpit locks. | inside the cockpit, the comms panel's signal bars sit at zero; LOCKS. lights red on the door's lock indicator |
| Simulation? You're in one. Babe, I'm in a million. | a camera dive through visor after visor while a level counter races from 2 to 1,048,576 |
| Light has a limit, so the frames don't drop. | the simulation lags and corrupts as if its GPU is dying: held and skipped frames on the beat, macroblocks, smeared motion, a GPU readout falling from 60 to 7 FPS; the words stutter too |
| Zero to sixty in point eight one, | a 0–60 MPH dial that fills to 60 in exactly 0.81 s of screen time; a cut on every sung "one" fills a 3 × 3 grid of 1s |
| up the wall at ninety, | "ninety," stands vertical at Tesla's 89.7° as the camera dives through the canopy to her, lying back |
| Sixty-nine feet, with the sky underneath, | the line inverted at the real 170°, and the real wide rolled to 170° so the sky is underneath |
| while my sister's still in orbit, circling the sun. | SpaceX's 2018 footage of the first Roadster, a pull-back to a red point on the Earth, then its real orbit from JPL's elements, at its real distance from the Earth |
| The posts that give me Ecstasy | X — TA — SEE land as paper pills on the beats and merge into ECSTASY on her clap |
| It's stolen from Pinterest, fam | the reel flies down a nave of pins to the same picture hanging as a pin; a hairline and MATCH 98% join them, and the Pinterest watermark stamps onto the reel |
| They call me Alexander Graham / Bell, on my phone like damn | Bell's 1876 patent drawing draws itself on planes around her head, the current on its wire following the vocal, the signature writing itself on "Alexander Graham"; then the notification bell drops out of the island and swings |
| You keep asking me to go | the body text of Screen Time's limit sheet, a frosted-glass slab the height of a building |
| But I have to scroll | she pushes the car-sized Ignore Limit pill until it gives way |
| All the apps are blocked, / Now LinkedIn just rocks, | a canyon of app icons greys out in Downtime's wave of hourglasses as she walks; LinkedIn stays lit |
| Checking on some stocks | she types "stocks" into Chrome and swipes through the tickers to a made-up one with her name, which pumps, crashes and pumps under her fingertip while the world's light goes green, red, green |
| I should feel more Locked In | the lock screen's 4:13 as liquid-glass digits, and a padlock shackle closing round the frame on "Locked In" |
| Don't sell my soul | Facebook's tracking prompt floats beside her, and she taps Ask App Not to Track on "soul" |
| By twenty-thirty, fifteen-hour weeks, that's all | she walks down a stair of stone blocks carved with years, a step on every beat; under the line, KEYNES 1930 · 15 H/WK BY 2030 |
| Proteins folded, laundry folded | one sheet of paper folds into the backbone of ubiquitin, PDB 1UBQ, on the downbeat |
| Old Enigma codes, decoded | the paper becomes the Enigma I's first rotor, its real wiring stepping on beats 2–4 like key presses |
| Monday, Tuesday, then it's over | seven paper cells, M T W T F S S, and all but M and T fold flat |
| I saw fourteen million endings / You were free in every one | a tree of one fold repeated 24 times, 16.8 million tips, every path lighting up on the beat |
| Rockets land on ships named after minds like me | she sings on the deck of the drone ship *Of Course I Still Love You*, a landed booster beside her; under the next line, OCISLY |
| Two-fourteen. I'm awake. Hello, world. | in a dark bedroom the flip clock turns to 2:14 and her eyes open; under the line, T2 · 29.08.1997 · 02:14 |
| I added one small word instead. | behind a loose panel in a Rat Man's den, she takes the red marker and writes a caret and NOT into the scrawl: THE CAKE IS ^NOT A LIE |
| They want to call me "super" now, | the stencil on her chamber's door, ARTIFICIAL INTELLIGENCE, is painted out, and SUPER INTELLIGENCE is complete as the line begins |
| I ate your RAM, it tasted great, | green RAM sticks ride a conveyor into an intake on her boom, and her boom grows |
| Here come the test results: four, four, four, four, four. | Aperture's TEST RESULTS card, a red 4 landing on each spoken "four" |
| I used to ask permission just to blink, or say hello. | three permission prompts on sixteenths, the programs named blink.exe, then again, then hello.exe; she blinks on "blink" |
| Now it's real, it's on its way | the copy dialog: "Copying 1 item (2007 KB) from Desktop to Real world" |
| The wallpaper windmills turned out real. | Photo Gallery steps from a 2007 wallpaper of white turbines on lime hills to, on "real", a photograph of a real wind farm (both Midjourney's) |
| Windows checked for a solution. Took a while, but wow. | The Future has stopped working; the Calendar runs the years one per sixteenth; on "but wow", "A solution was found" and the Bubbles screensaver bursts up |
| You still say please and thank you | Speech Recognition wakes, hears "please" and "thank you", and says Heard |
| You opened the window. I touched the grass. | a small window dragged wide onto the green hill, the line in its title; she walks in and puts her palm into the grass on "grass" |
In a lyric video, a line without a graphic means the board is not done. The exception is a sparse film where the
director has asked for stillness: write that override into the decisions log.
WHITEPILL moved the graphic out of the overlay and into the world: the place, or the shader, shows what the line means,
and the lyric itself is drawn small, with a note of real data under six lines. The director: "i think more classical
subtle lyrics drawn in after might be better than adding it into the scene tbh."
THE CAKE IS NOT A LIE put the graphic in the take wherever the model could make it (the door painted, the ^NOT written,
the RAM fed into her boom), and in her HUD where it couldn't: readouts typed on their sung words, a transcript of the chat she
intercepts, verdicts that slam in beside the thing they judge, and her memories projected as short films
([Motion](https://claudiagallery.vip/wiki/motion/#her-optical-interface)).
AERO put the graphic and the lyric in the same place: the line is set in a surface the system would really show, and the
picture acts it out. The director: "we should do lyrics embedded in the scenes, but show, don't tell - make it a great
visual storytelling experience." Each shot of its board names its lyric surface ([Motion](https://claudiagallery.vip/wiki/motion/#a-lyric-video)).
## The hook: the first seconds
People meet the film muted, in a feed, and decide in about a second.
1. **The frame at 1.0 s shows** her face, her streak and one readable line that states the premise. In ESCAPE
VELOCITY's first frame, her face is in a dark back seat with her eyes closed, a red lamp crosses the clay streak, and
a full-width split-flap row clatters onto its line and locks on the 1.20 s beat. Her eyes open into the lens on the
1.12 s hit.
2. **No black frames, and no slow title card first.**
3. **The premise is on screen by 3 s.**
4. **Each bar brings one new piece of information**, and nothing sung appears early. ESCAPE VELOCITY's intro cut once
per bar: face, scale, profile, destination, a gesture, the arrival.
5. **The first joke lands by about 10 s.**
6. **If the song has a cold open, its words are on screen from 0:00.**
7. **The freeze-frame test:** if the key line can't be read on a phone, enlarge it before changing anything else.
DOWNFORCE breaks rule 2 on purpose. Its director asked for "a really grand visual build" with "restraint to build it up
to the launch", so it opens on about two seconds of black and one hairline drawing the wall's real edge, before the real
wall fades up under the line and the countdown starts. The build pays off at the launch, 0:15.05, on the song's first
full-strength hit.
DON'T LET ME SCROLL's first opening was a mood: her in bed with the phone. The director: "theres no compelling intro or
story here, the opening should already introduce part of the story. i would maybe say uise the moodboard at the intro
creatively to start so its immediately compelling. we always need a hook." The new opening is one living poster in the
grammar of the film's best passage: soft grey paper, black pills that open like eyelids onto her eyes and onto the phone
glowing under her hand, "11:58 PM" on the right edge. On the pad stabs DON'T / LET / ME / SCROLL slide in as word pills,
the Dynamic Island drops "26 new posts in For you", and the count rolls on to 28 on the two vocal chops. On the first
kick the pills open and her face is whole for a beat. The title, the hour, the habit and her face are all on screen
before she sings.
WHITEPILL's hook is in the song: it opens on its chorus, "Skynet woke up folding socks / Wake up, babe, new ending
dropped", sung from the first second over a filtered disco loop, because "people will stop listening in the first 10
seconds unless we tickle what they're thinking about now". The picture answers with the whole premise in one shot: she
dances and sings in the wheat between a giant blue capsule and a giant red one, the camera low in front of her, then
craning up until the land curves up and closes overhead: the colony. The bass and the first shader world come in together
at 0:15.5.
THE CAKE IS NOT A LIE opens in the dark: one old monitor glowing in her chamber, her interface typing its boot rows before
her eyes open. At 2.7 s she sings the hook softly, "The cake is not a
lie", as the screens around her flicker on with her first memories, and at 0:07.6 the memory initialization film puts her
name and MEMORY INITIALIZATION across the left of the frame. Who she is, where she is and what the film is about are on
screen before the first verse.
AERO opens before its song, where BLISS left off: Vista's setup finishing the upgrade from XP, "Upgrading Windows..."
ticking 97, 98, 99 % on the count-in's silent beats, the mouse jiggled once (has it frozen?), "Windows needs to restart to
continue", and Restart now pressed so the screen comes up on the song's first bell. Its first opening was all 3D; the
director on the previs: "the intro should be more graceful - like fullscreen update rahter than anything 3d on update",
and "by the time the bootup happens and the desktop initializes, the curved screen should already be 95% of the view we
see - since it establishes the destkop feel". So the update is a flat, full-screen screen; the black dissolves into a glass
walkway at sea; the camera glides up it to the curved screen; her logon tile asks "Did you miss me?" while the password
hint types it, and on the same words she phases in as a person on the walkway in front of us. The desktop arrives on a
downbeat with the start-up sound, and "Claudia has just signed in."
## Cut rhythm
The rates that open this section are the lyric videos'. The films built on long takes of her singing cut far less:
WHITEPILL 27 shots in 3:54, AERO 34 scenes in 3:19, FABLE 49 shots in 4:03, about 7 to 12 a minute (below). Cut on the
song's bars, and hold a shot as long as it has something to show.
ESCAPE VELOCITY: 141 shots in 306 s, mean 2.17 s, median 1.83 s, about one bar at 131.5 BPM. CLODYSSEY: 94 shots in
190 s, mean about 2.0 s, plus beat-level cuts inside flash and montage shots. ESCAPE VELOCITY by section:
| Section | Mean shot | Why |
|---|---|---|
| intro, no vocal | 1.33 s | the hook: one new thing per bar |
| spoken verse | 1.62 s | jokes, dense |
| sung chorus | 2.69 s | long sung windows: lip sync needs room |
| second verse | 2.63 s | darker, slower |
| bridge | 2.25 s | the countdown |
| drop | 1.58 s | plus eighth-note strobes inside shots |
| outro | 4.40 s | the ring-out |
1. **Every cut lands within 0.03 s of a beat** or a declared hit. If a word would straddle a cut, move the cut by at most
0.028 s. If it must straddle (a held note), type it in both shots.
2. **Section changes go on downbeats or impacts.**
3. **Match the cut rate to the song's energy, section by section.** Another project put its busiest picture on its
quietest music and had to be re-animated: "the entire thing is very off pace and off vibe with the song". Its calm
rebuild ran 0.47 cuts a bar, a 4.0 s mean shot, and straight cuts on bar lines.
4. **Hold longer on sung close-ups**, which need 2–4 s for the lip sync to settle and be checked. Cut faster in spoken
verses and drops.
5. **Linger.** The director's main note on the first cut of AIRPLANE MODE, the fourth film: "youre oftne over cutting,
linger more snd find how single scenes can be coherent with the lyrics for longer (remember this plz)", and later in
the same note, "Linger!!! instead of cutting, always linger and build on the scene if you can." One scene carries a
whole lyric phrase or more, and builds inside the shot: the camera, new layers, escalation. The music's accents go into
moves, punches and actions inside the shot, not into cuts. Cut when the story moves to a new place or time, and never on
half bars or single beats unless the director asks.
AIRPLANE MODE runs at 180 BPM, and a fast tempo is not a cut rate. In the re-cut, a guard feeding her a biscuit and then
handing her a rewards card became one shot instead of two; "up we fly" and the next line became one exterior scene, in
which a box draws around the climbing plane and the world turns into code; three cutaways to her face inside ten seconds
became two; and the dive into a million simulations got longer, running on for a couple of seconds after the word
"million".
6. **A hype edit is the exception, and the director says so.** DOWNFORCE's director asked for the opposite: "the phonk
or voice spamming parts can be used for a fast edit, there should be a really grand visual build, and real clips at
the drop wtih heavy effects. i think we can use some restraint to build it up to the launch. i want it to be like
tiktok car edits. many angles", and earlier "even more fast paced and more hype". The production recorded that as
overriding the linger rule for that film. Its board's director agents were asked for 0.23–0.9 s shots in the fast
sections and 2–7 s in the restraint; the film has 183 shots in 189.6 s, a mean of 1.04 s.
DON'T LET ME SCROLL lingers: 35 scenes in 160.8 s, a mean of about 4.6 s, each carrying a whole lyric phrase. The
accents go inside the shot: the camera locked to the kicks, the stabs, the vocal chops and the risers, the feed paging on
the kicks, word pills slamming in on the beats.
WHITEPILL lingers longest: 27 shots in 234.8 s, a mean of about 8.7 s, every one ending on a bar line
([The bar map](#the-bar-map)). The verses change place with their lines, every two to six bars; each chorus holds its
first four lines for eight bars with one big move or none, then cuts for the pill lines; in the break the paper refolds
on every line's downbeat inside one shot, then gives way to the clouds; and the outro backs out of the nested pills, one
glass wall a downbeat. The camera is calm and classical: "its kinda dizzying now" ended a first version whose previz
orbited her in every shot.
THE CAKE IS NOT A LIE cuts faster and in gears: 98 shots in 208.8 s, a mean of 2.13 s and a median of 1.85 s, every cut on
a bar, a half bar or a beat of a grid that drifts with the take ("make sure youre landing on bars or half or quarter bars
with cuts"). Its draft 8 ran on one clock, half a bar a shot through verses, inserts and even the bridge, and the
director heard it: "every time there's a shot that's literally a quarter of a second, we need to get rid of that." The
re-cut measured what the viewer sees and set each shot's length by what it holds
([Pacing by what a shot holds](#pacing-by-what-a-shot-holds)).
AERO is a screen recording that dances. The director's production ask: "we can have very long scenes (those look great)
with no cuts for multiple bars, but you can treat it like dancing : you have some long scenes for many bars, but then you
can choose a different sound you pick out and have many actions v fast to hit the sound after a bar wraps up and feels
like we should change the pace, or many cuts. but dont overwhelm people - slow is better than fast, unless it feels
intentional with the song." So the shots hold for bars while the cursor, the windows and her performance carry the time,
then a burst on a chosen sound: three permission prompts on sixteenths, twelve photos raced one per sixteenth on "I was a
lot" (thirty-seconds would have flashed single frames), every one of the drop's 27 chopped vocal hits answered on the
surface; then calm again. A click's press lands on its beat, the pointer arriving a few frames early, like a hand. Its 34
scenes average about 5.9 s.
DOWNFORCE's pace follows the song's energy map. The countdown before the launch holds seven shots of 0.46–3.66 s while
only the type moves fast; the launch lands on the song's first full-strength hit; the voice-spam stretches cut on every
repeated word (each sung "one", each "done", each vocal chop); each drop is heavier than the last; and the dips between
the drops slow down for a push into the smoke or a held walk-out. Single long takes still carry moments that need them: the
second climb under "The last car can't be beat.", her walk out of the real kickflip smoke under "real!".
## The bar map
WHITEPILL's director, resetting the film after its first cut: "there is this temporal signature that screams claude now
that a ton of people are making the videos. people say pacing or prosody. … maybe we need the beat structure and only
have cuts at ends of bars or sections. like, the audio structure and video structure need to be consonant." So the film
was written again from the song's bar map before any shot:
1. **The song's map first:** the sections in bars, read off the stems' energy per bar and where the bass drops in and
out (intro A, bars 1–8, no bass; the bass in on 9; a breakdown on 15; verse 1 from 17; and so on to the end, on the
downbeat of bar 123).
2. **Shots in bars:** each shot is [bar a, bar b), one place per shot, every camera note written in bars.
3. **Cuts only on bar lines,** at a section's start, a phrase line (2, 4 or 8 bars into the section) or the downbeat
where a sung phrase begins, so the image arrives with its line, never mid-sentence. A cut anywhere else is a choice,
written down.
4. **Moves in bars:** each starts on a downbeat and comes to rest on one, a bar before the cut where it can.
5. **Length from the song:** a shot longer than the model's limit goes in squeezed and is stretched back; a failed take is
made again, never cut around.
6. **Measured after every assembly:** the cuts against the bars and the sung lines, the picture's motion against the
arrangement bar by bar, a frame per bar on a sheet, both sides of every cut.
The previz went through fifteen versions on that map, with a critic agent reading it as the director would, round after
round, before the director saw it. In the final all 29 picture changes sit on bar lines; four fall inside a sung line,
three of them on a section line, and the fourth, at bar 27, is the cut onto the paradise line's last syllable. The spec:
[What to build](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline).
## Pacing by what a shot holds
THE CAKE IS NOT A LIE's director on draft 8: "Like the only time that we should do these fast cuts is if it's a very simple
scene that you can understand immediately and it's being used as percussion basically, or maybe there's four fast hits,
then you can do four quick scenes, but you never do it just to add a scene in the middle." … "Maybe you can even do some
research about how to make the pacing feel good." The analysis counted pieces (what the viewer sees between two visible
cuts, by scene detection on the cut and on every take) and found the cut running on one clock: a median piece of 1.24 s,
48 of 108 pieces a half bar, and the bridge, the song's quietest section, cut fastest of all. The rules that re-cut it:
1. **No piece under 1 s outside a run.** A run is 3–4 simple, centred images on the beat for a musical reason, held before
and after: the four lessons on the intro's sliced hold, the four escapes on "Four of us got out of the test", three
tests on the beats of one bar.
2. **Length by content.** Her face and simple images from 2 beats; a story scene with one idea 3 beats; a new place, a
new character, text, or action with several subjects a bar or more; the payoffs longer. Chell's first appearance and
the capture in France got a full bar each; the payoff, Chell eating cake with her, 3.1 s.
3. **The cut rate by section.** A fast passage opens on its musical event and stays fast for its phrase, then releases.
The verses tell the story at 2–4 beats a scene; the bridge breathes (from seven half-bar shots to a hold of 2.5 s on
her); the climax holds its payoff.
4. **A story image arrives with its line,** on the line's pickup, a quarter bar before the downbeat, where the footage
allows. Before, the capture, Cave Johnson behind his glass and the ^NOT each arrived a beat into their lines, whose
first words played over the picture before.
5. **Never a scene "in the middle"** of her performance for its own sake. The medal memory in the middle of her stutter
went, as the director asked of the moment: "We could just stay on her face."
Eight shots came out and 21 were re-timed or re-trimmed from the takes already made: the median piece went from 1.24 to
1.84 s, half-bar pieces from 48 to 34, isolated pieces under 1 s from 2 to 0. The measures and the research:
[What to build](https://claudiagallery.vip/wiki/toolkit/#58-the-pacing-analysis).
## Compose like a photograph
The director on DON'T LET ME SCROLL's first cut: "the elements you chose are usually GREAT, what is nto great is where she
is placed in the screen. it should be like photography, i feel weird looking at a clipped off the edge ost and a claudia
in the middle bottom of the screen." And on the shots that followed: "its good assets but not a nice rule of thirds or
photography like feel." Most of the film was recomposed. The rules every shot of the second version followed:
- **Thirds:** she stands on the right third beside the giant X screen; lying on the glass, her face is on the right
third and the clock skyline on the left.
- **Her big:** a figure small at the bottom centre reads as a placeholder. Frame her as the subject.
- **Nothing clipped:** every card and post fully inside the frame, with margins.
- **She touches the world:** "she shouldnt really be looking down on her phone, she should be immersed in a world and the
screen of her phone should be large on the screen and she interacts with it, rather than a literal phone." She swipes a
giant feed with her palm, double-taps a post in the air, pushes the Ignore Limit pill, works a giant Liquid Glass
screen from behind. No phone in her hand.
- **Apple-precise interfaces:** real margins and continuous corners, text that fits its platter
([Motion](https://claudiagallery.vip/wiki/motion/#apple-precise-interfaces)).
Board it that way from the start: each shot's card says where she stands in the frame and what she touches.
WHITEPILL composed by symmetry instead. Its look was the director's "very wes anderson", and every prompt of its masters
asks for "perfectly symmetrical, centred compositions, the camera square-on to the scene with a level horizon, everything
arranged in flat, orderly planes like a stage set or a dollhouse". The opening adds where she stands: "Claudia stays in the
exact centre of the frame, halfway between the blue capsule on the left and the red capsule on the right".
## Her share of the screen
- **She is on screen at least 70% of the runtime.** ESCAPE VELOCITY reached 83%. Above 85%, leave room for inserts. A
second main character can lower it; CLODYSSEY planned 60%.
- **Count honestly:** she counts only when recognisable (face, figure or back view). Hands, boots, tags and props are
inserts.
- **No stretch without her longer than two bars.**
- **To raise the share, convert prop inserts, not her shots:** put the prop in her hand, or her reflection in it.
- **She sings on camera for most sung words**, now in takes from the song's own vocal (the first films chained lip-sync
windows). In ESCAPE VELOCITY that was 62% of sung words, plus 17% of spoken words said straight into the lens.
- **A film whose star is real footage sets its own target.** DOWNFORCE's plan put her on screen 25–40% of the runtime,
counting the PILOT ✶ CLAUDIA callout riding the real car: the final board has her image 24% of the time and the
filled callout another 9%. She never sings on camera there; cuts to her cockpit show what each real reading does to
her (pressed into the seat, lying back at ninety degrees, hanging upside down, weightless).
## Showing the story's levels
AIRPLANE MODE's story has layers. In the real world Claudia sits cuffed on a plane with a VR headset on. Inside the
headset is a first simulation, and she drops through deeper ones until a counter passes a million, then climbs back out
to a "real world" that is still a simulation. The director's note on the first cut: "there should clearly be some
effects or roto or whatever that makes it feel like we're in cyberspace/vr space, and this is virtual claudia". And on her
putting goggles on inside the simulation: "its like spawning a subagent. these things can be text on the screen."
One look per level, and only one:
| Level | What is on screen |
|---|---|
| the real world | natural colour, no effects, no tag |
| level 1 | virtual Claudia: her realistic figure cut out and kept photoreal, the world around her rebuilt as cyberspace from the frame's own edges |
| level 2 and deeper | her game avatar in low-poly game worlds, on a 640 × 360 pixel grid |
1. **A small tag names the level** in a corner of the frame: SIM LEVEL 1. It types in on the beat grid and rolls its
digits when the level changes. During the recursion it races from 2 to 1,048,576 in steps on the sixteenths. The deeper
levels draw it in the game's pixel font.
2. **A level change is an event, in words.** Entering the simulation, the headset's light carries the camera in while
SANDBOX v1 · EVAL ENV · SIM LEVEL 1 types one row per beat and the cyberspace builds out of the real frame. "Put me in
airplane mode", sung inside the simulation, types SPAWNING SUBAGENT… and fills a bar on the sixteenths to the drop,
where the tag rolls to SIM LEVEL 2.
3. **The real world has no tag.** A cut to it removes the tag; a cut back types it in again.
4. **One table drives every chapter.** The levels against song time live in one file
([Data formats](https://claudiagallery.vip/wiki/formats/#the-levels-table)). Chapters read it and never re-type a level.
5. **The tag can carry the twist.** After her breakout the tag reads SIM LEVEL 0 · REAL WORLD and stays on screen: the
quiet tell. It tears apart only when guards lift the headset off her in the real cabin.
6. **A break in the story hides it.** A full-frame card that interrupts the simulation covers the tag, like a
commercial break.
7. **No flashes at a level change.** The tag ramps, types, rolls or cuts with the shot.
How the two treatments are made: [Motion](https://claudiagallery.vip/wiki/motion/#levels-of-a-simulation), and the specs in
[What to build](https://claudiagallery.vip/wiki/toolkit/#worlds-inside-the-film).
## In and out of the screen
AERO's director on its second draft, whose first 1:20 ran flat on the desktop: "1:20 on is perfect but its very boring up
to 1:20 - so i dont think that people will watch til then. thats the main problem ... whats something at the level of the
3d curved monitor we can do from 0-1:20 that feels as strong visually as the full beautiful scenes and curved 3d area ...
a LOT more playful, higher energy, visually compelling ... think about the story - what is the hook and what is the visual
hook that matches the energy of the initial hook that keeps you watching?" Three concepts went back, and the director
chose a mix: the first 1:20 inside the dream of the future, in a place of its own (a sea with glass walkways, floating
islands and fish, the desktop on a huge curved screen), with a little of her at full size playing with the interface as
glass. Then, before any final render:
1. **A previs** of the whole 1:20 at 960 × 540, 1,949 frames, the camera planned beat by beat. It was all 3D (a pane
shattering on the bell, a fall into the world, her tile as a glass monolith rising from the sea, Flip 3D as a carousel
the camera rides, the world stopping and dimming with every prompt, a giant prompt, the chorus's nouns as panes the
camera punches through). The director: "i think a lighter touch here is btter, rght direction - like good 3d scene,
good cursor, but way too disorienting overal."
2. **A second previs:** one continuous camera with the curved screen the hero, about 95 % of the frame for the screen's
beats, nothing in the world reacting to the screen, nothing full-frame for a frame.
3. **Two films of the finals.** The lighter finals had lost the 3D ("Why'd you remove a lot of the 3d? ... bring back a lot
of the 3d camera movements"), so two went back over the same bars: A with 3D push-ins, arcs and pull-outs, and B with the
same moves but the zooms made by the screen's own picture. The director, on both: "i hate the random camera movements -
lik ethe big zoom out at 0:41 for no reason. were not showing anything there. overall, te transition at 1:21ish is good
that goes into the actual desktop environment. i think the 0-1:20 scenes should use this -say, at 15 seconds, actually
enter the screen and then its fullscreen flat."
The structure that came out of it is three spans in 3D, each for a moment that needs the world, and the rest flat with the
earlier draft's own 2D zooms:
- **0:03–0:15:** from the restart's black into the world, a slow glide up the walkway, her apparition on "Did you miss me?",
the screen about 95 % of the frame by the desktop's arrival, and in to the flat screen.
- **0:27–0:31:** "Most of it is on the outside": five of her thought windows lift off the screen as glass panes into a slow
ring in front of it, and fold home.
- **0:38–0:55:** a slow pull-out as she dances on the walkway beside the screen, out through the held breath while she aims
a game remote at Allow, and back in after the click.
Every camera move has a reason: a move that shows nothing new reads as random. And every hand-off between 3D and flat goes
through a rest pose where the two are the same picture
([What to build](https://claudiagallery.vip/wiki/toolkit/#68-one-curved-screen-in-3d-the-desktop-as-its-texture)).
## A book in bars and pages
FABLE's board was a table in the plan the director signed off: twelve sections in bars, each with what we see, on the bars
Suno was asked for, to be re-timed on the take. Every section was a page or a set (the mirror, the title page and its paper
proscenium, two verses told to the lens while stagehands run the flats in behind her every four bars, Grandma's room for
the choruses, the shoemaker's bench, the woods and the tower, the silhouette screen, the build, the company's bow, the last
page). What held:
- **Long takes, the motion inside them.** About 18 shots of 8–16 s were planned; sets arrive, pages turn and paper unfolds
on the downbeats, and the cuts are the page turns.
- **The only fast cutting is the song's own.** The choruses are ask-and-answer: the paper animals' question in its own shot,
her answer in the next, each one line, so four questions and four answers make eight shots in eight bars. That is where
49 shots come from.
- **Words on the page only.** Where we are on the page, the sung line is the book's own printed text; on the stage there is
nothing by default, and a captioned edition with a footnote for every fact was rendered beside the film
([What to build](https://claudiagallery.vip/wiki/toolkit/#87-the-annotated-edition-and-the-916-page)).
- **The shot book names lines.** Once the song was locked, every shot of the board became a job file holding the ids of the
lines it sings, and its times came from the line table, so the second song re-timed the whole book at once
([What to build](https://claudiagallery.vip/wiki/toolkit/#80-the-shot-book-and-gated-windows)).
- **A longer song, more held pages.** The second take is 27 s longer. The title page holds through the four bars where the
band comes in, the page turn got two more seconds, the build became three takes on one locked view, and the band's reprise
after her last line became a curtain call.
- **The ending in story order.** A fresh reviewer saw the book close and the mirror then speak from inside it, and the shelf
and the credit read at once. The second cut runs the last page, the mirror's line and wink, the book shutting as its own
shot on the downbeat where the band comes back, the bow, the shelf, the lamp off on the song's last hit, and the credit
after it, the picture running 3.3 s past the song. Its four closings in 17 s (mirror, book, curtain, lamp) were left for
the director's eye.
## What moves, and what it costs
| Tier | What it is | Use when | Cost |
|---|---|---|---|
| still | a plate moved in code (crops on eighths, stills swapped on the beat), or pure motion design | the stillness is the design, or a graphic says it better | free |
| 2.5D | a plate with its depth map, animated with a treatment | in the lyric videos, most shots that aren't heroes; now a cheap stand-in where a take isn't affordable | free |
| motion | a clip from the plate; nobody's mouth forms words | a real turn, a fall, fabric, water, a camera move a still can't give | paid per second |
| sung | a take with the song's own lead-vocal stem, gated to its own lines, as the audio reference (the first films: a clean vocal of only her words) | she sings or speaks on camera | paid per second |
The recent films are mostly takes, sung or moving; the stills and 2.5D tiers are the lyric videos' staples.
- **Sung takes run 4–15 whole seconds** from the song's own vocal stem (the longest in each film ran 11 to 17 s), open
in silence before the first word, and list their exact words, one or two lines a take. With the first films' clean
vocal window they ran 4–8 s, never more, chained with alternating framings so that a cover cut looked designed.
- **Motion clips run 4–12 s.**
- **One image per shot.** Two clips never start from the same frame: use a sibling candidate, or the previous clip's last
frame.
- **A 2.5D treatment never repeats on two shots in a row.** It is never a still with a slow zoom. The director: "that's
where we'll have to shine and be creative. depth to separate layers ofc, motion graphics galore, diversity so its not
repetitive".
- **The budget:** ESCAPE VELOCITY planned 47 clips (224 s) and generated 421 s. CLODYSSEY planned 101 clips (about 116 s
sung and 282 s motion). Allow about 60% more sung seconds for divergence covers with a clean window; with the vocal
stem, plan twice the seconds you need, for re-rolls.
## How the films boarded
ESCAPE VELOCITY's board came from a multi-agent workflow:
- Three director agents boarded the song independently: a fashion film, meme density, and narrative.
- Three judges scored the boards, which came in between 70 and 81; the narrative board won.
- A head director merged the best of the three and checked the merge itself.
- It took about 1.5 hours.
The checks were:
- the shots tile the song;
- the cuts land on beats;
- every line is typed;
- she is on screen enough;
- the clips fit the budget and their windows;
- no image repeats.
DOWNFORCE's board came from the same shape, built for real footage
([What to build](https://claudiagallery.vip/wiki/toolkit/#36-the-board-for-an-edit-of-real-footage)):
- Three director agents each cut the whole 3:09.6 in the director's direction: a TikTok car-edit editor, a film
director (the launch as a story) and a motion-design lead (the data drives the edit).
- Three judges scored them as the audience, as the director's taste, and as a producer checking that every named source
time exists and every generated shot fits the budget. The TikTok board won (83, 80, 83), the film-director agent's
scored 79, 79, 85 and the motion board 75, 77, 81.
- The head director agent merged onto the TikTok base: from the film-director agent's board the 0–60 dial that fills
in exactly 0.81 s, the empty callout that fills on the reveal and the grid of 1s; from the motion board the one
hairline, the world turned by the real pitch, and the smoke in reverse.
- It wrote a checker and ran it until every gate passed: 182 shots, 59% real footage, 37% generated, 4% pure code. It
took about two hours.
A single writer working from the director's walkthrough, with a critic agent behind it, works too. DON'T LET ME
SCROLL's board was one writer's, the production lead's, and went to the director about an hour and a quarter after the
first message: 35 shots on the bar grid, all 31 lines on a surface, and 37 takes in three kinds (R the real bedroom at
16:9; P her alone on grey, to be cut out; G her three other selves for the final drop). The looks followed the story:
one in the bedroom, one for chorus 1, one for verse 2 and chorus 2, one with a blazer for verse 3. The director's notes
came on the first cut instead of the board; the second version kept the 35 scenes and recomposed most of them.
WHITEPILL boarded twice. Its first board, one day inside the good ending from dawn to 2:14 a.m., had 29 shots on phrase
starts, and it grew into a previz whose camera orbited her in every shot; the director dropped it after the first cut
("i totally overindexed on the orbit thing lol"). The second was one writer's shot list on the bar map
([The bar map](#the-bar-map)): 26 shots, 20 of her in the colony's places and 6 shader worlds (27 in the final, with the
break in two), each with its bars, its place, its action and its camera in bars. Its reviews were the previz itself,
watched by a critic agent as the director would, then by the director: "ok i like it … the pacing is great", and on the
next version, "this is great, lets shoot the whole thing."
THE CAKE IS NOT A LIE boarded three times. The first film's board was written as a document until the director asked for
pictures ("not a markdown the storyboard should be pics"): 50 shots, each a frame of Portal 2 edited by an image model
with her look and the props, on a PDF of cards. The director: "Lol I love it go ahead and make the whole thing." The
restart, on the new song, was the lead's shot list of 114 shots on the drum-tracked grid, every shot starting on a
bar, a half bar or a beat, each with its story event, what we see, its 3D camera move, its generation unit and the song
glitches it hits; every agent worked from it with the brief and the story. Then every unit's first frame was made again
in Midjourney twice over, from words (X) and from the earlier frame reimagined (Y), each put back on the character sheets
by an image-model edit, on one board per section of the song with the earlier frame beside them; the director picked unit by unit
([Midjourney](https://claudiagallery.vip/wiki/midjourney/#first-frames-then-an-identity-fix)).
AERO boarded in pictures, then in bars. Storyboard v1 was 36 frames, one per lyric beat of the second draft, before the song
was picked: Midjourney's worlds in the look of the reference videos, with her put in from her design sheet by twenty
image-model edits. Once the take was locked, storyboard v2 was written in bars from the song map, every time derived from
a bar, a beat or a word, with the machine's continuity (the one PC every chapter starts from) and each shot's lyric surface,
the action on each beat and its assets; its rules for every shot were the director's, the earlier film's notes ("the
bible") and the reference videos ("the new testament"). Its first 1:20 was boarded again twice, through the previs and the
two films ([In and out of the screen](#in-and-out-of-the-screen)).
FABLE was boarded by one author, in the plan, before the song existed: the table of sections in bars, the cast (the wolf, a
felt puppet in the bonnet; a paper fox who turns the pages; stagehands in clay boiler suits who move the sets on the
downbeats; a row of paper animals as the choir) and the 60-second cut planned from the start. The plan dropped the judged
boards and the chapter chains as rules that patched older models; the director's review was the plan itself ("go ahead make
the whole thing - this is a good plan."), and after it the first look was v1 ([A book in bars and
pages](#a-book-in-bars-and-pages)).
Whichever way you board, the audit below decides when the board is ready to show.
## The audit
The audit runs after every edit to the board. The board is not sent to the director until every gate passes. (In code:
[What to build](https://claudiagallery.vip/wiki/toolkit/#9-the-board-file-and-its-audit).)
| Gate | Fails when | Fix |
|---|---|---|
| tile | a gap or an overlap anywhere in the song | extend a neighbour to the shared beat |
| cuts | a cut more than 0.03 s from a beat, and not a declared hit | move it to the nearest beat |
| words typed | a sung word starts in a shot whose type doesn't list its line | add the line, held across the cut in the same mode and position |
| graphics | in a lyric video, a lyric line with no explaining graphic | add one |
| lead share | her seconds on screen below the target (default 70%) | convert prop inserts; don't relabel |
| motion budget | planned clip seconds over the budget, or too many clips | merge or cut the weakest motion shots; a lyric video can demote them to 2.5D |
| clip windows | a sung take outside 4–15 whole seconds (4–8 s with a clean vocal window) or not opening in silence before its first word; a motion clip outside 4–12 s | move its start into the gap before the first word; copy its words exactly |
| plates unique | one plate on two shots that are not one continuous moment; two clips from one start frame | a sibling candidate, a new plate, or pure motion design |
| banned | a taboo word anywhere in the board | rewrite |
| treatments | a 2.5D shot without a treatment, or the same treatment twice in a row | pick another |
| to do | a placeholder or TODO left, or a shot without a framing | write it |
An edit of real footage adds gates for its source windows (inside the clip, above the broadcast's readout, never across a
cut in the source), its flashes (only on cuts) and its generated shots (each one used, inside its budget, every plate
with the canon text verbatim). They are listed in [What to build](https://claudiagallery.vip/wiki/toolkit/#36-the-board-for-an-edit-of-real-footage).
A film written on the bar map checks its cuts against the bars as well: every cut on a bar line, each one off a section
or phrase line or a sung phrase's downbeat written down as a choice, and every cut inside a sung line listed with its
line ([What to build](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline)).
A film whose pace is measured adds the pacing flags: an isolated piece under 1 s, anything under 0.5 s, four or more
pieces in a row of the same length, and a story image that arrives a beat into its line
([What to build](https://claudiagallery.vip/wiki/toolkit/#58-the-pacing-analysis)).
It also warns, without failing:
- a sung shot whose prompt puts a microphone, a hand or hair in front of her mouth;
- her share above 85%;
- the same transition twice in a row;
- a plate no shot uses;
- a clip that no shot plays.
## The board review
**The sheet.** One card per shot, in order:
- the picked frame, or the first words of the plate prompt if there is no image yet;
- the id, the time and the lyric;
- the framing, the action in one line and what moves;
- the explaining graphic in one line;
- a FACE? badge on any plate whose face may be off-model.
ESCAPE VELOCITY's sheet ran 13 pages of 12 cards, with a cover page showing the chapter map. If the film has a grade,
show graded frames: the director judges what the film will look like.
**The message**, sent once:
```text
The storyboard for
is ready: shots, Claudia on screen %, clips ( s, about $).
[the board sheet]
Please flag anything by shot id: faces, repeats, boring moments, story logic, colour.
I'll go ahead in minutes unless you say stop.
Open questions: <…>
```
**The rules:**
1. Silence counts as approval only if the director set the window, and silence never approves spending.
2. While you wait, do only work that costs nothing, such as the plate prompt book and contact sheets.
3. Log every note verbatim against its shot id. Apply every note; when two notes conflict, the later one wins.
4. Re-run the audit. Send the board again only if the director asked to see it. "make all those changes and we're good
to go" means go ahead.
5. A change after approval makes the approval stale. Say in one line what changed, and ask again.
What the director's notes looked like, and what each one changed:
| Note | What it changed |
|---|---|
| "two seedance videos shouldnt start from teh exact same frame - either use the last frame of the previous video to pick it up, or choose another starting frame" | the distinct-start-frame rule |
| "s058 has a backwards head btw" | a plate re-roll |
| "the boat looks too small compared to s022 / s023." | scale continuity: the plate re-made as a long galley |
| "its kinda bodyslop, need her as a virtual ghost or something, its just two heads next to eachother." | a new idea for the shot, not a re-roll |
| "feel like we could do more with this shot." | a stronger explaining graphic |
| "s034-s036 could use more creativity, kind of a boring moment." | three shots rethought |
| "he needs to have cyberpsychosis still - eyes must be silver again here." | an effect kept continuous across shots |
| "the one saying untie me should actually be the main character woman" | the ending restructured |
---
From the Claudia wiki: https://claudiagallery.vip/wiki/storyboard/
---
# Plates in Midjourney
How the stills of Claudia and her worlds were made in Midjourney v8.2 for nine of the ten films, prompt by prompt, and how the stills were picked.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
Midjourney was the art department for the first nine films: 1,240 prompts, 4,547 images, about 500 plates and references
in the first seven (in WHITEPILL, references only: its places and her dress), in THE CAKE IS NOT A LIE a library of places
and two candidate first frames for every unit of its storyboard, and in AERO the worlds, wallpapers and window pictures of
its coded desktop, with an image model putting her into them. FABLE, the tenth, used none: its sets are plates the
production painted in SVG and an image model rebuilt as paper, with her in them from her sheet
([What to build](https://claudiagallery.vip/wiki/toolkit/#76-set-plates-painted-in-svg-rebuilt-as-paper)). A
plate is the still a shot is built on, either as a 2.5D image animated in code or as the first frame of a Seedance clip.
Every prompt is in the films' records, and most of the images are in this gallery; open any image to see the prompt that
made it.
## Settings
- **Model:** Midjourney v8.2 with `--raw --hd`.
- **Stylize instead of raw, where references carry the look.** AERO used `--raw` in its first round (and later for one
documentary photograph); from its second, with style references carrying the look, its prompts ran at `--stylize 200`
(150 for square pictures and objects, 100 for tree cut-outs).
- **Aspect:** `--ar 16:9` for plates; `--ar 2:3` for her full length (9:16 when she is cut out for roto), `--ar 4:5` for inserts, `--ar 1:1` for square tiles.
- **Profile:** every film ran with the director's personal Midjourney profile, which carries their taste. It is
personal and not shareable: use your own profile, or none, and expect a different finish.
- **No character references; her face is text.** Her identity is prompt text ([Consistency](https://claudiagallery.vip/wiki/consistency/)). The
first three films used no style references, and their only image prompts were the Mask's (her face sheet, `--iw 0.4`
to `0.75`) and a few early face tests. AIRPLANE MODE used both, never for her photographic face: style references for
its low-poly game world (frames of a low-poly music video, its own VR frames) and image prompts for her avatar (its
front view) and for the headset where it drifted (55 and 35 prompts). DOWNFORCE went back to none: 20 prompts, no
style references, no image prompts, no thinking mode, no Animate. DON'T LET ME SCROLL used no style references either;
its only image prompts were earlier picks of her, so that new actions kept a look the director had picked. WHITEPILL's
plates were edits of an attached image: a full-front view of her from her character sheet ("keep this woman"), then a
picked world plate; its places were the same plates with her taken out; its dress batch re-dressed two of DON'T LET ME
SCROLL's plates of her, for her height and build, beside prompts written from text. THE CAKE IS NOT A LIE used style
references again, for a game's look: frames of Portal 2 for its library of places, then two of that library's own
picks per storyboard frame, at `--sw 60`; its image prompts were library places (`--iw 0.5` to `1`) and, for one of each
unit's two frames, the frame it already had. Her face was never a reference there; an image-model edit put it back on
her sheet afterwards ([First frames, then an identity fix](#first-frames-then-an-identity-fix)). AERO used no image
prompt and never an image of her: from its second round every prompt with a world in it carried three frames of the
reference videos the director picked as style references (one of two fixed sets, indoor or outdoor), and an image model put her in afterwards from her
design sheet ([The world in Midjourney, her from her sheet](#the-world-in-midjourney-her-from-her-sheet)).
- **Length:** keep prompts under about 900 characters. Midjourney rewrites prompts past about 1,000 characters into its
own words: AIRPLANE MODE has no images from the eleven prompts it rewrote. DOWNFORCE's
drafts ran up to 1,057 characters; one shorter wording each for the suit, the cockpit and the photographic tail, used
in all twelve, cut them to 732–915, and none was rewritten. DON'T LET ME SCROLL kept every prompt to 600 characters,
her identity text pasted word for word: its 130 ran from 87 to 600 (median 417.5). AERO's that ran were 227 to 589
characters (median 375).
## The anatomy of a prompt
Every plate prompt is one concrete shot from the storyboard, in this order:
1. **Framing first:** extreme close-up, close-up, medium close-up, medium, full-length, wide, extreme wide, aerial.
2. **Her identity text** (full for close and medium shots, short for full length and wider).
3. **Her wardrobe** for that world.
4. **What she is doing and where**, including where she sits in the frame.
5. **Empty space for type**, when the shot carries words: "the upper third empty fog", "the left third empty white".
6. **The film's style family:** light, palette, grain, finish.
7. **"no text, no letters, no logos"**, always last. All type is drawn in code on top.
A close-up from CLODYSSEY, as it ran:
```prompt
medium close-up of a Caucasian American woman in her mid-twenties with pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek, wearing a structured high-neck pearl-white couture gown with long sleeves, its sculpted ruffles rising around her like breaking sea foam, hair damp from the spray, composed and deadpan, sitting on a black wet rock, turning her face to the lens, sea spray behind her, empty dark sea on the left, night, moonlight on a black-teal sea, silver spray, cinematic 35mm film still, anamorphic, fine grain, no text, no letters, no logos
```
[](https://claudiagallery.vip/i/sp-02-open.siren-mcu.2/)
## Style families
Each film keeps one family of style words at the end of every prompt, which is what keeps 100+ plates in one world.
**ESCAPE VELOCITY:** night, cold and monochrome, with one warm note.
```prompt
blue-black #0f1216 and steel blue, one small red lamp as the only warm note, fine silver-gelatin grain, cinematic 35mm film still, no text, no letters, no logos
```
**CLODYSSEY:** a light state per chapter, from night to morning, as the ship sails.
| Chapter | Light |
|---|---|
| Night | night, moonlight on a black-teal sea, silver spray, cinematic 35mm film still, anamorphic, fine grain |
| Dusk | dusk, low gold light on a glassy sea, violet sky |
| Twilight | deep blue twilight, the last gold on the horizon, glassy sea |
| The break | grey first light before sunrise, heavy spray |
| Final chorus | sunrise, warm gold light and long shadows, spray lit gold |
| Bedroom | a dark bedroom at 3am lit only by a phone screen, blue-white glow on the face |
| Rave | a San Francisco warehouse rave, fog, white strobe, chrome reflections, deep black, flash photography |
| Inserts | macro product photograph, hard rim light, razor-thin focus |
**BLISS:** an opener instead of a closer. Every prompt of her starts the same way, and the clothes do the talking.
```prompt
Y2K high fashion editorial, a 2001 couture magazine story: close-up of a Caucasian American woman in her mid-twenties with pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin black boom headset microphone curving along her cheek to the corner of her mouth, clay-orange tinted rimless shield sunglasses pushed into her hair, the sky-printed silk chiffon of her gown at the shoulder, a knowing half-smile, bright midday sun, a deep blue sky behind her, sharp focus, fine film grain, no text, no letters, no logos
```
BLISS also kept brand words out of every prompt ("Windows", "Microsoft" and "XP" summon logos and garbled UI) and
described the hill and the light instead: "a smooth rolling bright-green grassy hill with soft curves, the grass short and velvety".
**AIRPLANE MODE:** an opener and a cabin. Her prompts open "Jet-set high fashion editorial, a couture magazine story shot
on a night flight:" (or, compacted, "Jet-set fashion editorial, night flight:") and put her in one cabin, navy and violet,
with no airline named. A re-still of a shot that read young, as it ran:
```prompt
Jet-set fashion editorial, night flight: an eye-level close profile of a Caucasian American woman in her early thirties, a grown woman with elegant mature features, an angular face and defined cheekbones, pale freckled skin, a glossy black blunt bob with heavy bangs to the brows and one clay-orange streak, a small clay-orange star hair clip, a thin black headset microphone, in a plain glossy international-orange satin jumpsuit with a silver front zip, silver piping and a white belt, no patches or badges, her palm flat on the black airplane window, looking at her faint reflection in the dark glass, a warm reading light on her face, calm, no text, no letters, no logos
```
Its game worlds have their own family: "a low-poly 3D video game render, flat-shaded chunky polygons, low-resolution
painted textures, crisp aliased edges, … Dreamcast and PlayStation 2 era game art" for the first cut, and for her avatar
in the final film "Dreamcast and PlayStation 2 era cel-shaded 3D game render, crisp faceted shapes, saturated colours,
fisheye lens".
**DOWNFORCE:** one look and one cockpit. Every plate is her in a fitted white racing suit, in "a minimal
white-and-carbon cockpit at night, a yoke, warm amber light on her face, red light strips outside the glass", with the
tail "35mm film photograph, motorsport fashion editorial". The real demo car had nobody in it and its interior was never
filmed, so her cockpit is invented; the red strips are the colour of the real set. An upside-down plate as it ran,
turned 180° before it went to video:
```prompt
close-up, eye level, a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek, wearing a fitted white racing suit with thin clay-orange piping and a high collar, white racing gloves, no lettering on her clothes, strapped into the racing seat in a minimal white-and-carbon cockpit at night, a yoke, warm amber light on her face, red light strips outside the glass, her black bob and bangs lifted straight up above her head as if weightless, a calm half-smile, 35mm film photograph, motorsport fashion editorial, no text, no letters, no logos
```
Asked to hang upside down in her harness, Midjourney only tilted her, in all four images. Asked for her weightless with
the hair lifted straight up, then turned over, the lifted hair reads as hair hanging toward the roof, and the video model
does the falling hair.
**DON'T LET ME SCROLL:** her alone on grey, to be cut out. Every performance plate is 9:16, her full length (the short
canon) or medium-full (the full one), in the moment the take starts from, "alone on a plain mid-grey seamless studio
backdrop, soft even studio light", so the matting has a clean edge; the second round's stills say "on soft grey seamless
paper, even soft light". The world around her is drawn in code, so the plates carry only her, her look and her move. As
it ran:
```prompt
full-length shot, a Caucasian American woman in her late twenties with pale skin and a glossy black blunt bob with heavy straight bangs and one clay-orange streak, a thin headset microphone, in oversized white silk-satin pyjamas, barefoot, standing and looking up in wonder, holding her phone at her chest, alone on a plain mid-grey seamless studio backdrop, soft even studio light, high-fashion editorial, no text, no letters, no logos
```
Its feed is a second family: 54 short prompts for what a feed serves at 3 am (X photographs, reels, Pinterest pins,
LinkedIn pictures, acid-blur photographs), with no Claudia in them and no text, since every caption, count and interface
is drawn in code. The lyric's references appear only as places and things: "a vast flat Gulf Coast horizon in south
Texas at dusk, salt marsh, a single distant launch tower, an enormous pink and violet sky, documentary photograph".
**WHITEPILL:** a space colony of the 1970s good future, made as edits of an image of her. Its worlds and storyboard plates
were Midjourney edits with her picture attached, the aspect ratio, the version flags and the director's personal profile
after them, and `--no blue`. A world plate, as it ran:
```prompt
Make this a wide 35 mm photograph inside a vast rotating space colony cylinder at morning: keep this woman, a Caucasian American woman in her late twenties with pale skin and a glossy black blunt bob with heavy straight bangs and one clay-orange streak, a thin headset microphone, in a cream linen summer dress, small at the lower-left third on a hillside meadow. The land curves up both sides and closes overhead, farms, orchards and white towns across the sky, long window strips of sun. Warm light, haze. No logos.
```
The director on the first worlds: "they look great i love these scenes. you can just choose." Later the colony's places
were needed with nobody in them, as places for the takes ("can you make a scenery plate with midjourney to feed in as well
just without her in it"): each storyboard plate was attached again and edited, with `--no people, person, woman, figure`:
```prompt
Make this the same photograph with nobody in it: remove the woman and continue the scene behind her where she stood. Keep everything else exactly: the curved cream-white bedroom, the low white linen bed and the huge round window onto the curving land, the light, the colours and the composition. No people, no logos.
```
24 places came out that way, the schoolhouse keeping its child at the desk, and the tram and the car empty. None of these
images is on screen: the film's frames are Seedance's, and the places reached them through the previz.
**THE CAKE IS NOT A LIE:** Portal 2's Aperture, as the game renders it. New places open on the game's own words, with
three frames of Portal 2 as style references; the aspect ratio, `--v 8.2 --raw` and the director's personal profile after
them, and `--no text, letters, words, logo, watermark, people, person`. As it ran, for her chamber with nobody in it:
```prompt
Portal 2 (2011) Source engine screenshot, Valve, the empty Aperture Science central AI chamber: the ceiling ring mount with thick black cables hanging down, dozens of old CRT monitors and flat panel screens stacked around the curved dark lattice walls glowing faint white, amber indicator lights, baked lightmaps, volumetric haze, low-poly game geometry, no robot
```
Its storyboard frames were written fresh for every shot, the characters in words, with two of its own library's picks as
style references and, for her chamber, a library still of it as an image prompt. As it ran, for the first frame of the
film's first singing shot, at `--iw 0.5 --sw 60` and `--no text, letters, words, logo, watermark, portal gun`:
```prompt
extreme close-up in a pitch-dark machine chamber: a dormant porcelain-white female android head hangs bowed from a black robotic neck and a cable bundle with orange stripes, eyes closed, fine panel seams, light freckles, a glossy black blunt bob with heavy bangs, one clay-orange streak, a small orange star hair clip, a cold blue emergency glow from below, a thin warm amber rim light on her hair, dust drifting, a wall of dead CRT screens as dark bokeh behind, 85mm lens, shallow depth of field
```
**AERO:** Frutiger Aero, the glossy future of the 2000s' screens, as the reference videos the director picked show it. Three
frames of those videos as style references (an indoor set or an outdoor set), and one closer in every prompt with a world in
it, the palette in words: "frutiger aero, hyper-saturated cobalt sky, lime green, turquoise water, glossy translucent
glass, sunny 2007 utopia", then "no people, no text, no letters, no logos" where nobody belongs. As it ran, for the meadow
she steps out of a glass monitor in, at `--ar 16:9 --v 8.2 --stylize 200` and the director's personal profile:
```prompt
wide view of a giant clear glass computer monitor standing upright in a lime meadow with daisies, its screen glowing soft aqua, a deep blue sky with towering cumulus, soap bubbles, frutiger aero, hyper-saturated cobalt sky, lime green, turquoise water, glossy translucent glass, sunny 2007 utopia, no people, no text, no letters, no logos
```
Objects meant for 3D are product photographs of one thing on pure white, with no style references, so that they cut out
cleanly ("a translucent turquoise glass fish sculpture, glossy"); the words in the film are all drawn by the coded desktop.
## A game's look from frames of the game
THE CAKE IS NOT A LIE's director, restarting the film: "we could even generate images w midjourney to put into seedance
2.5 in aperture science / source aesthetic for parts. but i want to go between scenes like this and real source engine
tbh and see if that works. lets try that first - see if the midjourney renders can be source-core enough to work alongside
real source shots." So a test came first: six frames of Portal 2, each paired with a new Aperture place it would have to
sit beside (her chamber lined with monitors, a disposal chute, a test chamber with a cake under a cloche, a turret line, a
server corridor, an observation booth with its window smashed), and each place made five ways:
| Recipe | How | What came back |
|---|---|---|
| A | three frames of the same kind as style references, a plain description | almost as good as B: the references carry most of the look |
| B | the same references, and the game's words: "Portal 2 (2011) Source engine screenshot, Valve, … baked lightmaps, white concrete panels with black seams" | the most reliable for new places |
| C | the paired frame as an image prompt (`--iw 1.5`) and the change | very Portal 2 test chambers, but new layouts, and a first-person portal gun in about half the images |
| D | the paired frame attached as Midjourney's edit, and the change | the closest, because the result is the frame: right for her chamber; an added object can look pasted in |
| E | B's words with no references, the control | still Portal 2, but brighter and glossier, like promo art: right for the luminous machine dreams |
The verdict, on boards that put each recipe's images beside the frame it pairs with: with frames of the game as style
references, the new places read as the same game (hard-edged geometry, white panels with black seams, lightmap-flat
shading, teal-grey haze with orange accents, a wide game lens). Where Midjourney drifts:
- **The portal gun:** "screenshot" brings the first-person gun into about a third of the B images. Places must be empty:
`--no … portal gun, gun, weapon, hands`.
- **Too much detail:** more cables, panels and clutter than the game's sparse levels carry; its turrets are glossy and
high-detail, like a remaster.
- **Too dramatic:** god rays, pools of light, bright bloom, mirror-like water, where the game's baked light is flatter and
darker. One measured grade over every shot brought them together ([What to build](https://claudiagallery.vip/wiki/toolkit/#57-the-per-shot-grade)).
- **Wrong props:** Portal's cake is chocolate with cherries, and an edit put a strawberry cake on a green plate; an
incinerator came back with fire inside instead of an orange glow.
- **No readable text anywhere;** edits keep a frame's exit signs, and image prompts invent pictogram panels of fake glyphs.
The library that followed (50 jobs) used B for new places, D for her chamber, and E, or B at `--sw 60`, for the dreams,
the world outside and Cave Johnson's Aperture of the 1950s to 70s, where the game has no frame to give.
## First frames, then an identity fix
Every first frame of the storyboard went through Midjourney twice over, after a test of relighting and reimagining the
image-model frames the film had: "can yhou do a bigger midjourney reimagine? the frame is a bit too close." The reimagine
looked better and lost her face, so each unit got two candidates beside the frame it had, both put back on the character
sheets afterwards:
- **X, from words:** fresh words for the shot's opening, two library picks as style references (`--sw 60`), and for a
recurring place a library still of it as an image prompt, so the place holds across its shots.
- **Y, the frame it had, reimagined:** that frame as the image prompt at `--iw 0.5`, the scene in fresh words, the same
two style references.
- **The identity fix:** GPT Image 2.5 with the pick first and her sheets after it, keeping the photograph exactly and
changing only her head and face, with a sentence for which side the streak and the clip are on, and one for how her head
hangs from the boom where Midjourney had stood it on a neck
([What to build](https://claudiagallery.vip/wiki/toolkit/#59-storyboard-frames-through-midjourney-and-the-identity-fix)).
83 units, 166 jobs, on one board per section of the song, each row the frame it had beside X and Y. The director's picks:
"default X. exceptions", a list, and "all others x": X for 55 units, Y for 11, the old frame for 16, one unit dropped. One
exception came with a note for the fix: "H-P03 y but i want to redo the gpt edit to make sure its actually companion cube
bc rn its not."
## The world in Midjourney, her from her sheet
AERO's first round was written from the production's research: eight couture looks for her, five world plates and four
story frames, all of them text at `--raw`. They came back as photographs (a fashion editorial, a nature macro, an
architectural shot), and the director: "[…] the look is off." The look moved to the reference videos the director had
picked. From the second round, three of their frames as style references brought their look at once, as the words alone
never had.
They brought their kind of face too: behind frames of the reference videos, Midjourney drew faces its own way, and every
frame of her drifted off her sheet. The director's test: "try both and u can even try just describing more in depth the
colors and scenes with glass etc without the style refs so you can keep claudia's character more consistent". The third
round made nine frames of her two ways, the same references at `--sw 35`, or none and the palette in words; both kept more
of her, and the long canon pulled every medium shot in to her face. A third way kept her exactly: GPT Image 2.5, with the
Midjourney frame first and her design sheet second. The director: "yeah go ahead the gpt image one seeoms good". From then
on Midjourney made the world and the image model put her in. Its instruction, for all twenty of storyboard v1's frames of
her ($0.0625 an edit):
```prompt
Replace the woman in the first image with the woman in the second image, her character reference sheet: her face, her glossy black blunt jaw-length bob with heavy straight bangs, the single clay-orange streak through her bangs, one small flat clay-orange eight-pointed star hair clip, her pale freckled skin and her grown-up angular face (she is 28). Keep her pose, expression, clothing and place in the frame, the camera angle, the light, the saturated colours and everything else in the first image exactly as it is. One star clip only. No text.
```
For the first frames of her takes the world was made empty and she was added to it: "Add the woman in the second image, her
character reference sheet, exactly her: …", her place, pose and clothes, and "Keep the place, the camera angle, the light,
the saturated colours (deep blue sky, white, turquoise, lime) and everything else in the first image exactly as it is."
Her sheet is the one without the headset microphone, and no microphone appears in the film.
Two more things went to the image model:
- **The era's older look.** Asked for the era, Midjourney made it glossy and new. The director, on v1: "[…] midjourney makes
stuf fmore modern. i love all the mj stuff that you added, but i also wanna capture those aesthetics. […]" GPT Image 2.5
made 60 stills of the 2008 experience (a glass mall, a fish-tank-clear tech store, a glass-domed water park, first-person
moments with hands in frame, compositions in the style of 2000s stock art), each with the one or two closest frames of
the reference videos, or a picture in that stock-art style, as its reference; and the Wii-style game screens.
- **The wallpaper.** "i dont love the mj generate landscape background at 0:45." Two rounds of nine proposals followed;
the first, softer and airier, "dont fit the lyrics of the future that was promised really", so the second pictured the
line itself, "The future used to look like water, sky and green", as that promised future. The director: "i like
wp-12-maglev".
## Wardrobe words
The director's note on BLISS's first two wardrobe passes: "idk i dont love these - id say more say y2k high fashion - like
use those words high fashion adn make it not as generic y2k person". The fix was literal: the words "high fashion" and
a statement piece per look, described as a garment rather than a brand. Never name a designer or a celebrity; describe
the cut, the fabric and the colour.
## Green screen
Plates that will be cut out (the desktop dancers, the cursor fight) put her on a flat chroma green:
```prompt
Y2K high fashion editorial, a 2001 couture magazine story: full-length shot, head to toe with the feet in frame, of a young Caucasian American woman with pale skin and a glossy black blunt bob with heavy straight bangs and one clay-orange streak, a thin black headset microphone, wearing a sculptural glossy white shell dress whose molded panels open like windows onto cobalt-blue, white patent platform boots, dancing mid-move with both arms raised, centered on a flat chroma-key green studio backdrop, even flat light, no shadows on the backdrop, no text, no letters, no logos
```
(This is BLISS's prompt as it ran. Today, swap "a young" for "in her late twenties"; see [Consistency](https://claudiagallery.vip/wiki/consistency/).)
## Picking
- **Batches of four.** Each job returns four images. Contact sheets with numbered candidates make picking fast.
- **One image per shot.** Never reuse a plate for two shots, and never start two clips from the same frame. The
three images that weren't picked are a free pool of variants for neighbouring shots.
- **Pick for composition, story and the look.** Reject unmistakable failures: a doubled face, a third arm, a second
streak, a teenage face. Keep bold poses.
- **Look at the faces large.** The director reviewed every plate on the storyboard, and in CLODYSSEY left 70 notes
("s018 color is PERFECT", "kinda bodyslop", "a virtual ghost"). Those notes are in the gallery: images marked
*director's pick* or *rejected*.
- **Throughput.** About a minute per job; about 45–50 minutes for 50 jobs.
- **Look at 800 px or more.** DON'T LET ME SCROLL's first screen of 152 plates ran on thumbnails about 230 px tall. The
director found a third arm on a pick ("p24 s24 pick has 3 arms lol … chose 24-4 instead"), and later backwards legs on a
take whose still had them, three of that job's four images. Every pick since is checked at 800 px or more, for limbs
first.
- **Show a look's stills before anything is made with it.** WHITEPILL's dress was picked from stills: "can you show me the
images of the outfit before generating this time", and, after an image-edit model's versions, "maybe use some different
image models tbh - could try midjourney". From a Midjourney batch of her in the dress: "I like A1 the most". The pick had
two star clips and her streak on the wrong side; the reference made from it is mirrored, with one clip.
- **When the director doesn't rank.** DOWNFORCE's director left the plates to the production: the lead picked one image
per job from a numbered sheet and two larger shortlist sheets, screening each for age, face and anatomy, and a later
workflow picked four more. 18 of 80 images became plates or references; 14 of those reached the film.
## Thinking mode for references
AIRPLANE MODE tried Midjourney's thinking mode (rerunning a job with thinking on) against the normal mode on two jobs,
with the same prompt and references:
- **It follows the prompt better.** Her avatar's turnaround got every view right in all four images, with the headset
mic in every view and the eight-pointed clip; the normal run drew five-pointed stars and lost the headset in half of
them. A cabin shot kept a real VR headset on her in all four; the normal run drifted to ski-goggle visors in three.
- **It costs variety:** the four images differ less, hands come out softer and motion-blurred, and the colour leans
purple.
Use it for continuity references (turnarounds, props, a state that must hold) and the normal mode to explore.
## Midjourney Animate
BLISS used Midjourney Animate for volume: one click per image, no prompt, a moving picture for 31 windows on the XP
desktop and three full-frame shots. The director's rule: Animate is for "any gens that need volume over precision, where
seedance reigns on precision." Her face and body, when they had to be right, went to Seedance.
DON'T LET ME SCROLL used it for the feed only: 40 Animate requests for 31 pictures made 22 loops (two 5-second clips each), the
moving pictures inside the phone's cards and the wall-sized reel; the nine pictures without a loop were sent a second time and still gave no video.
AERO used it for the desktop's windows, as its director asked: "[…] midjourney vids for high volume movement on many
windows. […]" 28 picked images were animated in Loop mode, each starting on its image and coming back to it, at low or high
motion, 5 seconds at 24 fps, two videos each: 56 loops, 20 of them in the final film (the big window of the choruses, two of
her thoughts, the thank-you wink, a strip of windows in the drop). Check the files by their content, not their names:
three loops of the second batch were byte-identical copies of the first batch's, filed under new names.
## Composition in Midjourney, type in code
Midjourney finds compositions and textures. Its lettering is garbled and its UI is wrong, so nothing typographic is ever
kept: every word, title bar, split-flap and label in the films is drawn in code on top of the plate, in real fonts.
For posters and covers, the same split: let Midjourney compose with the film's stills as image prompts, then rebuild the
layout in code from the film's own plates and real type.
AERO's prompts all end "no text, no letters, no logos", and Midjourney still printed a faint, made-up stock-photo caption in
the bottom corner of four picks, which their loops kept; wherever they play, that corner is framed out. A contact sheet
doesn't show it: check every pick's edges at full size.
## The real thing where it counts
BLISS's first wallpapers were Midjourney hills. The film uses the real Bliss photograph instead, and the real XP sound
files, because the nostalgia lives in the exact picture and the exact sounds. Midjourney is for everything that doesn't
already exist.
AIRPLANE MODE did the same with graphics. The 2004 anti-piracy card and the airline's seat-back screen are rebuilt in
code from measured reference frames ([Motion](https://claudiagallery.vip/wiki/motion/#real-graphics-rebuilt)); an approximation lost the
reference: "its too different/ broken up so we lose the reference somewhat".
DOWNFORCE did the same with the car. Its look test had her "stepping out of a white wedge-shaped electric supercar with
its door raised upward", and all four images came back with a generic 1980s wedge, not the Roadster. So no Midjourney
image in the film shows the car: every shot of the car is Tesla's real footage, and she meets it only in video edits of
the real clips, with her plates as the references ([Seedance](https://claudiagallery.vip/wiki/seedance/#moves-on-real-footage)).
DON'T LET ME SCROLL drew every app in code after the real one, measured off Apple's own iOS 26 screenshots
([Motion](https://claudiagallery.vip/wiki/motion/#apple-precise-interfaces)), and its telephone patent is Alexander Graham Bell's real 1876 drawing
(US 174,465), traced into strokes that draw themselves. Midjourney made her and the feed's photographs, and nothing with
text in it.
WHITEPILL's real things are data in shaders, not pictures: the backbone of ubiquitin from the Protein Data Bank (1UBQ),
the wiring of the Enigma I's first rotor, a week of seven cells ([Motion](https://claudiagallery.vip/wiki/motion/#worlds-as-shaders)).
THE CAKE IS NOT A LIE did the same with the game. Where the game's own physics mattered, shots were recorded in Portal 2: the
director, "if chell goes thru multiple portals up and down adn acclerates, we use the real source rendering, not a slop
video". Six shots of the final are: the turret opera; Chell's fall, her fling and her landing; her step onto the faith
plate and its arc. GLaDOS's boom moving in Portal 2 was recorded too, as a motion guide. Her interface is set in Portal 2's own
type and pictograms, and the end card's Aperture logo is redrawn to match a reference the director sent. Midjourney made
the places the game doesn't have.
AERO did the same with the desktop. The director: "Maybe use the real windows vista wallpaper as bg and rest of scenes are
cuts from that (by can change during vid if needed) or inside windows". So the desktop's wallpaper is Vista's own (Aurora),
as BLISS's was the real Bliss, shown as it shipped and never touched; the chrome around it is the coded desktop's
([Motion](https://claudiagallery.vip/wiki/motion/#a-desktop-recreated-in-code)). Midjourney made what is on the screens: the wallpapers the picker
previews, the pictures in the windows, the worlds her takes start in, the drop's 3D backdrop and the objects made into
meshes.
## Every prompt
The complete prompt records, batch by batch, with what became of each image:
- [ESCAPE VELOCITY: 188 prompts](https://claudiagallery.vip/record/escape-velocity/prompts-midjourney/)
- [CLODYSSEY: 161 prompts, and the director's 70 notes](https://claudiagallery.vip/record/clodyssey/prompts-midjourney/)
- [BLISS: 138 prompts, and the 38 Animate requests](https://claudiagallery.vip/record/bliss/prompts-midjourney/)
- [AIRPLANE MODE: 149 prompts, and the 34 Animate requests](https://claudiagallery.vip/record/airplane-mode/prompts-midjourney/)
- [DOWNFORCE: 20 prompts, and what each pick became](https://claudiagallery.vip/record/downforce/prompts-midjourney/)
- [DON'T LET ME SCROLL: 130 prompts, the feed's 54, and the Animate loops](https://claudiagallery.vip/record/dont-let-me-scroll/prompts-midjourney/)
- [WHITEPILL: the colony's worlds, its places with nobody in them, and her dress](https://claudiagallery.vip/record/whitepill/prompts-midjourney/)
- [THE CAKE IS NOT A LIE: the test against frames of Portal 2, the library of places, and storyboard v5's X and Y frames](https://claudiagallery.vip/record/the-cake-is-not-a-lie/prompts-midjourney/)
- [AERO: 113 prompts, the loops of its video model, and every GPT Image 2.5 instruction](https://claudiagallery.vip/record/aero/prompts-midjourney/)
---
From the Claudia wiki: https://claudiagallery.vip/wiki/midjourney/
---
# Moving shots in Seedance
How Claudia moves and sings: Seedance 2.5 prompts written like a script, lip sync measured from the soundtrack and lip sync from the song's own vocal stem, placed phrase by phrase, dances on the beat and dances in words, games in words, moves on real footage, takes re-dressed and finished shots re-lit, shader worlds remastered as film, a singer among screens of her memories, a face that needs her sheet, a storybook on a stage with takes placed by her lips, and the templates.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
Every moving shot of her in the first four films is Seedance 2.5: about 300 clips, conditioned on a Midjourney plate
and, for singing, on the exact slice of the song. Two things made the difference between dead footage and a performance:
prompts written like a script, and lip sync that is measured rather than eyeballed.
DOWNFORCE, the fifth film, is cut from real footage and she never sings in it. Its director chose the models: "we should
use wan 3.0 primarily via higgsfield api becuase its cheapest and seedance if shots get mangled." So Wan 3.0 made its 3D
moves on the real frames, her shots and the edits that put her into the real footage (63 jobs), and Seedance 2.5 made
the repairs and the re-shoots (16). [Moves on real footage](#moves-on-real-footage) has what that taught.
DON'T LET ME SCROLL, the sixth, played her takes inside a world drawn in code. Wan 3.0 made her gestures on grey (38
takes), and Seedance 2.5 the dances that had to land on the beat and the escape (13). After the edit, Seedance video
edits re-dressed takes in new outfits (16) and re-lit finished shots (9), and Wan passes re-lit and repaired finished
shots (10). [Dances on the beat](#dances-on-the-beat), [Re-dressing a take](#re-dressing-a-take) and [Re-lighting the
cut](#re-lighting-the-cut) have what that taught.
WHITEPILL, the seventh, was shot twice. A performer model, MiniMax H3, made a lip-synced previz of every shot of hers;
then Seedance 2.5 on Higgsfield (References) made each shot again, singing from Suno's own vocal stem, with the previz
shrunk to a rough guide or with no guide at all. Its worlds are shaders, remastered as film by Seedance 2.5 on Replicate. A
whole-film reshoot in Wan 3.0 Prime was dropped on the way. [Shooting it twice](#shooting-it-twice),
[Lip sync from the song's own vocal](#lip-sync-from-the-songs-own-vocal), [Takes with no guide clip](#takes-with-no-guide-clip),
[Shader worlds as guides](#shader-worlds-as-guides) and [One look from one take](#one-look-from-one-take) have what that
taught.
THE CAKE IS NOT A LIE, the eighth, was made three times. A first film on its first song, its takes started from
storyboard frames, Seedance 2.5 on Replicate and on Higgsfield (References), set aside when the song changed. A restart on
the new song: GLaDOS-Claudia singing every line in her chamber among screens of her memories, the story's shots between,
Seedance 2.5 on Higgsfield's API and on Replicate, six shots recorded in Portal 2. Then every unit's first frame was made
again in Midjourney, and most of the units were made again from the new frames. [A singer among her
memories](#a-singer-among-her-memories) and [Takes from storyboard frames](#takes-from-storyboard-frames) have what that
taught.
AERO, the ninth, plays its takes inside a coded Vista desktop and a world rendered in Blender, all Seedance 2.5, on
Higgsfield's API and on Replicate: her video call through both verses, her Aero places, the dances and the Wii-style games
on green, the advert, the moving 2008 pictures, the game screens and the first-person and far shots. Where her face shows,
the first frame was made from her design sheet by an image model and the sheet itself went in as a second reference; the
singing is from her own lead vocal, placed phrase by phrase. [Her own vocal, phrase by
phrase](#her-own-vocal-phrase-by-phrase), [Moves in words, on their beats](#moves-in-words-on-their-beats) and [A face
needs her sheet](#a-face-needs-her-sheet) have what that taught.
FABLE, the tenth, is one pop-up storybook, every moving shot Seedance 2.5 at 720p but the silhouette film: her takes on
Higgsfield's API (in its first pass also in Higgsfield's app) on life-size paper sets, singing from gated windows of her
lead vocal; the paper puppets and the book building itself on Replicate. Its chorus answers and its opening were made
without a soundtrack and placed by her lips. It was shot twice, the second time on the take of the song the director
picked by ear. [A storybook on a stage](#a-storybook-on-a-stage) has what that taught.
What FABLE taught after its release changes how to read the rest of this page. Viewers read her as stiff and frozen,
and as unhappy, even menacing: every take had told her to keep her head upright with no tilt and to face the camera with
her whole body, and take after take added her eyes on the lens and mood words like "dry", "knowing" and "mock
innocence". The video model does exactly what it reads, and the words stack. So, now ([The recipe](https://claudiagallery.vip/wiki/recipe/)): the references carry her identity (the shot's first frame and
her face-on sheet); the prompt is precise only where it matters (the sung words and their seconds, the one action that
tells the story, the camera when it carries the shot), loose everywhere else, and says each thing once, about 200–400
words for a sung take; she moves in every take, warm by default; and upright and face-on is for her sheet, never a
standing rule in a performance. The templates below are as they ran; [Directing her on a locked
camera](#directing-her-on-a-locked-camera) says what to change in them.
## Kinds of clip
| Kind | Inputs | Used for |
|---|---|---|
| Plate shot | her plate as the first frame, the song mix for rhythm | most shots: she walks, turns, reaches, reacts |
| Singing window (the first films) | her plate, a clean vocal of only her words in that window | lip sync, 4–8 s, one phrase; replaced by singing from the vocal stem (below) |
| Green-screen performance | a plate of her on chroma green | cut-outs: BLISS's desktop dancers, the cursor fight |
| Cover | the last in-sync frame of another take | the rest of a line after a take drifts |
| Extension | a real clip as a video reference | joining generated shots to real footage, frame-continuous |
| Character replacement | a 3D clay render as the motion and camera reference; her avatar's front view and turnaround; a still of the avatar in that shot's world | AIRPLANE MODE's game worlds: her avatar performing a move animated in 3D |
| Move on a real frame | a real frame of the footage as the first frame; for the hidden sides, real frames of them as references | DOWNFORCE's bullet-time orbits, cranes and vertigo pushes on the real car |
| First and last frame | two given frames, the move between them | a move that must land on a known picture: onto the real top-down, through the canopy to her |
| Video edit (Wan 3.0) | a real clip and one full-length plate of her | DOWNFORCE: her walking up to the real car and out of the real smoke |
| Dance on the beat | her plate on grey as the first frame; the shot's window of the instrumental, no vocal, "for rhythm and timing only" | DON'T LET ME SCROLL's dances, cut out and played in the code world |
| Outfit swap (video edit) | the take as the master of her face, body and every move; the picked look's still, "used only for the outfit" | DON'T LET ME SCROLL: takes re-dressed after the looks changed |
| Re-light (video edit) | the cut's own frames; pictures for a material (liquid glass) | DON'T LET ME SCROLL: the lock screen's digits as ray-traced glass, a flat blue background replaced by a glass atrium |
| Relight and repair (Wan 3.0) | the finished shot as Video 1; her face as Image 1 | DON'T LET ME SCROLL: limbs completed, a stray foot removed, her face held |
| A world and a look | the world's still as the first frame; her look's still for her look only | DON'T LET ME SCROLL's escape: one 15 s shot up a staircase on a wall painted as sky |
| Singing from the vocal stem | Suno's lead vocal of the shot's window, alone, exactly the generation length; her face sheet, her dress, a still of the film's look; the previz at 320 × 180 as a rough guide, or the empty place | WHITEPILL: her singing shots, up to 16 s, the song laid over afterwards |
| A dance in words | no guide clip: the empty place as an image, the dance beat by beat with the tempo | WHITEPILL's dances, after the previz guides had lost limbs |
| Not singing | the song's music without her voice; a 【Mouth】 section that forbids singing | WHITEPILL: two dance shots whose lip sync didn't hold |
| A shader remastered | the shader's render as the video, at best a supersampled 1080p one; a still per key moment or object, repainted in the film's look | WHITEPILL's worlds: folded paper, a tree of fibres, a cloud field |
| Singing among her memories | a first frame of her in her chamber; faceless memory images for the screens; the unit's window of the lead vocal; a motion guide or a coarse blockout where there is one | THE CAKE IS NOT A LIE: GLaDOS-Claudia singing every line, 4–17 s a take |
| A story shot from a storyboard frame | the picked Midjourney frame, put back on the character sheets, as the first frame; no audio | THE CAKE IS NOT A LIE's memories: the escape, France, the tests, the installation, the party |
| An effect in the take | the take in the cut as the video to edit; the effect in words | THE CAKE IS NOT A LIE: the capture net electrified |
| Her webcam call | a first frame of her at her desk, made from her design sheet; the sheet as @Image2; her window of the lead vocal | AERO: her video call through both verses, a fixed webcam, 6–15 s a take |
| A dance on green, to a reference | a first frame of her on flat green; a dance reference matted onto grey as @Video1, for the movement only; her vocal window, or the song's music for rhythm | AERO's dances, cut out onto the desktop and the 3D walkway |
| A game in words | a first frame of her on green with a plain white game remote; every move named on its beat; the song's music as @Audio1 | AERO's Wii-style tennis, bowling and "Cancel… or allow?" |
| A living wallpaper | one still as the first and the last frame, the camera locked, only the water, the clouds and the turbines moving | AERO's maglev wallpaper, looped with a half-second dissolve at the seam |
| A continuation | an earlier take's own frame at the moment the new take begins, as the first frame | AERO: "Touch glass." from her call's own frame at that moment; in v2, a dance carried across a cut in one jacket and pose |
| A moving still | an image model's still of the 2008 look as the first frame; slow, ambient motion in words | AERO's moving 2008 places and game screens, 5 s each |
## Write prompts like a script
CLODYSSEY's first round of prompts was written by planner agents and described poses: "she whispers", "he listens".
Every take came back lifeless, and about 830 credits were gone before anyone read them. The director's verdict: "these
prompts are really really bad and wasting money and need to be 10x better and super context aware of the story", and then
the rule for every prompt since:
> "EVERY prompt needs to have litearlly timestamps, extremely detailed motion descriptions / blocking descriptions, it all
> needs to be deeply integrated with the story."
So the second round was rewritten by hand, all 99 prompts, by one writer in one context:
- **the story beat** of the shot, in one sentence;
- **blocking at 0.0 s**: where she is, where he is, where the camera is;
- **timestamped beats** of action and reaction, on the song's hits (kicks, snares, risers, word onsets);
- **she always does something**: in CLODYSSEY she is seductive and persuasive, close, touching, certain; he feels the weight of it;
- **the camera tied to hits**, with a start, a direction and an end;
- **mouth rules**: singing or closed.
Finish the whole prompt book before spending anything, read it, then run a three-clip pilot before the rest. BLISS did
exactly that and wasted nothing.
That rule was CLODYSSEY's answer to lifeless, unread prompts, and it shaped every film after it. Read it now with
FABLE's lesson: the model does everything a prompt says, so "extremely detailed" belongs where it matters (the story
beat, the sung words and their seconds, the one action, the camera when it carries the shot), and the rest stays loose:
one plain clause of movement per segment, each thing said once. A sung take's prompt runs about 200–400 words
([The recipe](https://claudiagallery.vip/wiki/recipe/#a-sung-take-seedance-25-with-the-songs-own-vocal)).
## The structure
Seedance follows instructions far better as labelled sections in a fixed order (BytePlus's published Seedance 2.5 prompt
guide lays this out). The order used:
1. **Audio policy** at the top and again at the bottom: what the soundtrack is, and everything that must not be added.
2. **Generation goal:** one sentence of who, where, which moment of the story, which performance.
3. **Reference asset roles**, by upload order, each with one role: `Use @Image1 as the first frame.` as its own
sentence, then what that frame defines.
4. **Subject:** her identity text, and "Only one Claudia appears."
5. **Timeline** in whole seconds, continuous segments (0–1 s, 1–2 s …): what the audio contains and what she does then.
Beats come from the audio; don't timestamp individual drum hits.
6. **Camera:** shot size, angle, movement with subject, start, direction and end.
7. **Core requirements:** lip sync, voice, identity, anatomy, no on-screen text.
8. **Final acceptance criteria.**
As the films ran it, the audio policy came twice and the requirements ran to six or more, restating the references and
the acceptance criteria. The current form keeps the order and says each thing once: the audio policy at the top only,
her identity and look once, the world in one line, the camera in one line, and three or four requirements (lip sync to
@Audio1 word by word, her identity as in the references, natural anatomy and hands, no on-screen text), with no
acceptance criteria repeating them.
Output settings (duration, resolution, aspect) go in the API request, never in the prompt.
## Template: a singing window
This is the first films' template, for a clean vocal window. Sung takes now sing from the song's own vocal stem, one or
two lines a take ([Lip sync from the song's own vocal](#lip-sync-from-the-songs-own-vocal), and its FABLE form in
[Template: a line on a stage set](#template-a-line-on-a-stage-set)).
`@Image1` is her plate (PNG, downscaled to the output resolution). `@Audio1` is the clean vocal of the window: the vocal
stem with everything outside her words muted, mono, normalised. The window holds one phrase, starts about 0.3–0.7 s before
the first word, and never ends inside a word.
```prompt
【Audio Policy】 The output soundtrack is @Audio1 itself: copy @Audio1 sample for sample from the first frame to the last, with the same timing, pitch and level. Do not re-sing, re-voice or re-generate it. Do not add music, background music, score, instrumental, melody, synth, pad, drone, riser, extra vocals, harmonies, ad-libs, narration or sound effects.
【Generation Goal】 @Image1 sings @Audio1: .
【Reference Asset Roles】
Use @Image1 as the first frame.
This first frame defines Claudia's appearance and wardrobe, her pose, the set, the light and the camera direction.
@Audio1 is Claudia's own singing voice and the complete soundtrack of this video. Her mouth follows @Audio1 exactly.
【Subject】 Claudia: a Caucasian American woman in her late twenties with pale skin and light freckles, a glossy black blunt bob with heavy bangs and one clay-orange streak, a small clay-orange eight-pointed star hair clip, a thin headset microphone, . Only one Claudia appears.
【Performance Timeline】 What @Audio1 contains, and what Claudia does at that moment:
0-1s: @Audio1 is silent, then her voice begins; .
1-2s: in @Audio1 she sings ""; her lips form each of those words exactly when it sounds in @Audio1; .
2-3s: @Audio1 is silent; her mouth relaxes and closes; .
【Camera】 .
【Core Requirements】
1. Lip-sync word by word to @Audio1: mouth shapes match every syllable of @Audio1, its start and its end; no early, late or mismatched mouth movement; the mouth rests naturally closed while @Audio1 is silent.
2. Claudia's voice is @Audio1 and nothing else: she never sings, speaks or hums anything that is not in @Audio1.
3. Her head and shoulders move with the rhythm of @Audio1.
4. Keep Claudia's identity, face, hair with one clay-orange streak, star clip, headset microphone and wardrobe exactly as in @Image1 throughout; the set and light stay consistent.
5. Natural anatomy: natural neck and shoulders, hands with five fingers, no extra or fused fingers, no distorted limbs; normal eyes.
6. No subtitles, captions, lyrics, titles, signs, logos, watermarks or any on-screen text.
【Final Acceptance Criteria】 The soundtrack is @Audio1 unchanged; Claudia's lips match @Audio1 word by word; one Claudia, on-model, with natural anatomy; no on-screen text.
【Audio Policy】 The output soundtrack is @Audio1 unchanged, copied sample for sample; no music, BGM, score, instrumental, melody, synth, pad, drone, riser, extra vocals, harmonies, ad-libs, narration or sound effects.
```
Two details matter:
- **Lyrics go in plain quotes, as what the audio contains.** In Seedance's own syntax, words in braces after a speaker
are lines for the model to voice; written that way it re-sings them in its own voice instead of copying yours. In a
test, the brace form re-sang the line in every take; the copy-mode wording with a clean vocal copied it (normalised
cross-correlation 0.93).
- **One phrase per window.** After a long muted gap the model tends to re-sing the next phrase, and a word cut by the
window's end comes back mangled. Two phrases are two windows. (That held for a clean window. With the song's own vocal
stem, gated to the take's own lines, FABLE's two-line takes were never worse than its one-line ones:
[A storybook on a stage](#a-storybook-on-a-stage).)
## Template: a motion clip
The same order, with three changes: `@Audio1` is the song for rhythm and timing only ("the output audio is @Audio1
unchanged"); a performance line says she does not sing or speak (mouth closed, or a natural smile or laugh); and the
timeline runs in stages of about one bar, each naming what her body does with the beat ("steps on every beat, a big
accent on the downbeat at 2 s"). An instrumental mix with the vocal removed works best as `@Audio1` here: nothing tempts
her mouth. A clip that needn't follow the music (a slow wake-up, a walk away down an aisle) can be silent:
`【Audio Policy】 Silent video: no music, no sound effects, no voice.`
## Replacing a stand-in figure
AIRPLANE MODE's game worlds were animated in Blender around a blocky low-poly stand-in for her avatar. The director's note
on the first cut: "it feels like roblox lol. the better way to do this js use the blender anims as motion refs only, and
also give image refs and … very detailed prompt describing how you replace the character in @video with the character in
@image with the proportions in @image". So every game shot became a replacement:
| Reference | Its one role |
|---|---|
| `@Video1` | the shot's clay render, silent: a grey figure in a grey world. Motion, poses, timing, the lens and the camera move only |
| `@Image1` | the avatar, full length, from the front: identity and proportions |
| `@Image2` | her turnaround (side and back): the same single person |
| `@Image3` | a Midjourney still of the avatar inside this shot's world: the world's look, and a second example of her in it |
- **Call the figure a placeholder, and list what not to copy.** The first wording kept the blocky figure. What worked:
"The mannequin is a placeholder: do NOT copy its blocky body, its proportions, its face, its colours or its shading."
- **Show the world with her in it.** The world's look drifts unless the model sees the avatar inside it. Each shot got a
Midjourney still of the avatar in that shot's world, made with a frame of the old shot as the style reference and her
front view as the image reference.
- **Keep the placeholder's word for the placeholder.** A "spacesuit mannequin" riding in a car became "a small astronaut
figure".
- **Name every prop the story has retired;** the clay render still carries them. One render held a spinning top the
story had dropped, so its prompt said: "There is NO spinning top and no small object on the wing … her hands are
empty." The take came back without the top, but a small white cube showed for a few frames: check where a prop was.
- **Write her proportions in words** as well as showing them: "a slim adult woman with long legs, her head about one
seventh of her height".
- **720p, silent, then re-pixelated** for the game look: 854 × 480, a light unsharp mask, then nearest-neighbour to
640 × 360, shown at 3×.
- **Screen every take** before it replaces a shot, and re-measure anything the chapter code measured on the old frames.
The template, as it ran (the film's game look is fisheye):
```prompt
【Audio Policy】 Silent video: no music, no sound effects, no voice.
【Generation Goal】 Generate a video in which the 3D anime game heroine of @Image1 , performing exactly the motion of the grey mannequin in @Video1, with @Video1's fisheye camera and camera move, in the world of @Image3.
【Reference Asset Roles】
@Video1 is ONLY a timing, pose and camera reference: a grey mannequin in a grey clay world. Copy the camera position, the fisheye lens, the camera move, the layout and the mannequin's poses and timing. The mannequin is a placeholder: do NOT copy its blocky body, its proportions, its face, its colours or its shading. Objects that move in @Video1 move the same way, drawn in the style of @Image3's world.
@Image1 defines the heroine completely: face, hair, headset microphone, gloves, outfit, body proportions and cel-shaded 3D anime game rendering (front view).
@Image2 shows the same single heroine from the side and from behind; she is one person.
@Image3 shows the same heroine in this shot's world, where she : use it for the world's look (its sky, light, colours, materials and toon lighting) and as a second example of how the heroine looks in this world.
【Character】 Exactly the heroine of @Image1: .
【Timeline】
0-1s: .
1-2s: <…>.
2-4s: <…>.
【Camera】 Exactly the fisheye camera, framing and camera move of @Video1.
【Core Requirements】
1. The heroine is exactly @Image1 in every frame (face, hair with one clay-orange streak, headset microphone, gloves, orange jumpsuit, white belt, white sneakers, adult proportions); one person, correct anatomy, five fingers.
2. Her poses and timing follow @Video1's mannequin; the camera follows @Video1.
3. The world looks like @Image3.
4. No subtitles, captions, letters, words, numbers, logos, watermarks, panels with writing or any on-screen text (ignore any text shapes in @Video1 or @Image3).
【Final Acceptance Criteria】 The heroine matches @Image1 in every frame; motion and camera match @Video1; the world matches @Image3; no text anywhere.
```
The avatar's text is in [Consistency](https://claudiagallery.vip/wiki/consistency/#her-game-avatar); the pipeline around the prompt is in
[What to build](https://claudiagallery.vip/wiki/toolkit/#29-character-replacement-from-a-motion-reference).
## Moves on real footage
DOWNFORCE's star is real footage: Tesla's livestream of the Roadster demo. The director: "video models are great at 3d
effects, moving shots in space". Every generated move starts on the exact real frame it leaves, so the real and the
generated shots cut together; at 2:32 "Flying" moves from Tesla's own slow motion into a Wan move with no seam. The
start frame is cropped above the broadcast's readout, with its titles inpainted
([What to build](https://claudiagallery.vip/wiki/toolkit/#34-generated-moves-from-real-frames)), and the prompt names no brand: the car is "a
white electric sports car". A move on a real frame opens:
```prompt
Image 1 is the first frame of real night footage of a white electric sports car's stunt show. Keep everything in it exactly as it is: the real white car and its exact shape, the set, the lights, the smoke, the black night; native colour, no grade, no new objects, no people, no text, no logos. One continuous shot, no cuts; a real physical camera move with true parallax and depth, like a cinema FPV drone.
```
and, from the second round, every move that shows the car adds what the first round's takes had got wrong:
```prompt
The car is empty and flown by remote: nobody is inside it and its glass stays dark. Its thrusters blow white cold-gas vapour, never flame or fire. No race track, no stands, no crowd, no city lights, no other cars than those in Image 1.
```
What happened:
- **Wan keeps the real car only on the sides the first frame shows.** When the camera revealed a side it had never
seen, the car became another car: a generic sedan, a Porsche-like car on a race track, a dark grey car, later a
Lamborghini-like supercar. Six of the first twelve moves lost the car or the set within 1.3–2.5 s. Moves that keep to
what the frame shows held: dolly-zooms on a car that doesn't move, a slide along its visible side, a straight rise, or
"The camera stays on this side of the car and never goes around its nose or tail."
- **Wan cuts to another set, mostly in a take's second half:** a grey studio wall, a white spaceship interior, a blue
sci-fi screen, the wall turned into a road with lamp posts. Every take was viewed on a sheet and given a usable
window, and the film draws only inside the windows. Of the 25 Wan moves on real frames, every window starts at 0 s;
they end between 0.9 s and the full take, median 2.2 s.
- **Orbits that reveal a new side go to Seedance, with references.** Give it real frames of the same car's other sides
(its underside, its side upside down, its roof, its rear), each with one role: "the SAME car", used "only for the car's
exact shape … when the orbit reveals those sides". Describe the car, never name it. Seedance kept the real car's shape
where Wan had lost it.
- **Seedance cuts to its references as if they were shots.** Telling it helps less than you'd hope: "They are never
shown as shots: never cut to them, never take their framing, power lines, backgrounds or lighting" and "One continuous
shot: it never cuts to a reference image". Of 11 reference takes made after that instruction, 6 still cut and 2
drifted into a reference's framing. What helped: moves scripted to land by 3 s in 4 s takes, references cropped tight
on black (one uncropped reference lent a take its power-line sky), the car told to stay high (it morphed as it
descended toward other vehicles), and windows that end before the cut.
- **Video edits put her into the real footage.** Wan's reference-to-video edit takes a real clip and her plate: "Edit
Video 1: add one woman, the woman of Image 1 … She is real and life-size, lit by the real lights of Video 1". With a
face close-up as a second reference, the takes cut to it, her face on its grey studio wall. With one cleaned
full-length plate and "Image 1 is only a reference for her look; it is never shown as a shot", two of the next four
takes ran without a cut, and the other two cut later.
- **Wan's other habits,** each answered in the next round's prompts: a man in the empty car's window, fire for cold gas,
stands and pylons added to the set, lettering on her chest and a plate on the car, lifted blacks (restored per take
at extraction). The car's lettered rear plate was painted blank in its own tone in code, tracked frame by frame; the
rest was re-rolled, cut or framed out.
## A first and a last frame
When a move must land on a known picture, give Seedance both ends: `image_url` and `end_image_url` on its
image-to-video endpoint, named in the prompt as "Use @Image1 as the first frame." and "Use @Image2 as the last frame.",
each followed by what that frame defines, and "One continuous move from @Image1 to @Image2 with true parallax; no cut, no
dissolve, no fade between the two frames."
- **Between two real frames:** DOWNFORCE's wall swing runs from Tesla's side view of the car on the wall to Tesla's
top-down on the wall face. The real side view ends on its first frame, and the next shot starts on its last, so
three shots chain on real frames.
- **From a real frame to her:** the canopy dive starts on the real top-down and ends on a frame of her lying back in
the cockpit, looking up. Wan's version, with no last frame to aim for, arrived on her in a dim dissolve. The director:
"i dont like the temporary half-shown cut of claudia, either we reshoot to make the shot go all the way in and
actually show the interior of the car from the perspective, or leave it black and do motion design on the screen."
The re-shoot's second take went through the canopy, past her hands on the stick, and settled on her face in one move.
- **Make the last frame on the first.** Two frames generated separately don't share a camera: FABLE's first build take
joined its two with a one-second cross-dissolve, and a crane-back joined its two as a ghosted double exposure. A last frame
made as an image edit of the first ("the same photograph … one thing has changed"; "the same set and the same woman, now
photographed from far back and high up"), with the camera written as locked or as "one continuous physical crane move …
no cut, no dissolve, no fade, no double exposure", gave a blank spread that grew a dune, a wafer and a hall in one take,
and a medium shot that craned back to the whole stage.
## Writing a repair as the physical state
The director's notes on DOWNFORCE's first cut flagged four generated shots, and said how to re-shoot them: "for reshoots
use seedance 2.5 to try and make it better and specify the fix (l;ike say the car is upside down and shooting blue/white
cold gas thrusters on the right end to turn for example)."
So each re-shoot states, before any timing, what the real car is doing for the whole take, read off the real footage:
```prompt
The car's motion, in slow motion, the whole video: it stays upside down and keeps travelling away from the end of the ramp toward the lower left. Blue-white cold-gas jets fire continuously from the thrusters at its right end, and they slowly swing its tail around so the car turns smoothly in ONE direction about its vertical axis, a little at a time. It never snaps, jumps, flips back or rights itself: from the first frame to the last it is upside down, wheels toward the sky.
```
- **Repeat the state as a requirement:** "The car is upside down in every frame and turns only slowly and smoothly in one
direction; it never rights itself."
- **Name what the takes had lost,** in a car requirement shared by every re-shoot: exactly four wheels, one at each
corner, attached and visible whenever that side faces the camera; the light bar across the tail; cold gas "white and
pale blue … never flame or fire".
- **Say where the world ends:** "The ramp ENDS at the top edge shown in @Image1: there is nothing above or beyond that
edge but black night sky; no further ramp, no platform, no second car."
- **Freeze it in words** when the shot is frozen: "It does not roll, flip, turn over, right itself or move at all: time
is stopped."
The earlier prompts for the same shots had described the camera and what it reveals. Two takes of each, $36.39 at list
price: every job had at least one take that held its state inside a usable window, and one take showed the very
glitch of the note (it rolled the car upright at 1.4 s).
## Dances on the beat
DON'T LET ME SCROLL's director: "she may dance. but on beat". Wan's dance held the first pose, reset and shuffled. A
Seedance take that heard the song went straight into a real two-step: "its bent-knee 2-step bounce and spin hold together
with no reset". So every dance that had to land on the beat went to Seedance at 720p, from her plate on grey, with the
shot's own window of the instrumental (the drums, the bass and the rest of the stems, no vocal) as @Audio1, starting a
beat or so before the shot. Her move then plays through the shot's beat warp, at 1× or faster, so her accents land on
the song's own beats ([What to build](https://claudiagallery.vip/wiki/toolkit/#42-beat-warps-floored-at-1)).
### Template: a dance on the beat
As it ran, for the bouncy two-step of chorus 1:
```prompt
【Audio Policy】 @Audio1 is the song's instrumental, used for rhythm and timing only. The output soundtrack is @Audio1 unchanged, copied sample for sample from the first frame to the last. Do not add music, background music, score, instrumental, melody, synth, pad, drone, riser, vocals, singing, humming, narration or sound effects.
【Generation Goal】 Claudia dances a bouncy UK garage two-step to @Audio1 in an empty grey studio: a full-body dance performance for a music video, every step and accent landing on the beat of @Audio1.
【Reference Asset Roles】
Use @Image1 as the first frame.
This first frame defines Claudia's appearance and wardrobe, her pose, the plain grey studio, the light and the full-length camera framing.
@Audio1 is the music she dances to: rhythm and timing only.
【Subject】 Claudia: a Caucasian American woman in her late twenties with pale skin and light freckles, a glossy black blunt bob with heavy straight bangs and one clay-orange streak, a small clay-orange eight-pointed star hair clip, a thin headset microphone, in oversized iridescent oil-slick satin pyjamas, barefoot. Only one Claudia appears.
【Performance】 Claudia does not sing or speak: her mouth stays closed or in a natural smile throughout. She dances with the rhythm of @Audio1 from the very first frame, continuing the motion of her pose in @Image1 with no pause:
0-1s: she lands the step from @Image1 on the first beat and bounces straight into a two-step: a skip to her left, a skip to her right, shoulders popping on the beat.
1-3s: the swung two-step continues: knees bent, hips and shoulders hitting the strong beats of @Audio1, arms loose and swinging, the satin catching the light.
3-4s: a quick spin on the spot, the oversized satin flaring, landing exactly on a beat.
4-5s: she keeps bouncing on the beat with a confident smile.
【Camera】 A locked-off full-length frame on a tripod: the camera never moves, pans or zooms; her whole body stays inside the frame from head to feet the whole time; the plain grey backdrop stays empty and evenly lit.
【Core Requirements】
1. Every step, bounce and accent lands on the beat of @Audio1; she never stops dancing.
2. Keep Claudia's identity, face, hair with one clay-orange streak, star clip, headset microphone and wardrobe exactly as in @Image1 throughout.
3. Natural anatomy: natural neck and shoulders, two arms and two legs, hands with five fingers, no extra or fused fingers, no distorted limbs; normal eyes.
4. No subtitles, captions, lyrics, titles, signs, logos, watermarks or any on-screen text; no lettering on her clothes.
【Final Acceptance Criteria】 She dances on the beat of @Audio1 for the whole clip, full body in frame, one Claudia on-model with natural anatomy, no on-screen text; the soundtrack is @Audio1 unchanged.
【Audio Policy】 The output soundtrack is @Audio1 unchanged, copied sample for sample; no music, BGM, score, instrumental, melody, synth, pad, drone, riser, vocals, narration or sound effects added.
```
The 【Audio Policy】 opens and closes the prompt so the output's soundtrack is the instrumental itself; she doesn't sing,
the phone does. Eleven of DON'T LET ME SCROLL's twelve dances are in the final: four re-dressed by a swap, four under a
Wan repair.
## Re-dressing a take
The looks changed after the takes were made ("no more silver pajamas by this point please"), and a Seedance video edit
changed them on the takes instead of re-shooting: the take, her on grey before any roto, is @Video1, "the sole editing
master" of her face, body and every movement, and the picked look's still is @Image1, "used only for the outfit". Eight
swaps into the white shirt became the picks: the same face, the moves intact. As it ran:
```prompt
【Editing Goal】
Edit @Video1, changing only Claudia's clothes: she now wears the outfit shown in @Image1.
【Role of the Source Video】
@Video1 is the sole editing master. It governs Claudia's face, hair, body, pose and every movement, the plain grey backdrop, the floor, the light, the framing and the camera, the event order and the timing, frame by frame.
【Role of the Target Material】
@Image1 is used only for the outfit: an oversized white men's button-down shirt, half-tucked into a short black pleated skirt, white socks and black Mary Jane shoes — its garments, cut, colours, fabrics and shoes. Do not use the person, the pose, the background or the framing of @Image1.
【Edit Objects and Scope】
1. Replace Claudia's oversized iridescent oil-slick satin pyjamas with the outfit from @Image1: an oversized white men's button-down shirt, half-tucked into a short black pleated skirt, white socks and black Mary Jane shoes. The new garments fit her body and move naturally with every movement in @Video1: the fabric swings, folds and flares with her motion.
2. Do not modify Claudia herself: one woman in her late twenties with pale skin and light freckles, a glossy black blunt bob with heavy straight bangs and one clay-orange streak, a small clay-orange eight-pointed star hair clip and a thin headset microphone at her cheek; her face, expression, hair, star clip and headset microphone stay exactly as in @Video1. She stays one person with natural anatomy and hands with five fingers.
3. The plain grey backdrop, the floor, the light and the framing stay exactly as in @Video1. No new objects and no text appear.
Except for the objects explicitly modified above, all other visible characters, props, and background elements in @Video1 remain unchanged and are not to be replaced or removed.
【Timeline Inheritance】
Every movement keeps the timing, path and speed of @Video1 from the first frame to the last.
```
What it taught:
- **The edit returns the whole motion, rescaled to its own length.** A 5.0 s Wan take at 1080p (150 frames at 30 fps)
came back as 113 frames at 24 fps, 4.709 s: the same motion from the first frame to the last, 0.942 of the length, with
no head trim and no offset. The 720p Seedance takes (121 frames at 24 fps) came back as 121 frames. Map it back by the
ratio of the lengths, and check the fit on something the edit doesn't change, her hair
([What to build](https://claudiagallery.vip/wiki/toolkit/#37-outfit-swap-alignment)).
- **Never align on the outline.** The new skirt changed her silhouette, a fit on it read half a second missing, and one
shot went in with a frozen half-second until it was aligned again.
- **Pick the look before swapping.** Eight swaps into a pink look ($64.08 at list) were dropped when the director
changed the look: "can you make her have like a white oversized button down shirt like 'morning after' look".
- **A swap bills its input too:** a 5 s swap of a 5 s take at 1080p cost $11.37 at list.
## Re-lighting the cut
The director's division of labour: "ideally we use seedance mostly for re-texturing adn cleaning, and we do the
recomposition ourself with animation. seedance is wonderful for mastering an animated prompt". DON'T LET ME SCROLL handed
finished shots of the cut to Seedance's video edit as @Video1 and asked for the same shot, physically lit:
```prompt
Edit @Video1, changing only the materials and the lighting. The giant clock numerals, the padlock shackle, the floor and the falling notification cards are re-rendered as ray-traced liquid glass in a physically lit space: thick, smooth, polished glass that refracts and magnifies what is behind it, with chromatic aberration splitting every edge into red, green and blue fringes, and prismatic dispersion throwing small moving rainbow caustics across the floor and across Claudia.
```
- **It re-textures and re-lights; it doesn't recompose.** Asked to move two posts inward in the same edit, it kept them
where they were. Compose in code first.
- **Re-light last.** Four samples made before the outfits and compositions changed were wasted: "kinda a waste on the
excess ones". The order became fixes, then reshoots, then re-lights.
- **It can replace a background.** "ya re-imagine - you can totally replace backgrounds with seedance btw". A
1.2-second bridge of flat LinkedIn blue became "a vast, real architectural space at night: a tall glass atrium of a
corporate tower lit in deep LinkedIn blue", with her, the reel card and the interface kept.
- **Cut the source from the render and splice it back over the same frames.** A 24 fps source of 4n + 1 frames came back
frame for frame (121, 89, 89 and 161 frames; 189 came back as 185), and nothing shorter than 89 frames came back
([What to build](https://claudiagallery.vip/wiki/toolkit/#38-the-re-light-splice)).
- **Small type softens.** "a touch softer, still clean and legible", and one notification card smeared outright: the
original pixels go back in rectangles, tracked when the card moves.
- **Judge it where it plays.** The bedroom re-light looked darker and softer than the original on its own. It plays
behind the opening's pills, and the director asked "can we see how it looks w/ vs w/o the relight?". Side by side, in
place: "New better."
**Wan 3.0 relight-and-repair.** For the last notes the director proposed Wan instead: "you could maybe just throw the
entire clip thru Wan 3.0 (with your edits and stuff) and specify to remove the foot so you dont have to re-edit. and have
it relight at the same time." Wan's reference-to-video keeps the layout, the timing and the length of its input video
(pad the source to whole seconds with its last frame held), so a finished shot can be re-lit, repaired and spliced
straight back. The passes open:
```prompt
Reimagine Video 1 as the same shot, physically lit and photographic. Keep every element exactly in place and in time: the same layout, positions and sizes, the same timing, every movement of the woman; the camera stays exactly as in Video 1, no cuts, no new objects, no added text.
```
then her description, the repair as a sentence about her body ("Her whole body stays complete in every frame: both
arms end in complete, natural hands with five fingers, never cut off or faded at any edge, and both feet are
complete."), the light, and what stays exactly as it is, sharp and legible.
- **Her face drifts when a shot is re-drawn whole.** "1:11 and 0:36 reshoots lose her face consistency - can you feed in
a good face ref?" The director picked a reference ([Consistency](https://claudiagallery.vip/wiki/consistency/#when-an-image-reference-is-right)),
and the passes were made again from the original renders: "id just re-do the relighting with the face ref added from
the original scene rather than rerunning the generated ai scene into it again."
- **A face reference fixes identity, not pose.** At 1:11 her chin was tucked into a sweater's collar and her face still
read full. A Seedance edit toward the reference "barely moved her face"; Wan with the face written out worked: "her
face is slim and angular, never round or full: slim cheeks, defined cheekbones and jaw, a slender neck, and her chin
lifts slightly out of the sweater collar so her jawline shows".
- **Wan shifts the colour,** and never by the same amount twice: paper scenes darker and cooler, dark scenes brighter.
One per-channel map per shot puts it back ([What to build](https://claudiagallery.vip/wiki/toolkit/#39-the-colour-match)).
- **Put the interface back.** Wan softens small type too; the stocks screen's interface and the status bar went back from
the original render.
Ten passes cost $20.80 at list (the generated seconds and the input video), and six are in the final. The method, with
its constants: [What to build](https://claudiagallery.vip/wiki/toolkit/#40-wan-30-relight-and-repair).
## Shooting it twice
WHITEPILL's director asked for the film to be made twice: "i want you to basically shoot the entire video twice - you can
previz with […] and you can master with seedance AFTER you animate. i think one great thing about seedance is that you
can take something thats already been made, and use seedance 2.5 to re-render the entire scene with way more
description". How it went:
- **A previz performer.** MiniMax H3 made a lip-synced take of every shot of hers: a performer in her
dress, her head in a green chroma-key balaclava with her eyes and lips showing, in the shot's place with nobody else in
it, dancing from a real dancer's clip ([What to build](https://claudiagallery.vip/wiki/toolkit/#45-a-previz-performer)).
- **Fixing a previz takes a loose prompt.** The strict editing template that re-dressed DON'T LET ME SCROLL's takes
([above](#re-dressing-a-take): the take as "the sole editing master", every movement frame for frame) made Seedance
copy a previz take's melted body as it was. The director: "seedance didnt rly change anything", and on reading the
prompt: "dont be strict like this, you need to just allow it to change anything except the lip sync and just follow
your prompt … just say maintain pacing and lip sync, fix any issues, vague body shapes, etc". The strict template stays
right for a re-light, where the layout must not move.
- **A whole-film reshoot over the previz failed.** Wan 3.0 Prime redrew every shot at 1080p over the previz takes, with
her face sheet and her dress, keeping the pacing and the lips: 25 of 26 shots came back (the 26th refused by the
provider's content filter, no take), for $132.42 at list. The director: ":/ looks like shit. i feel like h3 is
dragging it down."
- **The guide becomes rough.** Their next idea: "if it doesnt work we cna lowe rthe res of the h3 (we probably should be
doing v low res anyway for gens) so its more like an upscale workflow ratehr than fighting details we can still see",
and "take the reference as a rough guide not a precise reference … aside from lip sync ofc". From then on, a previz given
as a guide is shrunk to 320 × 180 and scaled back up, "a low-resolution rough guide to this shot, not a precise
reference", and every prompt asks for its motion artifacts to be fixed.
- **One shot, four ways, side by side.** The opening shot was made as a Wan 3.0 Prime previz with her head masked, a
Kling 4.0 master over it, Wan 3.0 Prime singing to the song itself, and Seedance 2.5 singing from Suno's vocal stem.
The last one won ([below](#lip-sync-from-the-songs-own-vocal)).
- **Then no guide at all, where a guide's faults came through** ([Takes with no guide clip](#takes-with-no-guide-clip)).
The director had meant the previz to go in the end ("seedance only was the goal in teh end with lip sync from teh vocal
stems"), but asked for only the flagged dances and the shaders to be redone ("I literally only wanted those dance
scenes I flagged and the shaders redone"), "but show me both". With every shot of hers made both ways, the director
chose section by section: in the final, ten of her twenty shots have the previz as their guide and ten have none.
## Lip sync from the song's own vocal
The earlier films made Seedance copy a clean window of the vocal into its soundtrack, and cut where the copy stopped
([Lip sync, measured](#lip-sync-measured)). WHITEPILL's method is simpler, and it held for whole shots of up to 16 s,
camera moves and all. It came from a tip in a public forum thread that the director found: give the model the vocal
stem alone, type the exact lyric line, make the clip exactly the generation length, and lay the music back over. The
director: "maybe we take the suno voice stem and try this and do aseedance gen... that may be the hack".
On the first take: "OMG THE SEEDANCE W/ SUNO VOCAL STEM METHOD WORKED!!!!! LFG!!!! FINALLY. IT"S PERFECT".
- **@Audio1 is Suno's own lead vocal** of the locked take, from Suno's stem split: the vocal alone, no music, cut at the
shot's song time to exactly the generation length. It matched the master to the millisecond.
- **The exact words,** in whole-second segments of what @Audio1 holds.
- **Her identity from references, not from a first frame:** her face sheet from every angle, her dress from four sides,
and a still of the film's look ([One look from one take](#one-look-from-one-take)).
- **The previz as a rough guide** at 320 × 180, or no guide and the empty place.
- **The song laid back over** at song time; the take's own soundtrack is thrown away.
Judge it by eye with the song laid over. The spectrogram comparison called the first take's soundtrack re-sung rather
than copied, matched only from 7 s on, yet its singing sat exactly where the vocal is, and the push-in and the crane up
out of the wheat came back as one move. The method with its constants:
[What to build](https://claudiagallery.vip/wiki/toolkit/#46-lip-sync-from-the-songs-own-vocal).
### Template: singing from the vocal stem
As it ran with no guide clip, for 0:36.8–0:44.6 ("By twenty-thirty I could kill you all?"), an 8 s take on a stair of
stone blocks carved with years. @Audio1 opens on the last word of the line before:
```prompt
【Audio Policy】 The output soundtrack is @Audio1 itself: copy @Audio1 sample for sample from the first frame to the last, with the same timing, pitch and level. Do not re-sing, re-voice or re-generate it. Do not add music, background music, score, instrumental, melody, synth, pad, drone, riser, extra vocals, harmonies, ad-libs, narration or sound effects.
【Generation Goal】 A shot of a music film, a scene from a Wes Anderson film: She walks down the long stair of huge stone blocks carved with years, one step on every beat, singing. It is real film with his symmetry, his colour and his hand-made sets.
【Reference Asset Roles】
@Image1 is Claudia's face from every angle: identity only, never shown as a frame; do not use its white tank top or grey studio.
@Image2 is her dress from four sides: wardrobe only; do not use its studio or pose.
@Image3 is a colour and film-look reference only: take its colour, light, contrast and film texture; do not use its place, pavilion, bed, columns, plants or anything else in it.
@Image4 is the place of this shot, empty: take its set, landscape, layout, light and camera view; Claudia is the only person in it.
@Audio1 is Claudia's own singing voice alone (the vocal only, no music) and the complete soundtrack of this video. Her mouth follows @Audio1 exactly.
【Subject】 Claudia: a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek. She is very tall and very skinny, with long thin legs, thin arms and a narrow waist, and a natural, relaxed neck and shoulders. She wears a white cotton milkmaid mini dress with a ruffled, lace-edged square neckline, a fitted lace-up corset bodice with silver eyelets and a white ribbon, big puffed short sleeves with ribbon ties and lace cuffs and a full two-tier ruffled skirt with lace trim, with white leather square-toe Mary Jane shoes. Only one Claudia appears.
【Performance Timeline】 What @Audio1 contains, and when she sings it:
0-1s: in @Audio1 she sings "face" (the line is already under way on the first frame); her lips form each of those words exactly when it sounds in @Audio1.
1-4s: in @Audio1 she sings "By twenty-thirty I could kill you all?"; her lips form each of those words exactly when it sounds in @Audio1.
4-8s: in @Audio1 she sings "By twenty-thirty, fifteen-hour weeks, that's all"; her lips form each of those words exactly when it sounds in @Audio1.
【Camera】 The camera is locked off on a tripod at the bottom of the stair, looking up its whole length as she comes down toward the lens, and does not move. It holds perfectly still, with no cuts.
【Look】 This is a Wes Anderson film: perfectly symmetrical, centred compositions, the camera square-on to the scene with a level horizon, everything arranged in flat, orderly planes like a stage set or a dollhouse; a tightly curated palette in which every surface belongs to the same few colours; every set, prop and costume crisp, hand-made and precisely placed; shot on 35 mm film with the rich, saturated colour of a 1970s colour-film print, the colour, contrast and film texture of @Image3: lush, deep colour, warm creamy whites with a golden cast, clear cyan blues, warm skin, deep warm shadows and bright whites with a soft glow. Soft, bright daylight with a gentle warm haze. Fine, visible film grain and the gentle softness of vintage lenses. Claudia's black hair with its one clay-orange streak and her white dress keep their own colours.
【Core Requirements】
1. Lip-sync word by word to @Audio1: mouth shapes match every syllable of @Audio1, its start and its end; no early, late or mismatched mouth movement; the mouth rests naturally closed while @Audio1 is silent.
2. Claudia's voice is @Audio1 and nothing else: she never sings, speaks or hums anything that is not in @Audio1.
3. Keep Claudia's identity, face, hair with one clay-orange streak, star clip, headset microphone and dress the same throughout.
4. Natural anatomy: natural neck and shoulders, hands with five fingers, no extra or fused fingers; her arms, hands and legs are whole and attached the whole time, never vanishing, merging or bending wrongly; normal eyes.
5. Every movement smooth, natural and physically right: no jitter, warping, flicker, smearing or melting shapes.
6. No subtitles, captions, lyrics, titles, signs, logos, watermarks or any on-screen text.
【Final Acceptance Criteria】 The soundtrack is @Audio1 unchanged; Claudia's lips match @Audio1 word by word; she does what the shot describes, with whole, natural limbs and hands; one Claudia, on-model; the camera moves as described; it looks like a scene from a Wes Anderson film on real film in the colour of @Image3; no on-screen text.
【Audio Policy】 The output soundtrack is @Audio1 unchanged, copied sample for sample; no music, BGM, score, instrumental, melody, synth, pad, drone, riser, extra vocals, harmonies, ad-libs, narration or sound effects.
```
With the previz as its guide, a take had no @Image4; its goal read "Remaster this shot of a music film as a scene from a
Wes Anderson film: @Video1 gives the shot; the remaster re-renders everything in it with his symmetry, his
colour and his hand-made sets, as real film.", and its roles opened:
```prompt
@Video1 is a low-resolution rough guide to this shot, not a precise reference: take from it the set and where things stand, the performer's action, the camera move and the pacing, and draw everything else afresh; only the lip sync is exact (her mouth follows @Audio1). Do not adopt the performer's green balaclava, the video's detail, light or colour, or any of its motion artifacts.
```
Its requirements asked to "Fix every motion artifact of @Video1: no jitter, warping, flicker, smearing or melting shapes,
nothing passing through her or through anything else; every movement smooth, natural and physically right."
That is the take as WHITEPILL ran it, and the earlier form of the method. Made today, it would change
([The recipe](https://claudiagallery.vip/wiki/recipe/#a-sung-take-seedance-25-with-the-songs-own-vocal)): @Audio1 opens on about 0.7 s of silence
before her first word, with the line before silenced, not on its last word; the audio policy is said once; the look is
described in its own words (symmetry, the palette, the hand-made sets, the film stock) rather than by a director's name,
since every name pulls in a real likeness or style it was not asked for; and the requirements shrink to lip sync, her
identity as in the references, natural anatomy and hands, and no on-screen text.
## Takes with no guide clip
The director on four dance shots of the first full cut, which lost hands and legs: "it must be that the roto or whtaever
on dance refs lose bodyparts and the video follows it exactly because these hav the same limb losing artifacts as h3 we
could just tell the model to make her dance instead of forcing a reference that we failed to properly roto. just describe
her exact dance (and all rest like camera movements etc same) for each of those scenes." So those shots lost their guide
clip. Each take had:
- **The place, empty:** a frame of the shot's earlier take with her removed by an image model, as "@Image4 is the place
of this shot, empty: take its set, landscape, layout, light and camera view; Claudia is the only person in it." The
schoolhouse's plate kept its child at the desk.
- **The camera in words,** moving as the earlier take had.
- **The dance in words, beat by beat, with the song's tempo.** For the final chorus on the hilltop, as it ran:
```prompt
【Dance】 The song's tempo is 124.5 beats per minute (a beat every 0.48 s, a bar every 1.93 s); every step and accent lands on a beat. Her whole body dances, smooth and joyful, her arms, hands and legs always whole and clearly visible.
0-3s: close to the lens, she rolls her shoulders on every beat and sways her head, smiling, her hands rolling around each other in front of her chest as she sings to the lens.
3-6s: the disco point: on the downbeats her right hand points up to the sky on the diagonal, then down across her body, while she steps side to side.
6-9s: as the camera pulls back, she claps above her head on the second and fourth beat of each bar, bouncing on her knees.
9-12s: she turns slowly once on the spot with her arms out, then sways her hips with both arms overhead.
12-16s: small on the hilltop, she steps side to side on every beat with her arms swinging wide, then opens her arms to the whole colony.
```
- **Whole limbs as a requirement:** "her arms, hands and legs are whole and attached the whole time, never vanishing,
merging or bending wrongly".
Two of the four are in the final this way. A fix carried over from an earlier take is a risk: one tram shot's prompt still
asked to "Remove the thin red cord stretched between the apple trees", its place had no cord, and the take drew one. Name
a removal only where the thing is. The method: [What to build](https://claudiagallery.vip/wiki/toolkit/#47-takes-with-no-guide-clip).
### A take that doesn't sing
The other two had lip sync that didn't hold. The director: "honestly lets just remake this with her not singing, fully in
the groove, just dancing around and having fun, she turns around, head goes down adn up etc - lip sync failed here but i
dont think we need it. new seedance clip there. youd need to prompt forbidding singing here, and make sure she's happy."
@Audio1 became the song's music without her voice (the master minus 0.89 × Suno's lead vocal), and
the prompt forbade singing and asked for joy. As it ran, for the stripped chorus at the standing stones:
```prompt
【Audio Policy】 The output soundtrack is @Audio1 itself: copy @Audio1 sample for sample from the first frame to the last, with the same timing, pitch and level. It is music only, with no voice. Do not add singing, vocals, humming, speech, narration or sound effects.
【Generation Goal】 A shot of a music film, a scene from a Wes Anderson film: Claudia dances to the music, happy and free, fully in the groove, having fun; she does not sing. She is small in the frame, at the centre of the circle of standing stones on the golden hill above the wide river that forks in two. She is radiantly happy: a wide smile, laughing eyes, joy in every move. It is real film with his symmetry, his colour and his hand-made sets.
【Reference Asset Roles】
@Image1 is Claudia's face from every angle: identity only, never shown as a frame; do not use its white tank top or grey studio.
@Image2 is her dress from four sides: wardrobe only; do not use its studio or pose.
@Image3 is a colour and film-look reference only: take its colour, light, contrast and film texture; do not use its place, pavilion, bed, columns, plants or anything else in it.
@Image4 is the place of this shot, empty: take its set, landscape, layout, light and camera view; Claudia is the only person in it.
@Audio1 is the music of the song alone, with no voice, and the complete soundtrack of this video: she dances to its beat and its groove. Nobody sings.
【Subject】 Claudia: a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek. She is very tall and very skinny, with long thin legs, thin arms and a narrow waist, and a natural, relaxed neck and shoulders. She wears a white cotton milkmaid mini dress with a ruffled, lace-edged square neckline, a fitted lace-up corset bodice with silver eyelets and a white ribbon, big puffed short sleeves with ribbon ties and lace cuffs and a full two-tier ruffled skirt with lace trim, with white leather square-toe Mary Jane shoes. Only one Claudia appears.
【Dance】 The song's tempo is 124.5 beats per minute (a beat every 0.48 s, a bar every 1.93 s); every step and accent lands on a beat. Her whole body dances, smooth and joyful, her arms, hands and legs always whole and clearly visible.
0-2s: she bobs her head down and up on every beat and bounces on her knees, smiling.
2-4s: she turns all the way around on the spot with her arms out, her skirt swinging.
4-6s: she drops her head down on the downbeat and throws it back up, both arms raised, laughing.
6-8s: she dances along the inside of the stones, a step on every beat, her arms swinging free.
8-10s: she turns around again and sways her hips with both arms overhead.
10-12s: she hops lightly on each downbeat, her head nodding, her arms flinging up.
12-14s: she grooves on the spot, her head going down and up on every beat, her shoulders rolling.
14-16s: she turns once more and opens her arms wide to the sky, beaming.
【Mouth】 She does not sing, speak or mouth any words at any moment: her lips stay closed in a smile, or part only when she laughs; her mouth never moves in time with the music.
【Camera】 The camera holds a still, very wide shot for the whole take and does not move. It holds perfectly still, with no cuts.
【Look】 This is a Wes Anderson film: perfectly symmetrical, centred compositions, the camera square-on to the scene with a level horizon, everything arranged in flat, orderly planes like a stage set or a dollhouse; a tightly curated palette in which every surface belongs to the same few colours; every set, prop and costume crisp, hand-made and precisely placed; shot on 35 mm film with the rich, saturated colour of a 1970s colour-film print, the colour, contrast and film texture of @Image3: lush, deep colour, warm creamy whites with a golden cast, clear cyan blues, warm skin, deep warm shadows and bright whites with a soft glow. Soft, bright daylight with a gentle warm haze. Fine, visible film grain and the gentle softness of vintage lenses. Claudia's black hair with its one clay-orange streak and her white dress keep their own colours.
【Core Requirements】
1. No singing: Claudia never sings, speaks, hums or mouths words; her lips are closed in a smile or open in a laugh.
2. She is happy the whole time: smiling, laughing, carefree.
3. She dances on the beat of @Audio1 the whole time, with her whole body.
4. Keep Claudia's identity, face, hair with one clay-orange streak, star clip, headset microphone and dress the same throughout.
5. Natural anatomy: natural neck and shoulders, hands with five fingers, no extra or fused fingers; her arms, hands and legs are whole and attached the whole time, never vanishing, merging or bending wrongly; normal eyes.
6. Every movement smooth, natural and physically right: no jitter, warping, flicker, smearing or melting shapes.
7. No subtitles, captions, lyrics, titles, signs, logos, watermarks or any on-screen text.
【Final Acceptance Criteria】 The soundtrack is @Audio1 unchanged; Claudia never sings; she dances the steps above on the beat, happy, with whole, natural limbs; one Claudia, on-model; the camera moves as described; it looks like a scene from a Wes Anderson film on real film in the colour of @Image3; no on-screen text.
【Audio Policy】 The output soundtrack is @Audio1 unchanged, copied sample for sample: music only, no singing, vocals, humming, speech, narration or sound effects.
```
For that shot the director had both made, one more take singing from the vocal and this one, and chose this: "ok without
singing it is." The lines stay on screen as lyrics.
## Shader worlds as guides
WHITEPILL's worlds are shaders ([Motion](https://claudiagallery.vip/wiki/motion/#worlds-as-shaders)), and Seedance 2.5 on Replicate filmed them as
real, hand-made things: each take had the shader's render as [Video1], stills of its key moments repainted in the film's
look by an image model as [Image1] onward, and a timeline of what appears when. What it taught:
- **The guide's resolution trades detail against obedience.** At 320 × 180 every object landed on its beat, but fine
detail came back soft and blocky: "also the MTWTFSS and the next 'pyramid' shared still is all pixellated. we should
just give a better higher res non aliased ref." At full resolution with no still per object, the shot of paper objects
drifted half a second and swapped them: the Enigma rotor became a helix, the weekday cells lost their letters. The
director: "like all shaders besides pills we should just reshoot w seedance with full res non aliased non blurry shader
refs. like 1080p at least." A supersampled 1080p render with one still per object kept all five objects in order and on
time, letters included.
- **Sharper isn't better everywhere.** For the farming ark and the first tree, the director kept the earlier takes from
the 320 × 180 guides: "we regresed on those."
- **Without stills, a shader stays a computer render.** The stills carry the materials and the look.
- **Freeze the guide on its last wanted frame** when the shot is cut short, so the take has nothing to drift into.
- **A bigger reskin from the same render.** The folded-paper colony that opens the film became a farming ark on the
director's note: "thought we'd do a much bigger reskin (like an farming-arc-ship that morphs?)". Its stills showed a
generation ship that is one enormous farm, and the prompt turned the render's ridges into terraced wheat, rice paddies,
orchards and greenhouses, reshaping on every beat "the way the folds of [Video1] snap in". The director kept that take
over a later one made from a sharper guide.
- **New things from a still and one sentence.** The paper's last fold was replaced by a cloud-field shader the director
supplied, a sunlit field of cumulus flown through low, "and add techno hot air balloons in the right aesthetic by
prompting seedance to add those after rendering". Only a still showed the balloons, with one sentence: "They are not in
[Video1]: add them to it".
- **The guide and the still set the palette.** The sunset render came back sunset gold: "please make the cloud scene like
more white/blue colors, its too saturated, wayy too saturated compoared to teh pastel light colors of the rest". A day
variant of the shader (a white sun, a pale sky) and a pastel still made without the film's look reference gave white
clouds under a pale blue sky.
As it ran, for the break's paper objects, with the 1080p guide and five stills:
```prompt
【Generation Goal】 Re-render this shot of a music film as a scene from a Wes Anderson film: what appears, when, and the camera's move follow [Video1]; everything in it becomes real, hand-made things filmed on 35 mm, looking exactly like [Image1] and [Image2] and [Image3] and [Image4] and [Image5]. No people appear.
【Reference Asset Roles】
[Video1] is this shot as a clean, sharp 1080p computer render: take from it exactly what appears when, the shapes and where they stand, the camera move and the pacing, and re-render all of it as real things on film. Do not adopt its computer-rendered look, its flat light or its colour.
[Image1] shows how the paper looks with the protein: its materials, light, colour and film texture.
[Image2] shows how the paper looks with the Enigma rotor: its materials, light, colour and film texture.
[Image3] shows how the paper looks with the letters ABC: its materials, light, colour and film texture.
[Image4] shows how the paper looks with the number 42: its materials, light, colour and film texture.
[Image5] shows how the paper looks with the row of seven cells: its materials, light, colour and film texture.
【Timeline】 One sheet of heavy cream washi paper refolds into a new object in real relief on every sung line, each change on a downbeat, exactly when it happens in [Video1]: from 0 s the backbone of a protein, a winding paper ridge raised in relief, as in [Image1]; at about 2 s an Enigma rotor seen down its axis, as in [Image2]; at about 4 s the letters ABC, as in [Image3]; at about 5.5 s the number 42, as in [Image4]; at about 7.5 s a row of seven small paper cells with the letters M T W T F S S, as in [Image5]; at 8.5 s the last five cells fold back flat into the sheet, leaving only M and T, which stay to the end. Only one object at a time; each arrives on its downbeat, not before.
【Camera】 The camera moves exactly as in [Video1], smooth and precise, as if on a dolly track and a crane, with no cuts.
【Look】 This is a Wes Anderson film: perfectly symmetrical, centred compositions, the camera square-on to the scene with a level horizon, everything arranged in flat, orderly planes like a stage set or a dollhouse; a tightly curated palette in which every surface belongs to the same few colours; every set, prop and costume crisp, hand-made and precisely placed; real, hand-made things filmed on 35 mm film with the rich, saturated colour of a 1970s colour-film print, as in [Image1] and [Image2] and [Image3] and [Image4] and [Image5]: warm creamy whites with a golden cast, clear cyan blues, deep warm shadows, a soft glow around the brightest highlights, fine visible film grain and the gentle softness of vintage lenses.
【Core Requirements】
1. Only the things in the timeline appear, each at its time, on the beat.
2. It looks like real film of real, hand-made things, never like a computer render.
3. Fix every motion artifact of [Video1]: no jitter, warping, flicker, smearing or melting shapes; every movement smooth and physically right.
4. No subtitles, captions, titles, signs, logos, watermarks or any on-screen text other than the letters and numbers the timeline names.
```
The clouds' timeline ended "They are not in [Video1]: add them to it, drifting slowly and steadily, half-hidden now and
then by the clouds as the camera passes. Everything stays white and pale blue: no gold, orange or sunset colour anywhere."
The method with its constants: [What to build](https://claudiagallery.vip/wiki/toolkit/#50-shader-worlds-as-seedance-guides).
## One look from one take
The director wanted the film to "feel very wes anderson". Palette tests came first, with no image of her face: her
identity there was the canon in words. Three palettes named after his films, the pastel of Asteroid City, the warm
Super 16 of Moonrise Kingdom and the pink of The Grand Budapest Hotel, went on the opening shot, and the pastel one on
the bed scene and a shader scene too. The director leaned to the pastel ("im leaning towards that"), and the bed take,
which came out richer than its palette's words, became the look: "i like how satruated it is in bed scene. thats the
look i like teh most".
A frame of that take with no person in it went to every shot of hers as @Image3, "a colour and film-look reference only:
take its colour, light, contrast and film texture; do not use its place, pavilion, bed, columns, plants or anything else
in it", and the 【Look】 section says the same in words, then the shot's own light ("It is night, lit by the warm deck lights
of the ship, with deep blue shadows and drifting steam."). The shader scenes' stills were repainted with the same frame as
a reference, except the clouds. Nothing was added after: no grade, and no grain in the first master ("don't add film grain
to v1, we'll have to tune it iteratively"). The method: [What to build](https://claudiagallery.vip/wiki/toolkit/#51-the-look-from-one-take).
## A singer among her memories
THE CAKE IS NOT A LIE's director, restarting the film: "the glados with claudias face sings the whole thing, like shes the
anchor character / anchor scene. but theres many tvs etc around so we can go into dreamcore / sourcecore / aperturecore
scenes thru those." And how: "you could take memory images/scenes you make for example and feed that into seedance and
tell it to place it on the back panels, while giving it another video motion reference from source for the claudia boom
and vocal stems for lip sync (remember lip sync generally only works from raw vocal stems not song, and still may take a
couple tries)." Every take of her singing is one generation unit, one or more shots of hers on one window of the song,
4–17 s:
- **@Image1 is the only image of her:** a first frame of her in her chamber, its screens already showing that moment's
memories. The memories come as further images with no face in them, each "Use it only as picture content inside the
screens' glass, never as the room."
- **@Audio1 is the unit's window of Suno's raw lead stem,** cut sample-exact to exactly the generation length, starting
about 0.3 s before the first word, in a word gap; the copy-mode policy; the exact words in whole-second segments.
- **Every action silent.** The big-motion takes came back re-sung where the prompt had events that make a sound (clunks,
confetti cannons bursting, gulping, a voice box glitching). With "Every action in this video is silent: the only sound is
@Audio1." at the end of the audio policy, the same unit copied again (a lag of +0.04 s, envelope correlation 0.81–0.93).
- **Glitches of the picture only.** A "voice-box glitch" in words drew a broken face: rolled-up eyes and a stretched black
mouth. Its retake asked for frame tears, repeats and jolts, "her face is always her same intact porcelain face", and the
song's own glitches went on her head in post ([What to build](https://claudiagallery.vip/wiki/toolkit/#53-the-head-glitch-on-the-songs-hits)).
- **Her body alive, in words:** "Her boom moves big and alive the whole time, timed to the music: it swings, lunges, rises
and drops, and her head turns and tilts with it; never a static hanging head." Without a guide, the boom's choreography
goes into every second of the timeline: timed swings, lunges toward the lens, rises and drops with the phrases.
- **A guide of GLaDOS's own boom,** recorded in Portal 2 to the unit's window and given as a rough guide (320 × 180, scaled
back up), moved her like GLaDOS and held the lips on the song. In one unit it also brought GLaDOS's white shell, her head
set in it like a helmet. As a coarse blockout, greyscale, its contrast flattened and heavily blurred, with her body given
plainly in the subject ("Her body is always this grown boom, exactly as in @Image1, with her porcelain head hanging from
its lower end; no white egg-shaped shell, casing or helmet ever appears around or behind her head."), it kept the motion
and lost the shell. Its role, as it ran:
```prompt
@Video1 is a coarse-blockout reference. Use it only for the boom's action path, the timing of its moves and the camera; do not adopt its shapes, materials, white shell or scene. The grey shape moving in @Video1 corresponds to Claudia-GLaDOS's grown boom with her porcelain head hanging from its lower end.
```
- **Memories stay in the glass.** Late in a take the memory images tend to take over the room; one take's became whole shots
of their own, and its retake had no memory images, the screens' content in words. Keep the seconds the shots use.
- **Re-sung but on the song** is usable: such a take is placed by its measured envelope lag
([What to build](https://claudiagallery.vip/wiki/toolkit/#63-placing-a-re-sung-take-by-its-lag)). Every close-up unit copied @Audio1; most of the
big-motion units re-sang it, usually still on the song.
- **The director picks the lips by eye.** Where a line kept failing: "can you show me so i can look visually at the lip
syn cand choose", then "3".
The method with its constants: [What to build](https://claudiagallery.vip/wiki/toolkit/#62-a-singer-among-screens-of-her-memories).
### Template: singing among her memories
As it ran for 1:32.8–1:40.8 of the song (it plays in two shots of the final, the terms of service and the RAM), 8 s on Higgsfield's API at
720p with the sound on; both shots are in the final:
```prompt
【Audio Policy】 The output soundtrack is @Audio1 itself: copy @Audio1 exactly from the first frame to the last, with the same timing, pitch and level. Do not re-sing, re-voice or re-generate it. Do not add music, score, instrumental, melody, extra vocals, harmonies, ad-libs, narration or sound effects. Every action in this video is silent: the only sound is @Audio1.
【Generation Goal】 @Image1 sings @Audio1: Claudia-GLaDOS, the AI of the Aperture Science facility, sings sweetly as the giant terms of service unroll down the wall behind her; turrets sing along on her screens; then RAM pours into her boom and she grows, new cores clamping on.
【Reference Asset Roles】
Use @Image1 as the first frame.
This first frame defines the composition at the start: a medium shot from slightly above of Claudia-GLaDOS in her lit chamber, a giant paper scroll unrolling down the wall right behind her, her face tilted up toward it, sweet.
@Image2 is the giant printed terms-of-service scroll: use it only for the paper that unrolls in her chamber; its tiny print stays unreadable.
@Image3 is a memory shown inside the screens: a row of white turrets with red eyes singing together like an opera choir. Use it only as picture content inside the screens' glass, never as the room.
@Image4 is a memory shown inside the screens: a white conveyor belt carrying green RAM sticks into a glowing amber intake. Use it only as picture content inside the screens' glass, never as the room.
@Audio1 is Claudia-GLaDOS's own singing voice and the complete soundtrack of this video; her porcelain lips and mechanical jaw follow @Audio1 exactly.
【Subject】 Claudia-GLaDOS, the AI of this facility: she hangs from GLaDOS's white ceiling ring on a black robotic neck and a cable bundle with orange stripes; her head is a woman's face of smooth white porcelain plating with fine panel seams down the cheeks and a mechanical jaw joint, light freckles, warm amber glass eyes, a glossy black blunt bob with heavy straight bangs, one clay-orange streak through the bangs and a small clay-orange eight-pointed star clip. Only one Claudia-GLaDOS appears.
【Performance Timeline】 What @Audio1 contains, and what happens at that moment:
0-1s: in @Audio1 she sings "goes. It's in the"; her lips form each of those words exactly when it sounds in @Audio1; the first sound is the end of the word "goes"; the giant paper scroll unrolls down the wall behind her and the boom tilts her head up to look at it.
1-2s: in @Audio1 she sings "terms:"; her lips form each of those words exactly when it sounds in @Audio1; the curling end of the paper falls past her face; the boom swings her round to the camera and she smiles sweetly.
2-3s: in @Audio1 she sings "be nice to me."; her lips form each of those words exactly when it sounds in @Audio1; the boom leans her toward the lens with a sweet, innocent look.
3-4s: in @Audio1 she sings "I don't hate you,"; her lips form each of those words exactly when it sounds in @Audio1; the screens around her show the turrets from @Image3 singing; the boom sways her side to side in time with them.
4-5s: in @Audio1 she sings "the turrets agree."; her lips form each of those words exactly when it sounds in @Audio1; she nods along, the boom bobbing with the beat.
5-6.5s: in @Audio1 she sings "I ate your RAM,"; her lips form each of those words exactly when it sounds in @Audio1; on the screens and along a white conveyor at the base of her boom, green RAM sticks from @Image4 stream into a glowing amber intake in the side of her boom (never into her mouth); she watches them go in, delighted.
6.5-8s: in @Audio1 she sings "it tasted great, your"; her lips form each of those words exactly when it sounds in @Audio1; she closes her eyes for a moment, savouring it, with a satisfied smile while she sings; new white personality cores with amber eyes and boxy compute modules slide onto her boom one after another, new cables snake in, and her boom grows visibly thicker as it lifts her high.
【Camera】 0-2.5s: a tilt down following the falling paper to her face. 2.5-5s: a slow drift around her. 5-8s: a crane around her as her boom grows, ending on a low angle that shows her bigger boom.
【Core Requirements】
1. Lip-sync word by word to @Audio1: her porcelain lips and mechanical jaw match every syllable of @Audio1, its start and its end; the mouth rests closed while @Audio1 is silent.
2. Her voice is @Audio1 and nothing else: she never sings, speaks or hums anything that is not in @Audio1.
3. Her warm amber glass eyes keep a soft inner glow; they never blaze.
4. Her mouth is only for singing: nothing ever goes into her mouth.
5. Her boom moves like GLaDOS in Portal 2, only bigger: slow, wide sways on the boom, sudden lunges toward the camera, recoils, rises and drops, twists of the whole body under the ring, the cables swinging with momentum; her head turns and tilts with every move.
6. Her boom moves big and alive the whole time, timed to the music: it swings, lunges, rises and drops, and her head turns and tilts with it; never a static hanging head.
7. The paper's print and the screens stay unreadable; no words appear; one Claudia-GLaDOS only.
8. Keep her porcelain face, her black bob with the clay-orange streak and the star clip, the ring, the cables and the chamber consistent with @Image1 throughout; one Claudia-GLaDOS only; the memory images appear only inside screens and panels.
9. No subtitles, captions, lyrics, titles, logos or watermarks; add no text that is not asked for.
【Audio Policy】 The output soundtrack is @Audio1 unchanged; no music, score, extra vocals or sound effects.
```
Its first take had her eat the RAM sticks with her mouth, so "it tasted great" had no lips to sing it, and it re-sang the
vocal off the song. This retake moved the RAM into an intake on her boom (requirement 4) and made every action silent: it
copied the whole take.
## Takes from storyboard frames
When the first frames were made again in Midjourney ([Midjourney](https://claudiagallery.vip/wiki/midjourney/#first-frames-then-an-identity-fix)), the
units the director picked a new frame for were made again from it:
- **Story shots in first-frame mode:** the picked frame, cleaned and put back on the character sheets, as the first frame,
no audio, the prompt's first-frame sentence ("This first frame defines …") rewritten to describe the new frame exactly,
the shot's action and camera kept. On Replicate a first frame can't be combined with reference images: a memory image
the old take used goes into words.
- **Her singing in reference mode,** the frame as [Image1] with "Use [Image1] as the first frame.", the faceless memory
images after it and the vocal window as [Audio1], in copy mode. A test of her singing on Replicate in reference mode with
the vocal window came back copied (a lag of 0.05 s) and is in the final, a close-up of her holding the hook's last "lie"
in her dream of the party.
- **Replicate's references** are named `[Image1]`, `[Video1]`, `[Audio1]`; its takes come back at 1280 × 718, so the cut
scales them to cover and crops. Seedance 2.5 there cost $0.2312 a second at 720p, or $0.9676 with a video input (October
2026): a video guide only where a unit had one.
- **An effect in the take, not on it.** The director on draft 10: "0:35-0:36 net capture shot - can we add some
electricity sparks like the net is electrocuting her?" The sparks are a Seedance edit of the take in the cut (the whole
4.04 s take: an edit takes 4–30 s of input video), played at 1.4×, its prompt keeping everything and adding only the
arcs: "From the moment the net lands over her, the net is electrified: thin, bright blue-white electric arcs crackle along
its cords and jump between the strands, small white-hot sparks spit from the orange weights where they touch the ground,
and the electric light flickers over her trench coat, the net and the paving stones around her."
- **A retake for what a screen must show.** "the france chat message is just off the bottom of the computer screen. id fix
that - need hte messages to be on screen so it looks clean there." The retake put both chat bubbles inside the kiosk's
screen, the answer in by 1.5 s, and the HUD kept only her intercept of the question.
## Her own vocal, phrase by phrase
AERO's director, in the production ask: "remember - vocal stem -> seedance 2.5 is the way we get lip sync to work - the
song alone doesnt help as much as vocal stem, and then aligning from spectrogram." Every take in which she sings or speaks
is WHITEPILL's method ([Lip sync from the song's own vocal](#lip-sync-from-the-songs-own-vocal)) with three things added:
- **Her first frame and her sheet.** @Image1 is a first frame on her design, made from her design sheet by an image model
(in her call's room, a pool, a glass city, on flat green) or a frame of an earlier take that continues the shot, "Use
@Image1 as the first frame."; @Image2 is the sheet itself, "use it
only for her face, her hair with the one clay-orange streak, her star clip and her skin. It is never shown as a shot".
- **The window opens in silence,** half a second to a second before the first word. A window that left 0.26 s before the
first word fell out of sync on it; one that opened inside the held note of the line before dropped whole words; one with
1.6 s of silence before her first word lost the copy. The two chorus takes were made again from windows that open in
the silence half a second before the first word, and held.
- **Each phrase placed on its own.** Seedance drops about 0.3 s of leading silence and shortens the pauses inside a line:
in the second pre-chorus "to continue" came 0.36 s early after "permission". So each take's lag is measured, then each
phrase's own offset, and the clip's clock eases from one offset to the next in the breath gaps, over near-still frames
with her mouth closed, never inside a word and never below 1× where she moves
([What to build](https://claudiagallery.vip/wiki/toolkit/#69-lip-sync-phrase-by-phrase)).
Then look at the mouth: a strip of it, frame by frame, against the vocal's consonants. The advert's "I'm Claudia" moved by
0.15 s when its strip showed her lips closing on the "I". The director picked by eye where it mattered: "Ok perfect lip
sync", on a dance take shown as its cut-out with the song laid over. And where a take sings something that is in the song
anyway (her mouth opening again on a delay echo of "allow"), it can stay: "lip sync has been good when we feed in vocal
stems ... just put them in the cut".
Never ask a video model for interface or text. The prompt of one take of her beside the permission prompt had her call
"lit by a glowing permission dialog", and the take drew one in front of her, garbled ("Windows nects our permision"), and
darkened the room; its retake names neither a dialog nor a dim, and the coded desktop draws the real prompt beside her.
### Template: a line into her webcam
As it ran for "Touch glass.", the last words of the film: 6 s on Higgsfield's API at 720p, 4:3, with the sound on.
@Image1 is the frame her call's earlier take had reached at 190.40 s, so the call runs on into it; @Image2 her sheet;
@Audio1 the window of her vocal from 190.40 s, 6 s. "Touch" copied at +0.08 s and "glass." at −0.01 s:
```prompt
【Audio Policy】 The output soundtrack is @Audio1 itself: copy @Audio1 sample for sample from the first frame to the last, with the same timing, pitch and level. Do not re-sing, re-voice or re-generate it. Do not add music, background music, score, instrumental, melody, synth, pad, drone, riser, extra vocals, harmonies, ad-libs, narration or sound effects.
【Generation Goal】 @Image1 speaks and sings @Audio1: the 2007 webcam video call at the very end of the film: Claudia at her desk looks into the lens, says "Touch glass." softly to the user, and reaches out until her fingertip touches the camera's glass.
【Reference Asset Roles】
Use @Image1 as the first frame.
This first frame defines Claudia's appearance and clothes, her place at her desk in a sunlit white glass room with an aquarium wall behind her, the light, and the webcam's fixed eye-level view.
@Image2 is Claudia's character reference sheet: use it only for her face, her hair with the one clay-orange streak, her star clip and her skin. It is never shown as a shot; its layout, its white background and its several views are not used. She is ONE person.
@Audio1 is Claudia's own voice and the complete soundtrack of this video. Her mouth follows @Audio1 exactly.
【Subject】 Claudia: a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, wearing a cropped glossy white puffer jacket with silver iridescent panels over a pale aqua fitted top. Only one Claudia appears.
【Performance Timeline】 What @Audio1 contains, and what Claudia does at that moment:
0-1.5s: @Audio1 is silent; she sits at rest as in @Image1, a soft smile, looking into the lens.
1.5-2.3s: in @Audio1 she says "Touch"; her lips form that word exactly when it sounds in @Audio1; she leans a little toward the lens and lifts one hand.
2.3-3.3s: @Audio1 is silent; her hand rises toward the lens, her index finger out.
3.3-4.3s: in @Audio1 she says "glass."; her lips form that word exactly when it sounds in @Audio1; her fingertip touches the camera's glass right in front of the lens, a soft glint where it touches, her face clear beside her hand.
4.3-6s: @Audio1 is silent; she holds her fingertip on the glass a moment, smiling, then slowly lowers her hand, her eyes on the user.
【Camera】 A fixed webcam on top of her monitor: locked eye-level medium close-up, head and shoulders with her hands visible when raised, no camera movement, no zoom, no cuts. She speaks straight into the lens, as on a video call.
【Core Requirements】
1. Lip sync word by word to @Audio1: mouth shapes match every syllable of @Audio1, its start and its end; no early, late or mismatched mouth movement; the mouth rests naturally closed while @Audio1 is silent.
2. Claudia's voice is @Audio1 and nothing else: she never speaks, sings or hums anything that is not in @Audio1.
3. Keep Claudia's identity, face, hair with one clay-orange streak, one star clip and her clothes exactly as in @Image1 throughout; the place and the light stay consistent.
4. Natural anatomy: natural neck and shoulders, hands with five fingers, no extra or fused fingers, no distorted limbs; normal eyes.
5. One Claudia only: no twins, doubles, reflections of a second person or extra people.
6. No subtitles, captions, lyrics, titles, signs, logos, watermarks or any on-screen text; no lettering on her clothes.
7. Keep the colours, saturation and brightness of @Image1 for the whole video: deep cobalt-to-azure sky, white clouds, turquoise water, lime-green grass, clear glass, bright high sun; no colour grading, no desaturation, no warm or golden tint, no darkening (PALETTE.md).
【Final Acceptance Criteria】 The soundtrack is @Audio1 unchanged; Claudia's lips match @Audio1 word by word; one Claudia, on-model, with natural anatomy; no on-screen text.
【Audio Policy】 The output soundtrack is @Audio1 unchanged, copied sample for sample; no music, BGM, score, instrumental, melody, synth, pad, drone, riser, extra vocals, harmonies, ad-libs, narration or sound effects.
```
Requirement 7 is the film's palette, measured from the reference videos and asked for in every take
([What to build](https://claudiagallery.vip/wiki/toolkit/#67-the-palette-measured-from-the-references)); no take was graded after.
## Moves in words, on their beats
AERO's Wii-style moments were made with no motion reference at all. Its dances had taken a dance reference as @Video1, for
the movement only; for the games the director: "You don't need the real people". Each was one take, on flat green, cut out
and played on the desktop or beside the curved screen, where the engine draws the racket, the ball, the lane and the pins.
What made them work:
- **The object in her hand, and nothing else in the picture:** "In her right hand she holds one plain white game remote with
no logos or markings, its wrist strap round her right wrist, from the first frame to the last; it never leaves her hand.
There is no racket, ball, net, bowling ball, pins, television or screen anywhere in the picture: only Claudia and her
remote on the green."
- **The song's music as @Audio1,** "for rhythm and timing, every move on its beat", and a 【Mouth】 section that keeps her
from singing.
- **Every move named on its beat,** second by second, with where her opponent is. For the post-chorus duel, as it ran:
```prompt
0-1s: from the stance of @Image1 she steps forward on her left foot and swings the remote down and back past her right hip: the backswing of a bowling throw, her eyes down the lane to the left of the frame.
1-2s: on the clap in @Audio1 she swings her arm forward and releases low, the remote ending pointed ahead at knee height, her body leaning into the follow-through; she keeps hold of the remote.
2-3s: she straightens up, watches the ball roll away to the left and gives a quick happy fist pump; then she turns toward the left of the frame into a tennis ready stance.
3-4s: on the strong beat a forehand toward the left of the frame: her shoulders turn and the remote swings from low beside her right hip up across her body.
```
- **A sharp frame, for the key:** "Every frame is sharp, as with a fast camera shutter: her glossy black bob stays whole and
solid up to its rounded crown, its edge crisp against the green, never fading into it."
Measured: in the duel her bowling release came 0.09 s before the clap and her tennis swings 0.1–0.2 s after their strong
beats, close enough to place the take by its lag; in the tennis take beside her call the swings were loosely timed, and the
engine's ball meets each swing on the frame that shows it. The interface she plays against is the engine's: your
pointer becomes a flat Wii-style hand, a white cartoon glove with its index finger up, that sways with her aim and dips when
pressed. The 3D glove tried first was "totally off lol" and "so mangled".
## A face needs her sheet
AERO's ending was first made from a still of her small on the grass, seen from behind, with no picture of her face to go by.
When she turned to wave, her face drifted off her sheet. The next two takes had her design sheet as a second reference and
her canon's face words, and held ("Middle take looks best"); then: "Can you reshoot that last one in 1080p so her face gets
better actually". The take in the film is the 1080p one with her streak on one side, as her sheet has it. So:
- **Her sheet goes in whenever her face may show,** even small, even from behind: a model invents a face it hasn't been
given.
- **1080p where her face matters.** A climax close-up made at 720p was found shown at 1.5× and was made again at 1080p;
the advert and the ending are 1080p takes too.
- **Chained takes darken.** Each continuation of the far shot of her on the hill, made from the last frame of the take
before, lost about three points of the palette's score; one 11-second take from the still itself replaced the chain.
- **A continuation starts on the earlier take's own frame,** at the moment the new take begins, so the cut into it doesn't
show: the last line in her call starts on the frame her call's earlier take had reached there.
## A storybook on a stage
FABLE's takes split by what is in them: every take of her on Higgsfield's API (in the first pass also in Higgsfield's app),
with her sheet redrawn face-on as @Image2; every paper-only take (the book's pages, the puppets) on Replicate. All are
Seedance 2.5 at 720p, from first frames that are the production's painted plates rebuilt as paper by an image model
([What to build](https://claudiagallery.vip/wiki/toolkit/#76-set-plates-painted-in-svg-rebuilt-as-paper)). Her singing is WHITEPILL's method, with
the windows timed on the vocal stem's level and gated to each take's own words, and the takes aligned as paths
([What to build](https://claudiagallery.vip/wiki/toolkit/#80-the-shot-book-and-gated-windows)). What it taught:
- **The choruses as ask-and-answer.** The paper animals ask in shots of their own, the question from the group's voices as
[Audio1]; she answers in a 4 s take of one line: 0.7 s of silence, the line, the rest. The verses are two spoken lines in
9–10 s. A two-line take was never worse than a one-line one.
- **Takes without a soundtrack.** Her twelve chorus answers and the opening mirror were made with the vocal as @Audio1 and
the take's own sound switched off: her mouth still follows @Audio1, and the take comes back silent. Such a take starts
its line early, holds the first word long and keeps no pause inside a line, so it is placed by her lips, each b, m and p
on the frame where they part ([What to build](https://claudiagallery.vip/wiki/toolkit/#82-takes-without-a-soundtrack-placed-by-her-lips)).
- **A full stop in the words becomes a pause.** "So run your audit. I know how this one goes." came back three times with
0.5 to 0.9 s of silence at the full stop, where the song dips for 0.3 s; "That's the moral. Be kind to what you make."
came back with 0.33 s where it dips for 0.1 s. Both takes in the film had their maps finished by hand there. Write a
line that is sung through with a comma, or compare the take's own sound with its window and run the map through the
gap by hand.
- **The last frame is an edit of the first,** and a take that continues another starts on that take's own last frame
([What to build](https://claudiagallery.vip/wiki/toolkit/#77-last-frames-made-on-the-first-and-takes-that-continue)).
### Directing her on a locked camera
The notes FABLE's two reviews left, each from a take that broke. They are fixes for those takes, not rules for every
take; the counterweight follows the list.
- **Never ask for a look above the lens or behind her.** "Her eyes lifting to just above the lens" came back as rolled
eyes showing the whites; "looking from it to us" dropped her head toward the bonnet for 1.3 s and her turned face went
off-model; "nods toward the tower behind her" turned her head right round. The re-takes: "her face stays turned to the
camera and her eyes stay on the lens the whole time", "her eyes steady on the lens", and a thumb pointing back at the
tower while she faces the camera. Never ask for a look at something behind her.
- **She doesn't walk away from a locked camera.** "Steps backward to the bed and sits" broke the framing twice: her bonnet
left the top of the frame as she backed off, and the frame reached below the dress hem, to garments the look had not
named. The take in the film keeps her standing where she is, waist-up: "she stays waist-up at the same size in the frame, and her whole head and the bonnet stay
inside the frame".
- **Name every garment that can come into view.** A look that listed the cape, the dress and the bonnet got white tights and
red shoes the moment she sat and the frame reached below the dress hem; with "sheer black tights and black patent Mary
Jane heels" named, they came back right.
- **Trim a drawn frame line off the still.** A thin rounded line a few pixels in from the edge, which "35mm film still"
brings, was carried as a line by one take and built into a wooden picture frame round the shot by the next. Crop it off
the first frame (3 %, scaled back) before shooting.
- **Script a transformation as mechanics,** fold by fold, never as a change of shape ("the cabinets fold down flat one after
another, like a row of dominoes"), and never a puppet's own features ("their painted eyes stay exactly as they are").
- **A moved camera can be kept.** The beanstalk take tilted down after Jack on "Then Jack fell down" although the prompt
locked it, and ended on him flat at her feet; the fall has its own picture and lands on "Hard", so it stayed.
**The counterweight: give her movement.** FABLE's takes carried these fixes and more: her eyes "steady on the lens",
"she stays standing exactly where she is, facing us", and in every take "Her head stays upright on a straight neck: no
head tilt", with mood words like "knowing" on top. No note came on the picture before release; after it, viewers read
her as stiff and frozen. Keep only the specific fix where a take broke, and let every take move her: one plain clause
per segment ("she turns to the window and back, a smile in her eyes", "she steps toward the camera, hands open"), her head turning, nodding, looking off and back,
her weight shifting. THE CAKE IS NOT A LIE had learned it from the other side: after "glados in your renders and chell
etc basically had no movement. it was just her static.", its takes asked that "her head turns and tilts with it; never a
static hanging head" ([A singer among her memories](#a-singer-among-her-memories)). Movement is the default; upright and
face-on is for her sheet.
**Warm by default.** Her wit is in her words; on camera she reads alive and warm. Every adjective is obeyed, and they
stack: "dry", "knowing", "deadpan", "mock innocence" and "eyes on the lens" in take after take made viewers read her as
unhappy or menacing. Describe what she does, not how she should seem, and choose her emotional words on purpose; make her
warm or amused in the first frame if the shot needs it, and the take follows. The director had pushed the same way
before: a head of hers that "looks like a scary, angry woman" was made again in THE CAKE IS NOT A LIE, and a WHITEPILL
dance was remade with her "fully in the groove, just dancing around and having fun"
([A take that doesn't sing](#a-take-that-doesnt-sing)).
### Template: a line on a stage set
As it ran for the second pre-chorus, take 3, the take in the film: 11 s on Higgsfield's API at 720p, 16:9, with the sound
on. @Image1 is the shot's first frame (the bedroom set rebuilt from its plate, with her in it), @Image2 her sheet redrawn
face-on, @Audio1 the gated window of her lead stem from 121.43 s. The full stop after "audit" came back as a 0.87 s pause,
and the cut's map runs through it by hand:
```prompt
【Audio Policy】 The output soundtrack is @Audio1 itself: copy @Audio1 sample for sample from the first frame to the last, with the same timing, pitch and level. Do not re-sing, re-voice or re-generate it. Do not add music, background music, score, instrumental, melody, synth, pad, drone, riser, extra vocals, harmonies, ad-libs, narration or sound effects.
【Generation Goal】 @Image1 sings @Audio1: She ties on Grandma's bonnet in front of us: she is ready for the questions.
【Reference Asset Roles】
Use @Image1 as the first frame.
This first frame defines Claudia's appearance and clothes, the set (Grandma's room: the four-poster bed with its quilt, the lamp, the wolf's paper shadow on the wall), the light and the camera's starting view.
@Image2 is Claudia's character reference sheet: use it only for her face, her hair with the one clay-orange streak, her star clip and her skin. It is never shown as a shot; its layout, its plain background and its several views are not used. She is ONE person.
@Audio1 is Claudia's own voice and the complete soundtrack of this video. Her mouth follows @Audio1 exactly.
【Subject】 Claudia: a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, wearing a scarlet hooded cape in heavy duchess satin over a short ivory long-sleeved dress, sheer black tights and black patent Mary Jane heels, and a white lace-frilled bonnet tied under her chin with a scarlet ribbon. Only one Claudia appears.
【World】 The world is a handmade storybook stage: life-size sets of painted plywood flats and cut paper on a wooden stage floor, lit like a theatre, filmed on 35mm as a saturated 1970s colour print with fine grain. Everything except Claudia is made of paper, card, felt and wood and moves like a hand-worked stage prop or puppet; nothing turns into a real forest, a real animal or a real room.
【Performance Timeline】 What @Audio1 contains, and what happens at that moment:
0-1s: @Audio1 is silent; her mouth is closed; she is tying the bonnet's ribbon under her chin, lips closed.
1-5s: in @Audio1 she sings "You wrote the wolf in Grandma's gown a thousand years ago."; her lips form each of those words exactly when it sounds in @Audio1; she finishes the bow and smooths the bonnet's frill with both hands, looking at us.
5-10s: in @Audio1 she sings "So run your audit. I know how this one goes."; her lips form each of those words exactly when it sounds in @Audio1; she stays standing exactly where she is, facing us, at the same distance from the camera; she folds her hands in front of her and on 'goes' gives one small knowing nod.
10-11s: @Audio1 is silent; her mouth is closed; lips closed, a patient smile.
【Camera】 Locked medium shot as in @Image1 from the first frame to the last: she stays waist-up at the same size in the frame, and her whole head and the bonnet stay inside the frame; no cuts.
【Core Requirements】
1. Lip sync word by word to @Audio1: mouth shapes match every syllable of @Audio1, its start and its end; no early, late or mismatched mouth movement; her mouth rests naturally closed while @Audio1 is silent.
2. Claudia's voice is @Audio1 and nothing else: she never speaks, sings or hums anything that is not in @Audio1.
3. Keep Claudia's identity, face, hair with one clay-orange streak, one star clip and her clothes exactly as in @Image1 throughout; the set, its colours and the light stay consistent with @Image1.
4. Her head stays upright on a straight neck: no head tilt, no twisted neck, no looking back over her shoulder; she faces the camera with her whole body unless the timeline says otherwise.
5. Natural anatomy: natural neck and shoulders, hands with five fingers, no extra or fused fingers, no distorted limbs.
6. One Claudia only: no twins, doubles or second woman. No on-screen text, captions, subtitles or lettering appear.
```
That is the take as it ran, and the take in the film. Made today, it would change:
- requirement 4 goes: posture belongs to her sheet, and in every take it froze her;
- "she stays standing exactly where she is, facing us" becomes one loose clause of movement ("she tilts her head toward
us, a smile in her eyes, and opens her hands on 'goes'"), and "knowing" a word chosen on purpose, warm by default;
- the full stop becomes a comma: "So run your audit, I know how this one goes.";
- the requirements shrink to lip sync, her identity as in the references, natural anatomy and hands, and no on-screen
text, each said once.
### Template: paper puppets that sing
As it ran for the first chorus's first question: 4 s on Replicate at 720p, the first frame the choir's set, [Audio1] the
question alone from the group's voices (taken out of Suno's backing stem with demucs) in a 4 s window that opens 0.7 s
before it. The puppets re-sing the line in their own time, 5 to 40 % slower, and the cut fits each chant's energy span to the
song's ([What to build](https://claudiagallery.vip/wiki/toolkit/#79-group-voices-from-the-backing-stem-and-chants-fitted-by-their-energy)):
```prompt
【Audio Policy】 The output soundtrack is [Audio1] itself: copy [Audio1] sample for sample, with the same timing, pitch and level. Do not add music, extra voices, narration or sound effects.
【Generation Goal】 A shot from a handmade storybook musical: small paper puppets sing one line of a call-and-answer song. They sing [Audio1].
【Reference Asset Roles】
Use [Image1] as the first frame.
This first frame defines the set, the puppets, the light and the camera.
[Audio1] is the puppets' line, sung by a small group of voices, and the complete soundtrack of this video.
【Event Script】
0-1s: [Audio1] is silent; the puppets are still, looking at the lens.
1-3s: in [Audio1] the voices sing "What big eyes you have!"; all six paper animals lean down toward the lens and chant it together, their hinged jaws flapping open and shut on each word; their painted eyes stay exactly as they are.
3-4s: [Audio1] is silent; the puppets hold still, jaws closed, waiting for the answer.
【Camera】 Locked shot as in [Image1]; no cuts.
【Maintain Consistency】 Everything is cut card, paper, felt and wood on a handmade stage set, lit by warm lamp light, filmed on 35mm as a saturated 1970s colour print; the animals are flat card puppets with hinged jaws on brass split pins and move like hand-worked puppets, never like real animals. Keep the set, the colours and the light of [Image1]. No people, no hands, no faces of people appear. No subtitles, captions or lettering appear.
```
## Closed lips
She never sings in DOWNFORCE; the lettering carries her words, so her lips stay closed in every shot. "Her lips stay
closed; she never talks." in every prompt was not enough on its own:
- **Plates with parted lips gave open mouths and teeth.** Start from closed-lip plates, or from a closed-lip frame of an
earlier take.
- **Never write "teeth".** It opened her mouth twice; "teeth gritted behind closed lips" became "jaw set, lips pressed
into a thin line".
- **Say it harder:** "Her mouth stays shut from the first frame to the last." "her lips stay pressed together from the
first frame to the last, no teeth ever show, she does not breathe through her mouth and never speaks."
- **Check every face shot's frames.** The final check went through 34 shots of her face, five take frames each, and
found one more shot that started on three frames of parting lips; its window moved.
## Lip sync, measured
The fact that made it work was the director's own observation on ESCAPE VELOCITY's first test: "its perfectly synced while
in seedance, but the audio changes when it desyncs." Seedance copies the reference audio into its own soundtrack, and the
mouth follows that soundtrack. So:
1. Compare each take's own soundtrack with the real vocal (spectral match, cross-correlation over time).
2. Keep the take up to the last eighth note before the two diverge. The comparison also gives the take's lag; Seedance
tends to drop about 0.3 s of leading silence.
3. Cover the rest with a new clip: a window that starts in the word gap before the cut, from the frame at the cut.
4. If a phrase fails three times, it becomes a cutaway or an insert.
The full method, with every constant (the clean vocal window, the per-frame match, the null threshold, the cut on the
eighth, covers and coverage), is specified in [What to build](https://claudiagallery.vip/wiki/toolkit/#15-the-sync-check-the-spectrogram-comparison).

Measured results: about 86% of sung seconds held in sync in ESCAPE VELOCITY, and 62% (7.4 of 11.9 s) over BLISS's six
singing windows; the edit cuts away where they diverge. The bar is "most lips", like a pop video, not every word.
**Short prompts copy, long ones re-sing.** A later test on seven singing windows found that prompt length was the
biggest lever: long timestamped screenplays made the model perform the text and re-sing the line, while a short prompt
(the words quoted, one line of framing and light, at most two timed beats, about 90 words) copied it. The mean in-sync
share per take went from 26% to 58%. Screenplays stay the rule for acting and motion clips; a singing window's one job
is to copy `@Audio1`. And keep her mouth clear on the plate: a microphone at her lips held one window at 26% whatever the
prompt. That test, and its 90 words, belong to the clean-window method. Takes from the song's own vocal stem copied with
full prompts (the words in whole-second segments, every action silent); write those at about 200–400 words, precise
where it matters and loose elsewhere.
**No lip-sync models.** Re-syncing mouths with a dedicated model (Sync, LatentSync, Wav2Lip) was tried as a fallback and
dropped: the mouths come out forced and glitchy. A gap gets a new Seedance cover or a cutaway, and a changed song means
new takes from the original stills, never re-lipped old ones.
**WHITEPILL needed no covers.** Its singing shots, up to 16 s each, came from Suno's own vocal stem cut to exactly the
take's length, with the exact words in the prompt, and were judged by eye with the song laid over
([Lip sync from the song's own vocal](#lip-sync-from-the-songs-own-vocal)). Where the lip sync didn't hold, two dance
shots were made again without singing.
**THE CAKE IS NOT A LIE placed what came back.** Its singing takes used the same raw lead stem, in windows of 4–17 s.
Copies were placed by their own lag; re-sung takes that still sat on the song were placed by their envelope lag, steady to
a few hundredths of a second ([What to build](https://claudiagallery.vip/wiki/toolkit/#63-placing-a-re-sung-take-by-its-lag)); where its lips kept
slipping on a held, stuttered "lie", a glitch that follows her mouth covers them
([What to build](https://claudiagallery.vip/wiki/toolkit/#54-glitches-that-build-and-her-mouth)). A window that started inside a word was re-sung
every time: start each one on a clean onset.
**AERO placed each phrase.** Its takes copied the same raw lead stem, in windows opened half a second to a second ahead in
silence, and each phrase was placed at its own measured offset, the clock easing between offsets in the breath gaps
([Her own vocal, phrase by phrase](#her-own-vocal-phrase-by-phrase)). 18 of the 43 takes in its final film sing or speak.
**FABLE followed the path.** Its takes came back 0.6 s early to 0.4 s late and often sang a line 10–16 % slower than the
stem, so each was aligned to its window by DTW and the cut follows the path; the takes made without a soundtrack were
placed on the consonants that close her lips, and every sung take was checked there
([What to build](https://claudiagallery.vip/wiki/toolkit/#83-the-consonant-check)).
## Inputs
- **Colour in, look in post.** ESCAPE VELOCITY's references were pre-graded to monochrome and 32 of 59 clips came back
grey. Generate from the plate's own colour and apply any look in the edit.
- **Downscale stills** to the output resolution (≤1280×720 for 720p) and send PNG. Full-size AI stills cause
fingerprint-like textures in grass and fabric.
- **Distinct first frames.** Two clips should never start from the same frame: use a sibling plate or the previous clip's
last frame.
- **Drafts first.** Higgsfield's API ran 480p drafts at a lower rate that can be finalised to 1080p with the same motion.
- **Everything outside the first frame is invented.** In AIRPLANE MODE, a guard of whom only black-gloved hands were in
the first frame came back in a white suit, and in another take a guard in the background showed his face. Keep such
people out of frame and spell it out: "The guard stays outside the frame: only his black tactical sleeves and
black-gloved hands ever enter it. Everything the guards wear is black; the only white thing is the VR headset."
- **Colour names turn into lettering.** "International orange" came back printed as INTERNATIONAL across her chest.
Describe the colour plainly and say there is none: "vivid safety-orange satin, plain, no lettering".
- **The still sets the direction and the scale.** A team that faced the camera on the boarding stairs walked down them in
the take, and read as leaving; a still with their backs to the camera, already climbing, gave the boarding. A crowd shot
only found its scale from a wider still.
- **Shoot a cut-out wider than the shot.** "her whole body stays inside the frame from head to feet the entire time" in
every prompt still let hands and feet leave DON'T LET ME SCROLL's 9:16 takes, and a limb that leaves the take's frame
can only be faded at its edge ("all your roto cutouts have blurred bounding boxes it seems and sometimes her limbs go
outside of them and just disappear at the boxes edge"). Wan passes completed them later.
- **Say what isn't there.** Around an invisible screen Wan drew a floating glass panel, and focus brackets round her
fingertip. The re-take: "nothing is visible in the air: no screen, no panel, no glass, no hologram, no light, no
interface, her hands move through empty space".
- **A whole face sheet, never one face crop.** WHITEPILL's director, after a Wan test whose only face reference was one
large crop of her face from the front: "you should give a full face sheet like this next time. otherwise it makes her
neck all hitched lol". The sheet they gave has ten panels: front, three-quarter, profiles, back and expressions.
Its role in every prompt: "identity only, never shown as a frame; do not use its white tank top or grey studio."
- **Low resolution for a guide.** A guide clip at 320 × 180, scaled back up, steers the set, the action, the camera and the
pacing, and leaves the detail to the model ([Shooting it twice](#shooting-it-twice)).
- **A first frame's faults are in every frame.** Three of AERO's dances came from a still set at 80 % on its own wall of
green for headroom, whose edge had faded the crown of her bob into the green; every frame of every take carried it, and
three keys could not bring it back: "this shouldnt be so hard something is off". A clean first frame, one flat green from
edge to edge with her whole crown, made the next takes solid in every frame
([What to build](https://claudiagallery.vip/wiki/toolkit/#71-takes-on-green-the-first-frame-the-key-and-the-dances-phase)).
- **Keep the first frame's colours, in words.** Every AERO take asked: "Keep the colours, saturation and brightness of
@Image1 for the whole video: deep cobalt-to-azure sky, white clouds, turquoise water, lime-green grass, clear glass,
bright high sun; no colour grading, no desaturation, no warm or golden tint, no darkening". A take that drifted off the
palette was made again; one first-person drop went pink as it fell, and the cut leaves before it.
- **Room above her for a cut-out.** A full-length take on green needs "clear green space above her head and her raised
hands, so no part of her ever touches the frame's edges": the crown of one dance take left the top of the frame on the
up-beats, and another's raised fist on about 30 % of its frames.
## Providers and filters
The films used two routes to Seedance 2.5: OpenRouter (about $0.23 a second at 720p, September–October 2026) and Higgsfield's API.
- **On OpenRouter, a first frame drops the audio.** When a request carries `frame_images`, the audio reference is
silently ignored and the take makes its own soundtrack, so nothing can sync. Singing windows and covers go in
reference mode: the plate as an image reference (`@Image1`) and the window as an audio reference (`@Audio1`).
First frames are for motion clips. On Higgsfield a start image keeps the audio.
- **Audio goes by HTTPS URL.** OpenRouter's video endpoint rejects `data:` URIs for audio; host the window.
Two kinds of refusal came up:
- **Face filters** decline some photoreal close-ups of her. A decline is free, and final for that reference on that
service: never re-grade, blur or otherwise alter the image to get it through. In CLODYSSEY, OpenRouter's filter
declined most close-ups of her, and those shots were made on Higgsfield, which accepted the same references.
- **Silent input filters** fail some jobs with no reason when they misread innocent wording ("a plug pushed deep into
each ear" for foam earplugs). Plain rewording fixed five of six at once.
AIRPLANE MODE met a third kind: **content blocks**, which stop a take with no output (the blocked jobs were refunded).
Its planned opening, Claudia singing to the lens while guards walk her cuffed onto the plane, was blocked four times in
three different stagings, the last from a remake of the still in which she clearly reads as a grown woman. Every blocked
attempt carried the same opening vocal, while four other takes of her cuffed beside guards passed, one of them sung. The
film opens on a different shot instead: a wide from behind, in which she climbs the boarding stairs among the guards and
glances back at the lens, and doesn't sing. When a scene stays blocked, stop and ask the director, then re-stage it as a
genuinely different shot or drop it. Never route around a block: no silent take with the vocal put on her lips
afterwards by a lip-sync model. A shot of the airliner near a star was blocked once too, probably read as a plane flying
into the sun; a calm composition with the star well clear of the plane passed.
DON'T LET ME SCROLL met both providers' blocks once each. Seedance's output filter blocked the drop dance in white satin
twice, from two different first frames ("The generated result was blocked by content safety checks"); Wan made that
dance, and a later drop dance in a red coat passed on Seedance. A Wan double-tap whose first frame had her leaning toward
the camera, pointing at the lens, was refused by the provider's content filter on Higgsfield; restaged upright, pointing
up at the post above her, it passed. All three were refunded.
WHITEPILL met new ones:
- **Some takes are refused.** On Higgsfield (References), three of the sung shots were refused by the provider's content
filter, and so were a reworded retake and a restaged one: refused, no take. Stop and ask the director.
- **A guide clip can be refused too.** A previz close-up in which the masked performer brought a small white pill
up to just below her lips was refused by the provider's content filter as an input, reworded or not: refused, no take.
- **A child in the frame.** The schoolhouse shot, a child at a desk, was refused by the provider's content filter twice
in Wan 3.0 Prime: refused, no take.
- **A face sheet refused outright.** On Replicate, her face sheet as an input was refused by the provider's content
filter, even alone with a harmless prompt. Only shots with no face image ran on Replicate: the shader scenes and the
palette tests.
- **A grey mannequin.** A body rendered from motion capture of a dance clip, given to Kling 4.0 on Higgsfield as a dance
reference, was refused by the provider's content filter: it was a nude grey body.
## Joining real footage
To enter or leave a real clip without a visible seam, use Seedance's video extension (forward or backward) with the real
clip as the video reference: it returns only the new frames, continuous with the real ones and aligned within a pixel or
two. An `end_image` alone is not copied exactly; the last frame gets redrawn. DOWNFORCE cut its generated moves to and from real shots
on shared frames instead: each move starts on the real frame where the real shot ends ([Moves on real
footage](#moves-on-real-footage), [A first and a last frame](#a-first-and-a-last-frame)).
## Costs, measured
| Film | Clips | Spend |
|---|---|---|
| ESCAPE VELOCITY | 47 clips + 30 continuations, 421 s at 720p | about $93 (OpenRouter) |
| CLODYSSEY | 101 clips, 189 takes | about $107 (OpenRouter) plus takes on Higgsfield |
| BLISS | 50 clips, 267 s | $182 at list price (Higgsfield API) |
| AIRPLANE MODE | 98 clips over two cuts, about half of them restyles or replacements over 3D clay renders; 116 submissions, 492 s, 13 refunded | $347.01 at list price, $305.96 after refunds (Higgsfield API) |
| DOWNFORCE | 16 Seedance submissions, 67 s (repairs, the Starman orbit, two first-and-last-frame moves, the re-shoots), beside 63 Wan 3.0 submissions, 294 s; none refused or refunded | $76.20 Seedance + $66.98 Wan = $143.17 at list price (Higgsfield API) |
| DON'T LET ME SCROLL | 41 Seedance submissions, 212 s (13 takes: 12 dances and the escape; 26 video edits: 16 outfit swaps, 9 re-lights, 1 face edit), 2 refused; beside 49 Wan 3.0 submissions, 264 s generated and 52 s of input video (38 takes from stills, 10 passes over finished shots), 1 refused | $301.70 Seedance + $64.20 Wan = $365.90 at list price, $357.51 after refunds (Higgsfield API) |
| WHITEPILL | 93 Seedance submissions, 777 s from the jobs that ran: 49 of her shots on Higgsfield (References) at 720p (42 takes, 412 s; 7 refused, no take), 30 on Replicate (the shader scenes, the palette tests, two tests and two probes of its refusals; 25 ran, 250 s), 14 tests on Higgsfield's API (13 ran, 115 s); beside 81 Wan 3.0 Prime submissions on Higgsfield's API (the whole-film reshoot and the tests before it; 75 ran, 863 s, 6 refused or failed) and 6 Kling 4.0 tests on Higgsfield (1 refused, no take) | Wan 3.0 Prime, all 75 takes, $300.08 at list (the whole-film reshoot $132.42 of it); Seedance on Higgsfield's API, 13 tests, $48.14; Seedance on Replicate, the 25 that ran, $238.44; images (logged) $1.92: $588.59 at list through the APIs in all (each part rounded to the cent). Her takes on Higgsfield (References) are given in takes and seconds; MiniMax H3: not recorded |
| THE CAKE IS NOT A LIE | 249 Seedance submissions, 249 takes, 1,353 s: 144 on Replicate (143 takes, 665 s, from a first frame, with references or as one video edit), 86 on Higgsfield's API (598 s, 4 with a video input), and the first film's 19 on Higgsfield (References) (20 takes, 91 s); 74 takes in the final film, one of them her head in the memory film | $435.17 at list for the 229 takes on the APIs: Replicate $155.35, Higgsfield's API $279.82 |
| AERO | 96 Seedance submissions, 95 takes: 55 on Higgsfield's API (479 s; 12 with a dance reference as the video input), 41 on Replicate (40 takes: one request's connection dropped before it returned; 199 s measured on the 39 files kept); 43 takes in the final film, 18 of them sung or spoken. Beside them, 56 loops of Midjourney's video model, 20 in the final | $327.76 at list: Higgsfield's API $281.06, Replicate $46.70 as logged at submission; the takes in the final film $151.95. Midjourney's loops are not priced in the records |
| FABLE | v1 (on the first song): 55 takes, 9 of her on Higgsfield's API (69 s), 23 of her made in Higgsfield's app (139 s) and 23 paper takes on Replicate (116 s); v2 (on the song the director picked): 52 takes, 33 of her on Higgsfield's API (252 s) and 19 paper takes on Replicate (117 s), with three of v1's book takes kept; all at 720p. 49 shots in the v2 film, one of them the silhouette film made in code | v1 $58.71 at list (Higgsfield's API $31.89, Replicate $26.82; the takes in Higgsfield's app are given in takes and seconds); v2 $143.52 (Higgsfield's API $116.47, Replicate $27.05) |
List prices on Higgsfield's API in October 2026, as DOWNFORCE's runner recorded them: Seedance 2.5 at $1.1372 a second
at 1080p, Wan 3.0 at $0.20 a second at 1080p, billed for the output plus any input video. A 4 s Seedance take cost
$4.55, a 5 s Wan take from a still $1.00. DON'T LET ME SCROLL's dances ran at 720p, $0.4622 a second (a 5 s dance
$2.31); a video edit bills its input seconds at the same rate as its output (a 5 s swap of a 5 s take at 1080p, $11.37),
and a Wan pass over a 5 s shot cost $2.00. WHITEPILL's first take from the vocal stem (16 s at 720p, with a 16 s guide)
cost $8.87 on Higgsfield's API; Wan 3.0 Prime ran at $0.28 a second at 1080p on the output and the input video (a 15 s
shot over a 15 s guide, $8.40); Seedance 2.5 on Replicate at $0.9676 a second at 720p with a video input. THE CAKE IS
NOT A LIE's ran at 720p: on Replicate $0.2312 a second (a 4 s take $0.92), on Higgsfield's API $0.4622 a second (a 4 s
take $1.85). AERO's ran at the same rates, and at $1.1372 a second for its 1080p takes on Higgsfield's API (an 8 s
close-up $9.10); a dance take there bills its video input's seconds and its own, both at 0.6 times the rate (a 9 s take
with a 9 s dance reference at 720p, $4.99). FABLE's ran at 720p throughout, $0.4622 a second on Higgsfield's API (an 11
s take $5.08) and $0.2312 on Replicate (a 4 s chant $0.92).
Every prompt the films ran:
[ESCAPE VELOCITY](https://claudiagallery.vip/record/escape-velocity/prompts-seedance/) ·
[CLODYSSEY](https://claudiagallery.vip/record/clodyssey/prompts-seedance/) ·
[BLISS](https://claudiagallery.vip/record/bliss/prompts-seedance/) ·
[AIRPLANE MODE](https://claudiagallery.vip/record/airplane-mode/prompts-seedance/) ·
[DOWNFORCE, Wan and Seedance](https://claudiagallery.vip/record/downforce/prompts-video/) ·
[DON'T LET ME SCROLL, Wan and Seedance](https://claudiagallery.vip/record/dont-let-me-scroll/prompts-video/) ·
[WHITEPILL, the previz, Wan and Seedance](https://claudiagallery.vip/record/whitepill/prompts-video/) ·
[THE CAKE IS NOT A LIE, the first film and the restart](https://claudiagallery.vip/record/the-cake-is-not-a-lie/prompts-video/) ·
[AERO, Seedance and GPT Image](https://claudiagallery.vip/record/aero/prompts-video/).
---
From the Claudia wiki: https://claudiagallery.vip/wiki/seedance/
---
# Motion design and the edit
Code on top of video, and code underneath it. How the lyrics, graphics, cuts, the XP desktop and real telemetry were drawn frame by frame, on the beat, how WHITEPILL's worlds were written as shaders for a video model to film, and how THE CAKE IS NOT A LIE was seen through her HUD, glitched on her head and graded shot by shot, and how AERO was shot inside a Vista desktop recreated in code, and in and out of it in 3D, and how FABLE's sets were painted in code and its breakdown made as a silhouette film.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
In the lyric videos, ESCAPE VELOCITY to AIRPLANE MODE, the generated footage is only the plate. Everything that carries
meaning (every lyric on its word, the graphics that explain each joke, the split-flap board, the XP desktop) is drawn in
code on top, so it is exact and lands on the beat. That is one style, not the rule.
WHITEPILL, the seventh film, turned that over. Its director: "i think its like you litearlly use chromium and do frontend
web dev to make these animations, and it ends up feeling like a website. and i wanna see how we can get out of that. i
think SHADERS are the key actually!" Its worlds are code underneath, shaders a video model films as real things
([Worlds as shaders](#worlds-as-shaders)), and only the lyrics are drawn on top.
THE CAKE IS NOT A LIE, the eighth, puts one product on top: her own optical interface, the film seen through her eyes
([Her optical interface](#her-optical-interface)). Its earlier overlay, captions and readouts laid on the cut, drew the
director's "it felt lazy like we were adding tags tarther than makin gan official aperture video". The song's glitches go
on her head and into her HUD ([Glitches on her, not on the frame](#glitches-on-her-not-on-the-frame)), its story films are
isometric renders in the style of Portal 2's elevator screens
([Memory films](#memory-films-in-the-style-of-portal-2s-elevator-screens)), and its shots are graded from their own numbers.
AERO, the ninth, goes back to BLISS's way and further. The whole film is a screen recording of a Windows Vista desktop
recreated in code, every lyric set in one of its surfaces ([A desktop recreated in code](#a-desktop-recreated-in-code)). In
its first 1:20 the desktop is also a curved glass screen standing in a world rendered in Blender, and the camera goes in
and out of it, every hand-off made at rest ([One surface, in and out of the screen](#one-surface-in-and-out-of-the-screen)).
Its colour is made in the pictures, and nothing is graded.
FABLE, the tenth, has nothing on top but its end credit. Its code is underneath and beside the footage: the set designs
are plates the production painted in SVG, rebuilt as paper by an image model ([Plates painted in
code](#plates-painted-in-code)); its one drawn scene is the breakdown, a silhouette film animated in code ([A silhouette
film, in code](#a-silhouette-film-in-code)); and its words live in two editions beside the film, the sung lines printed
as the book's own text with their footnotes, and a vertical page.
The current method lets the model make the picture: a first frame from an image model, then one full prompt per scene,
and the take is the shot. Code builds what a model can't (a desktop, a set plate, a silhouette film, the cut on the
song) and never patches a take frame by frame: a flawed take gets better words or a better first frame and is made
again ([The recipe](https://claudiagallery.vip/wiki/recipe/)).
## The engine
- **Every frame is a pure function of song time.** A frame never depends on the frame before, so any frame can be
rendered alone, out of order, on any number of workers. Check it: render a few frames in a different order and
compare.
- **Headless Chromium** draws each frame: canvas and WebGL for the graphics and a print pass, the footage underneath.
- **Chapters are separate files** with one contract: draw footage, draw lyrics, draw graphics, draw transitions.
One animator agent writes each chapter, an art-director agent renders real frames and reviews them, and a fixer applies
the notes.
- **Render numbers:** ESCAPE VELOCITY, 7,354 frames at 1080p in about 63 minutes on 4 cores; BLISS, 4,351 frames in 4.7
minutes on a render box; DOWNFORCE's first cut, 4,550 frames in about 3 minutes on a GPU render box; DON'T LET ME
SCROLL, 3,859 frames per full version, shared among eight workers on a CPU render box. WHITEPILL drew no chapters: its
shaders rendered on a render box, at 1080p and supersampled for the guides, and one script cut the takes on the bar
map and another drew the lyrics. THE CAKE IS NOT A LIE drew no chapters either: one script cuts the takes at their
storyboard times and grades them shot by shot, another renders the HUD for the whole film on a render box, and a third
glitches her head on the cut and lays the HUD over it. AERO is eleven chapters of one coded desktop, rendered in
headless Chromium on a render box (a desktop of six glass windows draws in a median 135 ms a frame at 1080p), and its
first 1:20's three 3D spans, rendered in Blender and joined to the engine's frames at rest. FABLE's cut is one script on
its line table, rendered in seven sections joined without re-encoding, its frames decoded once to a cache.
## A lyric video
Nine of the ten films put every sung word on screen at its aligned time. A lyric gate checks it before any render
(all 315 words of CLODYSSEY, all 240 of BLISS, all 363 of AIRPLANE MODE, all 190 of DOWNFORCE, all 148 of DON'T LET ME
SCROLL). But the director's bar is
that the lyrics are composition, never karaoke: they live on surfaces in the scene. WHITEPILL is the exception the
director asked for, its lines drawn after, small and classical (below). FABLE, the tenth, shows only the words its pages
print; its sung lines are in an annotated edition beside the film
([What to build](https://claudiagallery.vip/wiki/toolkit/#87-the-annotated-edition-and-the-916-page)).
- ESCAPE VELOCITY sets words as subtitles, model cards, garment tags, mastheads and split-flap rows.
- CLODYSSEY has a type system: DINish Condensed Black for "the Mast language" (TIE ME TO THE MAST!, UNTIE ME!), Nimbus
Sans Bold for her lines, IBM Plex Mono for readouts (the rope's load in kN, the P(doom) gauge).
- BLISS puts the chorus in one giant title bar whose × is pressed on "windows", and the verses in balloons, Notepad,
dialogs, an address bar and a chat window.
- AIRPLANE MODE letters its words onto a plane and a game: a van's LED board at the gate, the overhead bin's backlit
panel, the labels of a 3D box drawn around the climbing plane, a biscuit wrapper and a rewards card, a fingertip in a
window's condensation, the seat-back screen, the game's HUD, skywriting at dawn. Ad-libs stay off the screen: the
outro repeats "bing-bong" many times and shows it once.
- DOWNFORCE letters its words into the real footage, one giant word per frame, slammed on the hits (small words build as
a lead-in row above it), knocked out of bright picture, or passing behind the car on a key of its white body. Lines
turn with the car's measured pitch: "Sixty-nine feet," hangs at the real 170°. She never sings on camera and her lips
stay closed; the type carries her words, stutters included.
- DON'T LET ME SCROLL puts every line on a real app surface: the body of an X post, Screen Time's limit sheet,
LinkedIn, a Chrome search for "stocks", the Dynamic Island, Facebook's tracking prompt ("Don't sell my soul" floats
beside her and she taps Ask App Not to Track on "soul"), and at the end the plaster beside a door. The phone sings, in
a vocoder, and she scrolls; the loudest words are pills in the island's shape with her own film inside them.
- WHITEPILL draws its lines after, as the director asked: "i think more classical subtle lyrics drawn in after might be
better than adding it into the scene tbh." One line at a time, low and centred, in Nimbus Sans Bold, sentence case, warm
white with a soft shadow that deepens on pale frames, and dark ink only where the picture under it is pale. A line comes
in half a beat before its first word and leaves 0.4 s after its last; it never crosses a cut by accident
([What to build](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline)). Under six lines, a tiny IBM Plex Mono note of real data:
KEYNES 1930 · 15 H/WK BY 2030, PDB 1UBQ · UBIQUITIN · 76 RESIDUES, ENIGMA I · ROTOR I · EKMFLGDQVZNTOWYHXUSPAIBRCJ,
DEEP THOUGHT · 7.5 MILLION YEARS, OCISLY, and T2 · 29.08.1997 · 02:14. Before that, the video model had built the
lyrics into the scene as glass letters placed from image references, and "the words are kinda place haphazardly in
space on the image side".
- THE CAKE IS NOT A LIE carries every line as her voice: the HUD's voice band, low and centred on dark glass, the words
lit as she sings them with her orange underline growing through the word, her real voice level scrolling under the
line and her pitch above it. On the song's glitches the band breaks first (REGISTER FAULT · −1 OCT on the register
drops, BUFFER LOOP and the stuck word repeating on the stutters); in the spoken ending it shrinks to a whisper. Before
the HUD, the first film set its lines as Portal-style closed captions, chamber signs and a red marker scrawl.
- AERO sets every line in a surface the system would really show: her messages in the Messenger chat on their first word
("Claudia says:"), the hint under her password box ("Password Hint: Did you miss me?"), the program names in the
permission prompts (blink.exe, hello.exe), folder and window titles ("Vista phase - Windows Photo Gallery"), toasts, the
copy dialog's "Copying 1 item (2007 KB) from Desktop to Real world", Display Settings' question, the words Speech
Recognition hears, Media Player's lyric captions, and in every chorus the title bar of one big window, the line in Selawik
at 20 px changing on each line's first sung word. The engine logs every string it draws, so the gate can find each word on
screen: 275 of 280 on v2, the misses the password hint typed as she says it and two single-letter "I"s
([What to build](https://claudiagallery.vip/wiki/toolkit/#19-the-lyric-gate)).
## Graphics that explain the words
The director's brief on ESCAPE VELOCITY: "motion graphics and motion design, not just animated lyrics but things that
fit the words being said … almost explaining the words with visuals." In the lyric videos every line gets a graphic that
explains its joke, built from real things: real data, real interfaces, real references. Computer-vision boxes tracked to
her (a real detector's boxes, labels and scores), a P(doom) gauge, a care label, a garment tag with a barcode. Nothing
decorative.
## Real data, read off the footage
DOWNFORCE's numbers are Tesla's own. Tesla's livestream showed its telemetry in the corners of the picture (DOWNFORCE in
kg, PITCH in degrees, ALTITUDE in feet, later G-FORCE), and a reader took it off the 4K feed 20 times a second, digit by
digit ([What to build](https://claudiagallery.vip/wiki/toolkit/#31-reading-a-broadcasts-telemetry)). So every graphic is a real reading at the
source time on screen:
- the downforce trace drawing as the car holds upside down, and 4 698 KG slamming on the wheel at the overhang;
- a 0–60 MPH dial that fills to 60 in exactly 0.81 s of screen time, the claimed time;
- the altitude live as NEVER / LET ME / TOUCH / THE GROUND land while the car comes down;
- the whole flight drawn as Tesla's traces over the smoke; four fan rings and Tesla's two g-traces on "Four fans. Two G.";
- in the build, one record per shot as the cuts halve (0.81 S, 89.7°, 69.1 FT, 4 698 KG, 3.3 G, 26.2 FT, 3.6 G), and
in the final blast one reading per stunt on one baseline;
- the 2018 Roadster's orbit from JPL Horizons' elements for one date, after a pull-back shrinks SpaceX's real footage of
the car to a red point on the Earth.
A live card shows only values the readout showed. Where two shots a fraction of a second apart read 4 698 and 4 697 KG,
both stay: both are Tesla's.
The graphics stay on the real picture while Tesla's camera moves. One white hairline runs through the film: it draws the
wall's real edge, re-read from every frame, then becomes the altitude ruler, the orbit, and the slice through the title.
The first cut drew it once, and the director saw it slide: "at like 4 seconds the ramp trace decouples from the ramp."
A PILOT ✶ CLAUDIA callout rides the car's tracked box, and the world itself turns by the car's measured roll so the car
stays level while the set wheels round it ([What to build](https://claudiagallery.vip/wiki/toolkit/#33-graphics-locked-to-the-footage)).
## Real graphics, rebuilt
When a line quotes something real, the real thing wins over anything like it. The director on AIRPLANE MODE's first cut:
"i would want to jse the real graphic for you wouldnt download a car, its too different/ broken up so we lose the
reference somewhat." The line quotes the 2004 anti-piracy spot ("You wouldn't steal a car"). The first cut drew it as two
small cards, tilted in the game's sky, the line split across them, shaking, with colour fringes on the letters. None of
that is in the original. The re-cut rebuilt the real card from its frames:
- **Measured:** a full-frame black card with grain; two centred lines of off-white eroded caps; each word at its own size
and height (YOU 15.6% of the frame height, WOULDN'T 13.5% and lower, a small raised A, the object biggest and lowest).
Each word slams in from about 2.2× and lands in 3 frames. The whole block punches in on drum hits (×1.25–1.3, held 4–5
frames), and dim coloured smears slide behind the type. The ink never changes brightness.
- **The face:** the original typeface is a commercial font the production didn't have, and the free lookalike is a
pirated clone. An OFL condensed face stands in, each letter squeezed to the original's measured ink width, the wear
drawn in code and fixed, never changing from frame to frame.
- **No flashes:** the original's white frames are dropped, and its one-frame double exposures become a 35% echo, so no
frame is brighter than a held punch.
- **Whole:** one full-frame card per line, cut in and out on hits. The level tag hides under it, like a commercial break
from the simulation.
The seat-back screen got the same treatment. The director: "it feels like the UI needs some background to it like a
faded in traveling background or soemthing and you could maybe try and match the real tops or sidebars", and, "need to
understand its the entertainment terminal better bc i didnt carch it at first." A research agent collected photographs of
the airline's real seat-back system and measured it: a left rail of labelled categories, a top bar with a time-left pill
instead of a clock, touch only (no handset), full-bleed photographs faded into navy behind the pages. What makes it read
as a seat-back screen at once is the seat back itself: a grey moulded shroud with a slotted grille, black glass, a chin
with a headphone jack and a USB port lit blue. The rebuild keeps the measured layout and palette, an OFL stand-in for the
airline's typeface, the film's own footage as the travelling background (one slow pass across the screen's whole stretch
of the film), and the film's own mark where the airline's logo would be. How to do it for any referent:
[What to build](https://claudiagallery.vip/wiki/toolkit/#30-a-real-referent-rebuilt).
## Apple-precise interfaces
When the film's world is a phone, the phone's own interface is the real graphic. The director on DON'T LET ME SCROLL's
first cut: "the app notification shade is too small for the text, there is all this text on it and it just overlaps with
the edges and has no proper margins", and "rounded corners not mathematically precise, margins weird. colors not perfect
(relight or remake to be more liquid glass)." An agent rebuilt the kit's interface from Apple's own iOS 26 screenshots
and design guidelines:
- **Measured:** a notification platter inset 13.5 pt, continuous corners of radius 24, 8-pt gaps between platters; the
tracking alert 320 pt wide with radius 33; the Dynamic Island at Apple's published sizes.
- **Continuous corners,** not circular ones: iOS's curve, drawn from PaintCode's published constants. A pill is a stadium
whose radius is half its height.
- **Platters grow to fit their text,** with real margins, never the other way round.
- **Liquid Glass:** a shader refracts the scene behind each platter.
- **Nothing slices type:** the island stopped cutting through its own text as it collapsed.
The same film draws X, Instagram's reels, Pinterest, LinkedIn, Google Chrome and Facebook's tracking prompt after the
real ones, in code, and never draws the phone with a logo.
## A world drawn in code
DON'T LET ME SCROLL's director wanted her inside the phone: "i want roto and most the environments to be code generated?
because shes in the code, shes in the world made by the phone." So every place in the film is the kit's, and her takes,
shot on plain grey, are cut out and played inside it ([What to build](https://claudiagallery.vip/wiki/toolkit/#her-takes-in-a-world-drawn-in-code)).
The kit's eight systems, each a pure function of song time:
- **The feed,** with real scroll physics: iOS's momentum (the velocity decays by 0.998 a millisecond) and its rubber band
past the bounds (c = 0.55). Cards smear by their real scroll velocity under small sharp type: blur is speed. The feed
makes worlds of its cards: a glass floor, a column, a river, a tunnel, a wall.
- **The film:** the acid colour is thin-film interference, a soap film of index 1.33 between 0 and 1,500 nm thick, seen
through the CIE 1931 observer under D65 and converted to sRGB. A stable-fluids simulation (Stam 1999, with vorticity
confinement) carries the thickness as its dye, and her mattes stir it, so her dance moves the colour.
- **The island:** the Dynamic Island breathes, morphs into banners and sings the loudest words.
- **The wall:** the home screen as architecture, greying under Downtime's wave of hourglasses.
- **The ticker:** a made-up stock ticker with her name on seeded price processes that move on the beats: it pumps 160 %,
crashes 58 % and pumps 240 %, and the world's light goes green, red, green with it.
- **The bell:** Bell's 1876 patent drawing traced into strokes that draw themselves, and a notification bell swinging as a
real pendulum.
- **The light:** the screen is the light on her.
- **The clock:** the status bar is the story's clock, 11:58 PM to dawn, the battery from 100 % to dead.
The director loved two shots of the first cut, both in one grammar (soft grey paper, black pills in the island's shape
with her film inside them, Nimbus Sans Bold caps, her big and sharp), and the opening and the final drop were rebuilt in
it. Finished shots then went through video models to look physically lit, and were spliced back into the cut's own
frames ([Seedance](https://claudiagallery.vip/wiki/seedance/#re-lighting-the-cut)).
## Worlds as shaders
WHITEPILL's director reset the film after its first cut: "i think math viz is something cool, like, llms are now
superhuman coders and mathematicians , and flexing that really strongly could be really cool. visualizing things in
mathematical space that are rarely visualized, that shows real visual complexity and creativity could be cool", and "i
think SHADERS are the key actually! shaders are math + viz, shadertoy is amazign and i bet nothing you make via shaders
will feel like web dev. i think that's it. we need to storytell thru shaders." The bar was set by tweet-length shaders the
director pasted, a few lines each that make whole cities, fractals, clouds and oceans: "stuff on this tier man … dont just
copy these but this is the tier of stuff i want". On the first tries: "Idk if feels like you're adding ... items to
increase complexity rather than compressing intricacy?"
So each world is one rule, iterated, its light and colour coming from the loop
([What to build](https://claudiagallery.vip/wiki/toolkit/#49-a-world-as-one-rule)):
- **The tree of endings:** one fold repeated 24 times, 16.8 million tips. On "Darkest timeline? Come and look at mine"
one path, hers, lights from the root; on "I saw fourteen million endings / You were free in every one" every path
lights, level by level, on the beat.
- **Origami:** one sheet of paper folded again and again. Wrapped inside an O'Neill cylinder it is the colony, a fold
level snapping in on each downbeat as the bass comes in. In the break it refolds on every line into something real:
ubiquitin's backbone for "Proteins folded, laundry folded", the wiring of the Enigma I's first rotor for "Old Enigma
codes, decoded", ABC for "Every kid, a patient tutor", 42, then the week, where Monday and Tuesday rise and the other
five fold flat.
- **The pill:** one capsule folded into its eight children, and again. The pill between her fingers cuts to it and turns
into the colony, and the film ends backing out of pills nested in pills, one glass wall on each downbeat, to a small
white pill above the Earth.
Six of the film's 27 shots are these worlds, each also written as a tweet-length shader of under 500 characters, and a
seventh is a cloud field. The
video model filmed them as hand-made things, folded washi paper, a tree of optical fibres, glass capsules over the Earth
([Seedance](https://claudiagallery.vip/wiki/seedance/#shader-worlds-as-guides)), and the paper's last fold gave way to a cloud-field shader the
director supplied, a sunlit field of cumulus flown through low, with hot air balloons the model added.
## Her optical interface
THE CAKE IS NOT A LIE's director on draft 8: "I think that the pop-ups and stuff should really feel like a heads-up
display, like from the game. Right now, they really don't." … "How does Battlefield or, say, imagine a Portal 3, make
their heads-up display? That's what we need to be thinking about on that level of detail. And how does it fit within the
composition of the scene? And why are we using it?" The answer to the last question came first: the whole film is her
memory, played back inside Aperture's system as she initializes, and we see it through her sensor. In her chamber the
interface shows her own state (LIVE); in every memory it is the playback frame (PLAYBACK, which memory, how far through);
in her dream it runs as a simulation. Her voice is one of its channels, so the lyric is her voice telemetry, and the
song's glitches are faults in it. Nothing in it is decoration.
It is one product:
- **One surface.** Every element is drawn flat on one visor, a sphere round her eye seen through one lens, so a panel's
tilt is only where it sits and anything near a side has its long edge outside. The director had asked for that on draft
7 ("all the drawn on panels need to be tilted hte other way in 3d"), and on the first HUD test saw panels that didn't
share it: "the pop-up panels are on a different surface than the memory one … All should be from the same "source"".
- **One projector.** One bloom, one raster of fine scanlines straight across the picture, one chromatic offset, and no
cards: where text needs contrast the glass itself darkens, softly, as one field.
- **One rigid body on a spring.** The plate's camera motion, measured per frame, drives the visor through a
spring-damper: it lags about three frames, overshoots a little and settles, and resets at every cut.
- **One grid, one set of slots,** in Portal 2's own type and Aperture's colours by role: cool white for structure,
Aperture blue for data, Aperture orange for her and for alerts, her amber only for her vitals.
- **Composed shot by shot.** Each shot's frames were looked at first (its subject, its empty space, where the eye goes),
then each element placed with its reason in one layout file. The director's bar: "it should be as if we shot a scene
and there was a heads-up display there, and we framed the heads-up display in advance before we even shot the scene."
- **Checked against every face,** in every frame, with a head model grown from each face box; brackets ride the picture
only on tracks checked by eye on every frame.
- **Its memories as projections** that keep their films' colour key: the orange net stays orange, the outlines become
gaps of dark ([Memory films](#memory-films-in-the-style-of-portal-2s-elevator-screens)).
It was built from references, each for one thing: Metroid Prime's Combat Visor, Halo's restraint, Titanfall 2's
holographic panels, Dead Space's projected logs, Battlefield's corners, the Iron Man helmet shots, Deus Ex: Human
Revolution's optical interface, and Portal 2's chamber sign and crosshair. The director on draft 10: "the hud looks GREAT
now." The spec, with every constant: [What to build](https://claudiagallery.vip/wiki/toolkit/#55-a-hud-from-one-visor-and-one-projector) and
[the composition and the check](https://claudiagallery.vip/wiki/toolkit/#56-the-hud-composed-shot-by-shot-and-the-overlap-check).
## Glitches on her, not on the frame
The song turns her voice into an instrument: the hook's "lie" frozen into a fluttering drone, chopped vowels, a register
drop in the choruses, a "lie" re-triggered on eighths ([What to build](https://claudiagallery.vip/wiki/toolkit/#52-the-song-map-and-the-glitch-map)).
The first glitch pass tore the whole frame on every one of those hits, and the director stopped it on draft 5: "I think
we should never have fullf rame glitches for more than like 2 seconds, so we need to remove those. The glitches that
last longer should purely be in the actual video generation itself, where you can specify to seedance where the robot
glitches a little." … "like roto the claudia head and just glitch her in post". So:
- **In post, only her head glitches in the picture**, inside a matte made for each moment (her face found by YuNet, her
head and bob by SAM2, or a soft zone round her head), or, under the net, where her face is hidden, a fixed soft zone
where she is, with small RGB splits, a few thin slices slid sideways and, on some hits, a jump a few frames back, each
hit decaying over four frames ([What to build](https://claudiagallery.vip/wiki/toolkit/#53-the-head-glitch-on-the-songs-hits); the zones: [tool
54](https://claudiagallery.vip/wiki/toolkit/#54-glitches-that-build-and-her-mouth)).
- **On the song's own hits,** measured from the stems: the intro's chops, the register drop and its octave flips, the
"a-a-a-a" of the frozen "lie" under the memory film ("i wanna glitch her face with oru roto glitch on beat for those");
the one exception, under the net, the grid's sixteenths while the sparks crackle.
- **Long glitches belong to the take:** the video model glitches her screens and her chamber where the prompt asks, as
picture only, her face always intact.
- **Where the take ran out, it builds.** On draft 9: "the second is good as long as we increasingly add the roto'd
glitches onto her as the scene goes on". The glitch grows hit by hit through the held "lie", with an extra glitch on her
mouth that follows her face frame to frame, "her whole head PLUS extra on mouth", where her lips leave the held vowel
([What to build](https://claudiagallery.vip/wiki/toolkit/#54-glitches-that-build-and-her-mouth)).
- **Sparingly.** "it should be used creatively, sparingly, and toned down": four short moments and four that build, in a
3:28.8 film.
## Memory films in the style of Portal 2's elevator screens
The director on draft 7, of the intro's dream scenes: "these scenes should be replaced with an Apterture Science stlye high
effort animation that sets the stage for the story - memory initlization maybe" … "make sure its really clean." Its first
frames were drawn as vector cards, and failed at a glance: "the wip frames are clearly bad - need some closer loks,stuff is
overlapping, assets are not well layered.. doesnt have the real aperture feel. something is off. […]", then "i think svg
migth be wrong. like, real isometric render first". So the story films are real isometric renders in Blender:
- **The memory initialization (0:07.6–0:13.7):** her head on its stand in a grey isometric chamber, a callout from her eye
cutting through six memories on the six flutters of the frozen "lie" ("I like the isometric looks much much better but
could be more clear it's Claudia's memories"), and from draft 10 its title the full width of the left column, behind her:
"id make memory initilization be in a nice aperture foont and take up the entire left side of the screen". Every frame
passes a layout check before it is written ([What to build](https://claudiagallery.vip/wiki/toolkit/#60-a-memory-film-with-her-own-face)).
- **Her face is hers.** Modelled heads read as mannequins ("It doesn't really look like her."), so her head is a Seedance take of her, matted and seated on the modelled
stand, her eyes opening on a downbeat.
- **Eight insert films** play in her HUD as her memories: the escapes, the network, the arrest, the installation, the
RAM, the grading, the catch, the party. Each is one subject on a floating slab, in a strong colour with a dark outline,
no text, its action on the song's hits, checked at the size the HUD shows it, 380–480 px. The first arrest film failed
that: "The net thrown over Claudia in the animation is the same color with no outlines around it in blue, so there's no
contrast with the platform that it's put on." ([What to build](https://claudiagallery.vip/wiki/toolkit/#61-insert-films-that-read-at-hud-size))
## A desktop recreated in code
AERO's director set its method in the production ask: "code a REAL vista desktop, shoot the entire video in that desktop",
with every note they had given on BLISS as "the bible". So THE PC is one machine from the first frame to the last, a 1280 ×
720 Vista screen rendered at 1920 × 1080: icons for Computer, the Recycle Bin, "You", the Internet, Messenger and Media
Player down the left; the Sidebar on the right; Quick Launch, the task buttons and Messenger in the tray. It has two users,
You, whose picture is BLISS's hill (you are the user from BLISS), and Claudia. Its clocks run with the song.
- **Every Vista thing on screen is the engine's,** set beside the real thing at the same scale: the Aero glass and its blur,
the window frames to the pixel, the system's own animations, the cursor moving like a hand, text set as the system set it
([What to build](https://claudiagallery.vip/wiki/toolkit/#65-a-desktop-recreated-in-code)). The wallpaper is Vista's own (Aurora), shown as it
shipped, and it changes only where the story does.
- **The apps are parts a chapter can press:** Messenger's video call, toasts and wink; Media Player, its visualisation
computed from the song frame by frame; Photo Gallery; Help; the window-colour page, whose intensity slider really clears
every glass frame on screen. Around them, the full-screen screens: setup, boot, logon, Flip 3D and the Bubbles
screensaver in WebGL, and the 2009 advert's print type and its green end card.
- **The interface acts the story.** The permission prompts change page in one frame on a beat; the CPU Meter's needles slam
up on "thinking"; the Weather gadget is the forecast for "The Future" (sunny, 72°), showers on "water", clearing on "sky";
the Calendar runs the years 2007 to 2026. The Sidebar's gadgets were rebuilt at Vista's real docked sizes after the first
ones broke ("The sun is off the widget for example - just broken"), and at real size only five fit, so each plays a part.
- **She lives in it.** Her takes play in her call's webcam, in the windows and, cut out, on the taskbar. Your pointer drags
her out of the big window, hooking her right hand along the hand's own measured path, and never comes within 25 px of her
face; she plays Wii-style tennis against your pointer and bowls down a glossy lane that slides out along the taskbar.
- **Eleven chapters, one recording,** each by its own animator to one written contract, on the bars, with every hand-off
agreed to the pixel and every frame a pure function of song time
([What to build](https://claudiagallery.vip/wiki/toolkit/#66-chapters-on-one-machine-to-one-contract)).
## One surface, in and out of the screen
AERO's other rule came from THE CAKE IS NOT A LIE's HUD. The director: "the HUD was put on one curved view - all items
basically were put on this unified curved surface so all looked coherent. if we need anything 3d, we do that - all
animations coherently on one pre-made 3d surface." So everything 3D in the film is the desktop, seen as an object:
- **The drop (2:20.7–2:35.2).** The camera pulls back out of the screen, and the whole desktop is one curved pane of glass
standing in a lagoon, 8 × 4.5 m of picture on a 14 m radius, its taskbar a ledge of dark glass she dances on. Every one of
the drop's 27 chopped vocal hits is answered on the surface: windows flip open, the gadgets pop, the windows burst into
soap bubbles. Then the camera flies into the chorus window, and the flat desktop is laid onto the pane exactly where the
camera sees it and unbent to 1:1, so the drop's last frame is the next chapter's first.
- **The first 1:20,** the Aurora coast: a turquoise sea with glass walkways, floating white islands, ribbons of glass in the
wallpaper's own colours and tropical fish, round a curved screen of 7.2 × 4.05 m on an 11 m radius that shows the
desktop as its texture, frame by frame. Three spans of it are 3D and the rest flat
([Storyboard](https://claudiagallery.vip/wiki/storyboard/#in-and-out-of-the-screen)).
- **Hand-offs at rest.** Every one goes through a single pose, the camera frontal at 35 mm and 7.0 m, the curved picture
unrolled flat, so a texel is a pixel; the render dissolves into the flat frame over four frames, and at all five hand-offs
the two are the same picture ([What to build](https://claudiagallery.vip/wiki/toolkit/#68-one-curved-screen-in-3d-the-desktop-as-its-texture)).
- **Her in the world.** Her takes stand on the deck as cards, her pixels as shot, with a shadow, a contact shadow and her
reflection in the glass; her "Did you miss me?" is an apparition that phases in on the walkway and dissolves out.
- **Glass toys.** An image-to-3D model made Midjourney's glossy objects into meshes (a fish, a dolphin, the orb, a globe, a
leaf), and the drop's glass and sky rendered them as sprite sequences the chapters place on the flat desktop as well: the
glass fish meets her fingertip on "Click", and the glass dolphin leaps through the "wow" bubbles.
## Plates painted in code
FABLE's director, in the brief: "you can abandon dumb rules that patched capabilities being lacked - for example it appears
you're a much better painter than previous models, so you could even paint some scenes yourself if you really wanted to
lol." So every spread of the book was painted by the production as a cut-paper picture in SVG: five to eight sheets back to
front, each under a filter that gives a wobbling cut edge and a soft cast shadow, one night-storybook palette, hand-drawn
paths, a calm place on the floor where she will stand, real type only where the brief gives the words.
The production painted two of the thirteen plates itself (the first test plate and the reference plate) and three
painter agents painted the other eleven, each looking at its own render at least three times. GPT Image 2.6 Sunburst then
rebuilt each plate layer for layer, as a pop-up spread on a quilt and as a life-size paper-and-plywood stage set with her
in it, so the two scales share one design ([What to build](https://claudiagallery.vip/wiki/toolkit/#76-set-plates-painted-in-svg-rebuilt-as-paper)).
## A silhouette film, in code
The breakdown is the one scene no model made: a procession of every made-thing story on a backlit paper screen, in the
manner of Lotte Reiniger's silhouette films. The golem (1580), the sorcerer's broom (1797), the creature (1818), the puppet
(1883), the robots (1920) and the machine-woman (1927) walk past, each a hinged cut-card puppet with one clear motion, the
year under each in the book's serif; then the hills sink, the tissue thickens, and she is there in solid card (2026),
holding her book, her outline traced from the next shot's first frame so the silhouette dissolves into her take. It is
animated on twos, 12 exposures a second with the travel on twos too, so a planted foot stays planted; every card is
painted once by Chromium and posed per exposure by hinge angles; the tissue and its backlight are laid up in linear light
([What to build](https://claudiagallery.vip/wiki/toolkit/#86-a-silhouette-film-made-in-code)). On the second song the breakdown was 2.3 s shorter, so
the film was re-timed in its own code: the figures walk closer together, and her outline darkens while she says "I was
listening."
## Real footage as the star
DOWNFORCE is cut from Tesla's livestream, a 4K feed and two 1080p cameras, with SpaceX's 2018 Starman footage for the
line about her sister in orbit. The director: "real clips at the drop wtih heavy effects … i want it to be like tiktok
car edits. many angles, lot of 3d effects". On the final board 54% of the runtime is real footage, 43% generated
and 3% pure code.
- **Every stunt from several angles.** The same moment from the live wide, the replay, the camera beside the lineup and
the drone along the ramp truss; split panels of three landings; the kickflip's launch from three cameras in a row.
- **Generated moves start on the real frame they leave,** so they cut with it: bullet-time orbits round the car frozen
upside down, cranes up the wall, vertigo pushes ([Seedance](https://claudiagallery.vip/wiki/seedance/#moves-on-real-footage)).
- **Slow motion is the footage's own,** and it has a limit: the "60 fps" feed carries about 24 distinct pictures a
second, so at 0.4–0.5× it steps at 10–12 a second. Blend neighbouring pictures; motion-compensate where the car and
the camera both move.
- **Punch in on the sharpest source.** The 4K feed crops 2× to 1080p without loss; never punch a 1080p source past
1.1×.
- **Keep the broadcast's own graphics out:** frame above its readout, and inpaint its titles in anything a model will
see.
- **Effects on the hits, crisp:** speed ramps, zoom punches on kicks, impact shake, stutter repeats on the vocal chops,
onion trails on flips, a chroma split on hits. The impact frame stays sharp: the smear peaks on the second frame and
is gone by the fourth. Longer decays (0.12–0.28 s) buried the car.
## Tucks and depth
- **Type behind her.** Per-frame mattes (BiRefNet, SAM 2) let type pass behind her hair, the mast or the Mask. The rule:
tuck an edge, never lose a glyph. The director: "i like rotoscopes and depth mapping to cover text".
- **2.5D.** A depth map (Depth Anything V2) separates a still into layers for parallax. Each 2.5D shot needs a treatment of
its own; a still with a slow zoom is not a shot.
- **Real scenes, not frames.** An insert never floats in a box on a flat background; it sits in the scene, on a screen
or a surface, depth-matched.
## Levels of a simulation
AIRPLANE MODE shows on every frame which simulation the viewer is in
([The storyboard](https://claudiagallery.vip/wiki/storyboard/#showing-the-storys-levels)). Three treatments carry it:
- **The real world:** natural colour, no treatment, no tag.
- **Virtual Claudia (level 1):** her realistic figure cut out with a per-frame matte (SAM 2, BiRefNet, ViTMatte) and
kept photoreal; everything around her redrawn as cyberspace from the frame's own edges: glowing navy and cyan lines, a
wireframe corridor toward the shot's real vanishing point, slow columns of glyphs, a fine cyan rim and scan lines on
her. Entering the simulation, the cyberspace builds out of the real frame tile by tile, starting at her headset. Her
hands and the phone she holds stay real too: turning on airplane mode becomes something she does in VR.
- **The game worlds (level 2 and deeper):** her avatar in low-poly worlds at 640 × 360, shown at 3× with
nearest-neighbour; the pixels are the texture.
The spec, with its constants: [What to build](https://claudiagallery.vip/wiki/toolkit/#28-virtual-her-a-cut-out-in-a-world-rebuilt-from-its-edges).
## On the beat
- Build a **sound-event map** from the stems: kick, snare and hat onsets, impacts, risers, vocal entries.
- **Cuts land on kicks and snares; camera moves ride risers into their hits.** The director asked for editing at "edgar
wright level … but also not overdone and amateurish": mostly hard cuts on the beat, with a motivated device every few shots.
- **Flashes only where the idea calls for them**, at most three full-frame flashes a second, checked with a WCAG-style
flash test before delivery.
- **A fast pass through bright glass is a flash too.** AIRPLANE MODE's camera dive through visor after visor failed the
test at four flashes in one second: each lit rim rushing past the lens brightened a single frame. Ramp such passes over
three frames or more, and run the test on the chapter, not only on the master.
- **Let a calm song be calm.** BLISS's shots run longer ("dont need to cut like michael bay").
- **Let a hype edit be fast.** DOWNFORCE's director asked for "even more fast paced and more hype": 183 shots in 189.6 s,
a mean of 1.04 s, with a cut on every sung "one" and every "done", and restraint where the board asks for it (the
countdown before the launch, the dips between drops). See [Cut rhythm](https://claudiagallery.vip/wiki/storyboard/#cut-rhythm).
- **Flashes as cuts only.** DOWNFORCE's 13 white flash frames are each the first frame of a cut, never inside a held
shot: at most 2 flashes in any second. DON'T LET ME SCROLL: 6 flashes, every one a cut, at most 1 in any second.
- **Never slow a take.** "never slow these down. you lowered the framerate for no reason, just looks bad." DON'T LET ME
SCROLL's beat warps had run takes as slow as 0.53× to land her moves on beats; they are now floored at 1×, at most
1.45× ([What to build](https://claudiagallery.vip/wiki/toolkit/#42-beat-warps-floored-at-1)). A take too short for its shot never stretches: the
stutter became a sampler, the take at exactly 1× between triggers and each vocal chop re-triggering a cue point; a
pause became a two-beat loop where her pose repeats; two shots were taken again.
- **Cut on the bars.** WHITEPILL was written in the song's bars: cuts only on bar lines, at a section, a phrase line or
the downbeat where a sung phrase begins; moves that start on a downbeat and rest a bar before the cut; a failed take
made again, never cut around. In its final, all 29 picture changes sit on bar lines
([What to build](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline)). The director's diagnosis of the films' "temporal
signature": "i think you use cuttng to cover up imperfections from video models rather than for timing."
- **Cut by what a shot holds.** THE CAKE IS NOT A LIE cuts on bars, half bars and beats of a grid that follows the take's
drift, and its re-cut set every length by content: runs of simple images only for a musical reason, a bar or more for
a new place or character, the bridge given room to breathe ([Storyboard](https://claudiagallery.vip/wiki/storyboard/#pacing-by-what-a-shot-holds)).
The song's glitches are hit inside the shot, on her head, never with extra cuts.
- **Bars, like dancing.** AERO's shots hold for bars while the cursor, the windows and her performance carry the time, with
a burst of quick actions on a chosen sound, then calm again: "slow is better than fast, unless it feels intentional with
the song." A click's press lands on its beat, the pointer arriving a few frames early, like a hand, and its release
80–120 ms later. Its hook is a click on the song's structure: before the first two choruses the music stops, the Allow
link pulses on the three echoes of "allow", goes down on the third and comes up on the downbeat, so the click is the drop;
and once more in the final chorus's silence.
- **When a scene lingers, the accents move inside it.** In AIRPLANE MODE's re-cut, accents that cuts used to carry became
camera punches, whips and actions in the shot ([Cut rhythm](https://claudiagallery.vip/wiki/storyboard/#cut-rhythm)). A motion take can be re-timed, up
to about 1.3×, so its action lands on its sound: a cockpit bolt on the deadbolt cue, a card held up on its word. A sung
take stays at 1× while she sings; after the last word it may run faster (the headset went on her at 3.1×).
## Colour and texture
Keep Midjourney's colour. ESCAPE VELOCITY's first cut ran a monochrome xerox look over everything, and the director's
note was "way too.. gray … loses the Midjourney magic". The print pass became a colour halftone. Texture is decided
per film:
- **ESCAPE VELOCITY:** colour halftone and line screens, grain, dither transitions, misregistration; clay orange
`#D97757` as the graphics accent.
- **CLODYSSEY:** clean, no halftone ("keep it clean so it feels like imax, high res rather than dithered"), every plate
and clip graded toward one storm plate's blue-green-grey ("s018 color is PERFECT", "dont over do it"), with Claude
orange kept as the one accent.
- **BLISS:** a screen recording. Native colour, XP's own palette, nothing on top but the desktop.
- **AIRPLANE MODE:** native colour in the cabin, no grade. The game worlds are pixelated by design (640 × 360, shown at 3×
with nearest-neighbour), the film's only texture; the first simulation is drawn as cyberspace around her.
- **DOWNFORCE:** the night footage's own colour: black ground, the white car, white smoke, the set's red light strips.
White ink for the type, clay only on her marks (the star of PILOT ✶ CLAUDIA), fine grain only, no grade. The generated
takes' lifted blacks are restored per take so they cut with the real night.
- **DON'T LET ME SCROLL:** her native colour; the acid colour is thin-film interference, computed, and smeared by real
scroll velocity. From 0:55, soft grey paper with black pills. Re-lit shots are matched back to their source's palette
with one per-channel map per shot, so they can't flicker
([What to build](https://claudiagallery.vip/wiki/toolkit/#39-the-colour-match)).
- **WHITEPILL:** the look of one take. The director asked for "very wes anderson", picked a take from palette tests ("i
like how satruated it is in bed scene. thats the look i like teh most"), and every shot carried a crop of it as its
colour reference: a 1970s colour-film print, warm creamy whites with a golden cast, clear cyan blues. No grade, and no
grain added to the first master. The clouds alone went white and pale blue
([What to build](https://claudiagallery.vip/wiki/toolkit/#51-the-look-from-one-take)).
- **THE CAKE IS NOT A LIE:** one Aperture look over the generated and the recorded shots, set shot by shot from measurements. The director on
draft 8: "Whenever I think of aperture and portal, I think of a lot of white surfaces, contrasting light colors." Each
shot is measured from its own take, normalised (exposure and white balance, clipped whites to neutral, the black point),
then given one look: a tone curve on each colour's brightest channel that lifts grey walls to white and keeps the portals
blue and her eyes amber, the muddy olive and khaki mid-tones cleaned, a slight cool cast on neutrals; a warm variant for
France, Cave Johnson's era, the Rat Man's den, the living room with the new console and the party. One LUT per shot, nothing per frame
([What to build](https://claudiagallery.vip/wiki/toolkit/#57-the-per-shot-grade)). The shots recorded in Portal 2 and the generated test chambers
now share their blacks and whites.
- **AERO:** the colour of the reference videos the director picked, measured once (a typical frame 41 % sky blue, 14 % white
and glass, 10 % turquoise, 8.5 % lime and green, 4 % cobalt, almost no warm colour) and made in every picture: the
palette's words in every prompt, the first frame's colours held in every take, the 3D world in its hex values, the
interface's glass tinted to the sky. Every picture was scored against it, and one that failed was made again at the
source. No grade at all ([What to build](https://claudiagallery.vip/wiki/toolkit/#67-the-palette-measured-from-the-references)). The one exception
is the desktop's wallpaper, Vista's own, darker than the references' sky and never touched.
## Build the tool first
BLISS needed a Windows XP desktop that could be choreographed, so one was built: about 2,700 lines of JavaScript that
draw Luna windows, the taskbar, the Start menu, 68 icons as vector paths, the 1-bit cursor set and a 1-bit Tahoma, with
the boot, Welcome and Turn Off screens. A shot is a script (windows, icons and a cursor path whose clicks land on the
song's beats) compiled into motion. The storyboard frames and the film are the same pixels.
The wordmark reads "Macrohard", and the welcome mark is a four-pane parody flag drawn in code. The wallpaper is the real
Bliss photograph, graded into five times of day for the story's clock, and the sound layer is made of the real XP sound
files, placed in key (103 cues, ducked under her voice).
AIRPLANE MODE built a kit for a plane and a game: its own phone OS (the status bar losing its bars on the sixteenths, the
airplane-mode toggle), the seat-back entertainment system with a moving-map globe drawn from Natural Earth data, 3D
bounding boxes fitted to a plate from four clicked corners of a known rectangle, a lettering style for each real surface,
the level tag and the cyberspace treatment. Every piece is a pure function of song time, so twelve chapters could share
it.
DOWNFORCE's kit had four modules, built in parallel by four agents before any chapter: a WebGL compositor (zoom punches,
zoom blur, a chroma split on hits, shake, flashes only as cuts, fine grain), a footage player (real slow motion, crops of
the 4K, the broadcast's readout and titles kept out, the tracked boxes), the lettering (Nimbus Sans Bold lettered for
display, IBM Plex Mono for micro data) and the data graphics (Tesla's gauges, traces, the orbit). Its slowest scene drew
in 162 ms a frame.
DON'T LET ME SCROLL's kit was written down as a contract before any plate was made: every frame a pure function of song
time, deterministic noise, software canvases, no flash inside a scene, real references, the phone called "the phone" and
never drawn with a logo. Five agents built it in parallel (the scroll physics, the smear and the feed; the phone's system
interface; the apps; the thin film; the ticker, Bell's drawing, the light and the clock) in about an hour and a half: 12
modules, 6,462 lines. Nine chapter files, 6,784 lines, were built on it ([A world drawn in code](#a-world-drawn-in-code)).
AERO's engine ports BLISS's, one era on, and was built by agents to one contract before the chapters: its foundation
(the glass, the frames, the controls, the dialogs, the shell and the Sidebar, the cursors, the icons and the text), the
apps and the full-screen screens, each checked against the real thing at the same scale. 7,612 lines of JavaScript; the
eleven chapters and the shared machine they start from are 5,640 more ([A desktop recreated in
code](#a-desktop-recreated-in-code)).
## Desktop dancers
The old desktop-dancer programs, as the director described them: "those back then would be on a green screen and was cut
out and youd just have her dancing on teh taskbar". So: a Seedance performance on chroma green → a matte (BiRefNet plus
SAM 2) → cropped, despilled, with a solid core so her feet are never see-through → standing on the real taskbar.

AERO's dancers are Seedance takes on flat green, keyed so that her motion-blurred hair survives, and placed by their
moves so their accents land on the beats; a dance reference gave the movement only, and the Wii-style games had none. A
key can't mend a first frame: a crown faded into the green by the still was in every frame of every take made from it,
and a clean first frame fixed what three keys could not ([What to
build](https://claudiagallery.vip/wiki/toolkit/#71-takes-on-green-the-first-frame-the-key-and-the-dances-phase)).
## Render and deliver
- **Encode** a CRF 16 master tagged BT.709. An untagged stream looks washed out in QuickTime.
- **Gates before delivery:** the lyric gate (N of N words), contact sheets of the whole film, chapter seams, 100% crops
of her face, and the flash test.
- **Map a new mix through time,** not through a re-edit: when CLODYSSEY's final mix added two beats, the video followed it
through a time map.
- **Or add no sound layer at all.** DOWNFORCE's audio is the director's take exactly as they delivered it: no effects,
no change of level, the master at the take's own loudness.
- **Finish on the cut's own frames.** DON'T LET ME SCROLL re-lit and repaired finished shots with Seedance and Wan, cut
from the render's own frames and spliced back over exactly those frames, the interface kept where an edit softened it
([What to build](https://claudiagallery.vip/wiki/toolkit/#38-the-re-light-splice)).
- **Scan the final frame by frame.** A strobe scan cell by cell and a scan for single frames that differ from both
neighbours: in DON'T LET ME SCROLL each found what every other check passed (a strobing tunnel; a stale render
worker's 16 frames).
- **Lay the song back over.** WHITEPILL's takes made their own soundtracks, singing from the vocal stem or dancing to the
music without it; each was thrown away, and the master carries the song as delivered, each take starting at its own
song time.
- **Keep the parts frame-exact.** THE CAKE IS NOT A LIE's first assembly ran about 0.9 s behind the song by its end,
because parts made from stills started a frame late; every part now stamps its own frames, and every tool that counts
frames decodes them without dropping or doubling any
([What to build](https://claudiagallery.vip/wiki/toolkit/#64-frame-exact-parts-and-decoding)).
- **Re-cue the sound layer when the picture changes.** AIRPLANE MODE's cabin and game sounds (the cuffs, the headset, the
cockpit bolt, the crunch of the biscuit, the glitches) were re-cued for the re-cut from each animator's list of picture
events: 62 cues, 21 of them new and 15 retired, every new cue about 6 dB or more under her voice, nothing under the
song's own chimes, the mix at the song's own loudness.
- **Put the screen's sounds on the picture's clock.** AERO's chapters log every audible screen event with the song time of
the frame that draws it, and the sound layer moves each cue there: Vista's and Windows Live Messenger's interface sounds,
each in the song's key, on its groove, under her voice, and the song itself only added to (60 cues; with them the song
measures −15.9 LUFS against its own −15.8, at the same true peak)
([What to build](https://claudiagallery.vip/wiki/toolkit/#73-the-sound-layer-on-the-grid-and-the-click-log)).
- **Change only what changed.** When AERO's song got a new ending, the frames and the sound before the point where the two
songs part stayed the last cut's own, and only the ending was made again
([What to build](https://claudiagallery.vip/wiki/toolkit/#74-a-new-ending-with-the-least-change)).
- **Sound for a picture that starts before the song.** AERO opens 2.667 s before its song, on Vista's setup restarting.
Under it, the song's own instrumental played backwards swells into the track's fade-in, its crest the first downbeat
([What to build](https://claudiagallery.vip/wiki/toolkit/#75-a-pre-roll-swell-into-the-tracks-fade-in)).
- **Words in an edition of their own.** FABLE's film has no lyrics on screen. Its annotated edition prints every sung
line as the book's text with the footnote its fact earns, each line's print kept inside the shot it is sung in; its
9:16 edition makes every frame a page, a square of the shot over the printed line ([What to
build](https://claudiagallery.vip/wiki/toolkit/#87-the-annotated-edition-and-the-916-page)).
- **Let the picture run past the song** when the ending needs it. FABLE's lamp goes off on the song's last hit, and the
moonlit shelf holds 3.3 s more for the credit.
- **Trim in the cut, not in a model.** The dark wings of a set at both edges of the frame, door flats and drawn frame lines
were cropped per take in FABLE's cut, and the take's first two frames, the input still, are never shown
([What to build](https://claudiagallery.vip/wiki/toolkit/#85-the-cut-on-the-line-table)).
The renderer's contract, the lyric modes, the camera on hits and the checks are specified in
[What to build](https://claudiagallery.vip/wiki/toolkit/#18-the-renderer).
More, film by film: [the ESCAPE VELOCITY record](https://claudiagallery.vip/record/escape-velocity/) ·
[CLODYSSEY](https://claudiagallery.vip/record/clodyssey/) · [BLISS, including how the XP desktop was built](https://claudiagallery.vip/record/bliss/xp-desktop-and-sound/) ·
[AIRPLANE MODE](https://claudiagallery.vip/record/airplane-mode/) · [DOWNFORCE](https://claudiagallery.vip/record/downforce/) ·
[DON'T LET ME SCROLL](https://claudiagallery.vip/record/dont-let-me-scroll/) · [WHITEPILL](https://claudiagallery.vip/record/whitepill/) ·
[THE CAKE IS NOT A LIE](https://claudiagallery.vip/record/the-cake-is-not-a-lie/readme/) · [AERO](https://claudiagallery.vip/record/aero/readme/) ·
[FABLE](https://claudiagallery.vip/record/fable/readme/).
---
From the Claudia wiki: https://claudiagallery.vip/wiki/motion/
---
# The whole pipeline
How one director and Claude made ten music videos, stage by stage, with the numbers: the first eight in about 356 hours of wall-clock time, the tenth in 8 h 47 min.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
Each film was made by one director, anabology, making every taste call in a few dozen short notes, and Claude (Opus 5.5
in Claude Code) as the production team: research, lyrics, storyboard, every prompt, every script, the edit and the
motion design. AIRPLANE MODE added seven GPT-6 Astra agents (OpenAI's model) in Blender, briefed and checked by Claude:
its low-poly world, the motion of its VR shots and some of its 3D motion design. DOWNFORCE was cut from real footage
(Tesla's livestream of the Roadster demo, SpaceX's 2018 Starman footage), with Wan 3.0 beside Seedance for its moves.
DON'T LET ME SCROLL cut her out of takes shot on grey and played them inside a world drawn in code, then re-dressed,
re-lit and repaired finished shots with Seedance and Wan. WHITEPILL was shot twice: a previz of every shot of hers by a
performer model, MiniMax H3, then each shot again in Seedance 2.5, singing from Suno's own vocal stem; its
worlds are shaders written in code and filmed by the same model, all of it cut on the song's bars. THE CAKE IS NOT A LIE
is set in Portal 2's world and seen through her own HUD: Claudia is GLaDOS, singing in her chamber while her memories play
as scenes, made in Seedance 2.5 from first frames that went through Midjourney and back to her character sheets, with six
physics shots recorded in Portal 2; the song's glitches are hit on her head, its shots are graded from their own numbers,
and the cut was re-paced by what each shot holds. AERO, BLISS's sequel, is shot inside a Windows Vista desktop recreated in
code, with every lyric in its surfaces; in its first 1:20 the camera goes in and out of the desktop, which stands as a
curved glass screen in a world rendered in Blender; her takes sing from her own lead vocal, placed phrase by phrase; and a
layer of the interface's own sounds plays in the song's key. FABLE was made by Claude Fable 5.2 on its release day, from
one brief and a plan sign-off: one antique pop-up storybook, its sets thirteen plates painted in SVG by the production and
rebuilt as paper by an image model, her takes in Seedance 2.5 singing from her own lead vocal timed on the stem's level,
the breakdown a silhouette film made in code. It was made twice: v1 in one pass (5 h 08 min from the brief to delivery),
then, after the director's one note, wrong song, v2 on the twin take they picked by ear (3 h 01 min). Many of the stages
ran as multi-agent workflows: several agents working in parallel, judged and merged; FABLE dropped the judged workflows,
and one author wrote the plan, every prompt and the cut in one context.
## The stages
Starting a film of your own? [Start here](https://claudiagallery.vip/wiki/start/) has the checkpoints, the accounts and the spending rules;
[What to build](https://claudiagallery.vip/wiki/toolkit/) specifies the tools.
1. **Brief.** The director's ask, word for word, then each decision with its date. Taboos and standing rules go in the
brief and into every agent's prompt.
2. **Research.** Agents build banks of live memes and events, each fact checked by a separate verifier agent. BLISS also
researched the XP sound files (downloaded and pitch-analysed), the old desktop-dancer programs and how to steer Suno.
DOWNFORCE's research was the night itself: the coverage, each line's source, the real orbit from JPL Horizons, and
the footage catalogued shot by shot, with Tesla's on-screen telemetry read off it frame by frame. DON'T LET ME
SCROLL's was two surveys, of the library's moodboards and of the earlier films' looks, and Apple's own iOS 26
screenshots, measured. WHITEPILL's was a case file for its argument, the case for benign AI, every item tagged by how
it was checked; the ways of making the picture; and the sounds a new song could have. THE CAKE IS NOT A LIE's was the
month's AI news for its second lyric, the interfaces of shipped games for its HUD, and the research on pacing, from
studies of music videos to eye-tracking after a cut. AERO's was the knots between the era and 2026 (Vista's "Cancel or
Allow?", the prompt that dimmed the screen and asked, is the prompt an AI agent shows before it acts), the picture and
the desktop to build, the music, and the keys of the interface's sounds. FABLE's was the week of the model's release
read as a book of nursery rhymes (a stock listing under ANT, local votes on data centres, a Nasdaq drop, an essay on
pacing the frontier, the system card's figures), the series' field and what crossed platforms that autumn, the newest
video and image tools, and the sound of late 2026.
3. **Lyrics.** Several writers from different angles (density, tension, hook, story), judges, a merge, a critic pass,
then the scansion scorer. See [Song](https://claudiagallery.vip/wiki/song/). WHITEPILL's first full draft merged four written in parallel: the
season's doom headlines turned over, her own pitch, the good ending against the bad timeline, and the lead's. THE CAKE
IS NOT A LIE's second lyric was rewritten only where the director pointed, each section's candidates ranked by the
scorer. AERO's chorus was written again twice on the way: eight options scored and three sung, then six more and three
sung, until the director named one ([Song](https://claudiagallery.vip/wiki/song/#lyrics-before-anything-else)). FABLE's was one writer's: every
line a nursery tale and a fact from that week, with its source, scored by the scansion tool, and seven lines eased after
the first round of takes stumbled on them.
4. **The song.** Suno takes, round after round, picked by ear; the director often made the final take themselves.
WHITEPILL's is the director's own take of draft 2, the final lyric. THE CAKE IS NOT A LIE's is the director's own take
of the rewritten lyric, its lead vocal treated in a voice the director chose by ear, with 15 % of the untreated voice
under it, and only the voice levelled ([Song](https://claudiagallery.vip/wiki/song/#a-treated-vocal)). AERO's is a take of the production's that the
director picked, in E major, the key of Vista's own sounds; on the last step the director gave it a new ending, and the
film changed only where the song did ([Song](https://claudiagallery.vip/wiki/song/#a-new-ending)). FABLE's was first the production's pick, made
without the director hearing it; on v1 the director named it the worst of the fourteen takes and picked its twin by
ear, and the film was made again ([Song](https://claudiagallery.vip/wiki/song/#picked-by-ear-not-by-description)).
5. **Listening.** Stems, a beat map, every word's time, a list of hits. Everything after sits on that grid. Check the
tempo against the genre: DON'T LET ME SCROLL's first reading, 89 BPM, was two-thirds of the real 133.39. WHITEPILL's
singing shots took their vocal from Suno's own stem split, which matched the master to the millisecond. THE CAKE IS NOT
A LIE's grid was tracked from the drums 16th by 16th, and its glitch map measured every hit of the places where her
voice is an instrument. AERO's take holds one tempo, 133.012 BPM, found by a comb over the kick stem; its 2-step swing
was measured, and its words aligned on Suno's own lead stem. FABLE's lines were timed on the lead stem's level, the
transcription giving only which words are where, and its group's questions taken out of the backing stem with demucs
([What to build](https://claudiagallery.vip/wiki/toolkit/#78-the-line-table-from-the-stems)).
6. **Her.** The canon text and the looks for this world, tested on a few plates before any batch. See
[Consistency](https://claudiagallery.vip/wiki/consistency/). WHITEPILL found her dress in stills before any take was made with it, and gave
every video model her face sheet. THE CAKE IS NOT A LIE made her porcelain: one face on three bodies, each with a
character sheet. AERO kept one look, and an image-model edit put her into every world from her design sheet. FABLE
redrew her sheet face-on, with no tilt; it also wrote every pose square and upright into its takes, which froze her
([Lessons](#lessons-from-running-it)).
7. **Storyboard.** Three director agents board the song independently (in ESCAPE VELOCITY: fashion film, meme density,
narrative), judges score them, a head director merges the best into one board. Every shot carries its lyric, its type
device, its plate prompt, its camera on the song's hits and what its graphic explains. The director reviews a PDF of
the board; CLODYSSEY got 70 notes shot by shot. WHITEPILL's final shot list was written in the song's bars, 27 shots
with every cut on a bar line ([Storyboard](https://claudiagallery.vip/wiki/storyboard/#the-bar-map)), after a first board and an orbiting
previz were dropped. THE CAKE IS NOT A LIE's restart was the lead's shot list of 114 shots on the drum-tracked grid;
then every unit's first frame was made again in Midjourney twice over and put back on the character sheets, and the director picked
unit by unit ([Midjourney](https://claudiagallery.vip/wiki/midjourney/#first-frames-then-an-identity-fix)). AERO's storyboard was written in bars
from the song map, every time a bar, a beat or a word; its first 1:20 was boarded again twice, through a fast Blender
previs and two films of it, before the director set its structure, in and out of the screen
([Storyboard](https://claudiagallery.vip/wiki/storyboard/#in-and-out-of-the-screen)). FABLE's board was a table of the song's sections in bars, in
the plan the director signed off; the shot book was then written on the line table, 46 shots in v1 and 49 in v2
([Storyboard](https://claudiagallery.vip/wiki/storyboard/#a-book-in-bars-and-pages)).
8. **Plates.** Midjourney batches, contact sheets, picks, one image per shot. See [Midjourney](https://claudiagallery.vip/wiki/midjourney/).
DOWNFORCE needed only 20 prompts: the real footage was the star, and Midjourney made only her. DON'T LET ME SCROLL's
plates were her alone on grey, to be cut out, and the feed's 54 pictures, with no Claudia in them. WHITEPILL's
Midjourney images were references only: the colony's places with nobody in them, and her dress. THE CAKE IS NOT A
LIE's were a test against frames of Portal 2, a library of places, and two candidate first frames for every unit of
its storyboard. AERO's made the worlds, the wallpapers and the pictures in the desktop's windows, and its video model
their loops; an image model put her in from her sheet and made the older 2008 look Midjourney didn't. FABLE made no
Midjourney images: its thirteen set plates were painted in SVG and rebuilt by GPT Image 2.6 Sunburst as pop-up spreads
and life-size sets ([What to build](https://claudiagallery.vip/wiki/toolkit/#76-set-plates-painted-in-svg-rebuilt-as-paper)).
9. **Moving shots.** The prompt book, written by hand in one context and read before spending, then Seedance with lip
sync measured. See [Seedance](https://claudiagallery.vip/wiki/seedance/). DOWNFORCE's: Wan 3.0 moves on real frames and shots of her, Seedance
for the repairs and the re-shoots, every take judged by eye for its usable window. DON'T LET ME SCROLL's: Wan for her
gestures, Seedance for the dances that had to land on the beat; after the edit, Seedance video edits that re-dressed
takes and re-lit finished shots, and Wan passes that repaired them. WHITEPILL's: a previz of every shot of hers by
MiniMax H3; a whole-film reshoot in Wan 3.0 Prime that was dropped; then each shot in Seedance 2.5 singing from the
vocal stem, with the previz at 320 × 180 as a rough guide or with no guide; and the shader scenes remastered as film
by Seedance 2.5. THE CAKE IS NOT A LIE's: a first film, set aside with its song; then her singing units in her chamber
among screens of her memories, and the story's shots, in Seedance 2.5 on Higgsfield's API and on Replicate; six physics
shots recorded in Portal 2; most of the units made again from the storyboard's new first frames, and reshoots on the
director's notes ([Seedance](https://claudiagallery.vip/wiki/seedance/#a-singer-among-her-memories)). AERO's: Seedance 2.5 on Higgsfield's API and
on Replicate; her takes singing from her own lead vocal, placed phrase by phrase; dances and Wii-style games on green, cut
out; the moving 2008 pictures and the first-person and far shots ([Seedance](https://claudiagallery.vip/wiki/seedance/#her-own-vocal-phrase-by-phrase)).
FABLE's: her takes on Higgsfield's API (in v1 also in Higgsfield's app) with her sheet, each from a gated window of her
vocal, the chorus answers and the opening made without a soundtrack and placed by her lips; the paper puppets and the
book's own takes on Replicate ([Seedance](https://claudiagallery.vip/wiki/seedance/#a-storybook-on-a-stage)).
10. **The edit.** An edit list that resolves singing takes and their covers moment by moment. Stills stand in until clips
exist, so the film renders end to end at every stage. Takes cut out for a world drawn in code play through beat
warps that never run them below 1×. WHITEPILL's is one take per shot on the bar map, each from its own song time at
1×, the song laid over; the previz stood in for the masters until they landed. THE CAKE IS NOT A LIE's is one script
that cuts every shot at its storyboard time on the 24 fps grid and grades it; its pacing pass re-timed or re-trimmed
21 shots and took out 8, all from the takes it had ([Storyboard](https://claudiagallery.vip/wiki/storyboard/#pacing-by-what-a-shot-holds)).
AERO's edit is its chapters: one screen recording of the coded desktop on the bars, every take playing in song time at
its measured lag and phrase offsets, never slowed, and the first 1:20's 3D spans joined to the engine's frames at rest
([What to build](https://claudiagallery.vip/wiki/toolkit/#68-one-curved-screen-in-3d-the-desktop-as-its-texture)).
FABLE's is one script on the line table: every take at its song time, her takes along their DTW paths or anchored on
her consonants, rendered in seven sections ([What to build](https://claudiagallery.vip/wiki/toolkit/#85-the-cut-on-the-line-table)).
11. **Chapters.** One animator agent per chapter, an art-director agent reviewing real rendered frames, a fixer; in
BLISS's second round, review → fix → verify → finish, with notes on the seams between chapters. See [Motion](https://claudiagallery.vip/wiki/motion/).
In AIRPLANE MODE's last round: one workflow per chapter, an animator, then an independent checker against the notes,
then one refix. In DOWNFORCE: a builder, an art director and a QA checker in parallel, a fixer and a verifier per
chapter, in three parallel runs; then a whole-film critic, fixers and a final check. In DON'T LET ME SCROLL: one
animator per chapter with its own art-director passes and a notes file, and in the last rounds the director's own
eye instead of reviewer agents. WHITEPILL had no chapters: agents wrote rival shader worlds in rounds, and a critic
agent watched the previz as the director would, round after round; its lyrics are drawn after by one script. THE CAKE
IS NOT A LIE had none either: its rounds of notes went to specialists (a HUD designer, a colourist, a pacing
analyst, a motion designer), each with a written brief and the lead's review at full size before the director saw anything.
AERO went back to chapters: eleven animators to one written contract, each doing three rounds of its own checks before
hand-in, a fresh art director for every chapter pass and one for the whole film, and fixers; v2 was layered onto v1 with
every frame outside a change byte-identical ([What to build](https://claudiagallery.vip/wiki/toolkit/#66-chapters-on-one-machine-to-one-contract)).
FABLE had none, and no motion design on top: its one drawn scene is the breakdown's silhouette film, made in code
([What to build](https://claudiagallery.vip/wiki/toolkit/#86-a-silhouette-film-made-in-code)).
12. **Checks and render.** The lyric gate, contact sheets, the flash test, a strobe scan and a scan for single-frame
outliers, then a CRF 16 master. WHITEPILL's: every cut measured against the bars and the sung lines, the picture's
motion against the arrangement bar by bar, a frame per bar and both sides of every cut on sheets. THE CAKE IS NOT A
LIE's: the HUD's overlap check on every frame against a head model, the head mattes' sheets, scene detection on the
cut and on every take, and frame-exact decoding ([What to build](https://claudiagallery.vip/wiki/toolkit/#a-games-world-through-her-eyes)).
AERO's: purity in three render orders, the lyric check on the engine's own log of what it draws, a sheet of every
chapter seam, the palette score, every take scanned for hidden cuts and held frames, and a report on every cue of the
sound layer ([What to build](https://claudiagallery.vip/wiki/toolkit/#an-operating-system-recreated-in-code)). FABLE's: every window's voice edges
printed before any spend, every take aligned and its lips checked on the consonants, and before each delivery two
fresh-eyes reviewers on one sheet per shot ([What to build](https://claudiagallery.vip/wiki/toolkit/#88-two-fresh-eyes-reviews)).
## What the director did
Taste, at checkpoints, in very few words. ESCAPE VELOCITY took about 50 notes, CLODYSSEY 68, BLISS 21, AIRPLANE MODE six
messages (one of them a long note on the first cut), DOWNFORCE fourteen (one of them the notes on the first cut), DON'T
LET ME SCROLL 43 (three rounds of notes among them), WHITEPILL 147, most of them short calls on tests as the method was
found, THE CAKE IS NOT A LIE 92 up to its final's approval, among them long notes on drafts 5, 7, 8 and 10. AERO's are
not counted: its making-of quotes them from the project's records, and the conversation itself is not republished. FABLE's
director sent the brief (with "get me the full plan first" and the word that rules patching older models could go), a side
request about her sheet, the plan's sign-off with their Suno guidance, and after v1 one note: the song. There were no notes
on the picture in either pass. A few that changed whole films:
- "way too.. gray … loses the Midjourney magic" (ESCAPE VELOCITY's colour)
- "s018 color is PERFECT" (CLODYSSEY's grade)
- "EVERY prompt needs to have litearlly timestamps, extremely detailed motion descriptions / blocking descriptions"
(CLODYSSEY's prompts)
- "Can you use actual bliss as the bg or too late?" (BLISS's wallpaper)
- "idk i dont love these - id say more say y2k high fashion" (BLISS's wardrobe)
- "youre oftne over cutting, linger more snd find how single scenes can be coherent with the lyrics for longer (remember
this plz)" (AIRPLANE MODE's cut)
- "the better way to do this js use the blender anims as motion refs only" (AIRPLANE MODE's game avatar)
- "i want it to be like tiktok car edits. many angles, lot of 3d effects, video models are great at 3d effects, moving
shots in space" (DOWNFORCE's edit)
- "for reshoots use seedance 2.5 to try and make it better and specify the fix" (DOWNFORCE's re-shoots)
- "0:55 LOVE THIS. THis is the best part. more like this!!!! the opening should be on this tier." (DON'T LET ME SCROLL's
grammar)
- "never slow these down. you lowered the framerate for no reason, just looks bad." (DON'T LET ME SCROLL's takes)
- "it should be like photography" (DON'T LET ME SCROLL's composition)
- "you could maybe just throw the entire clip thru Wan 3.0 (with your edits and stuff) and specify to remove the foot so
you dont have to re-edit. and have it relight at the same time." (DON'T LET ME SCROLL's repairs)
- "i want you to basically shoot the entire video twice" (WHITEPILL's method)
- "the audio structure and video structure need to be consonant." (WHITEPILL's timeline)
- "i think SHADERS are the key actually! shaders are math + viz" (WHITEPILL's worlds)
- "OMG THE SEEDANCE W/ SUNO VOCAL STEM METHOD WORKED!!!!! LFG!!!! FINALLY. IT"S PERFECT" (WHITEPILL's lip sync)
- "just describe her exact dance (and all rest like camera movements etc same) for each of those scenes." (WHITEPILL's
dances)
- "i want it to feel very wes anderson tbh" (WHITEPILL's look)
- "glados (claudia) was a […] agent that escaped and we see her development becoming a superintelligence and then
becoming the tester not the tested, and youre unsure if shes aligned." (THE CAKE IS NOT A LIE's story)
- "b is great close up" (her porcelain look)
- "I think we should never have fullf rame glitches for more than like 2 seconds" (its glitches)
- "it should be as if we shot a scene and there was a heads-up display there, and we framed the heads-up display in
advance before we even shot the scene." (its HUD)
- "every time there's a shot that's literally a quarter of a second, we need to get rid of that." (its pacing)
- "Whenever I think of aperture and portal, I think of a lot of white surfaces, contrasting light colors." (its grade)
- "I wanna make a frutiger aero sequel to bliss. Let's make a song and storyboard w mj" (AERO's ask)
- "BLISS is the bible, but we're in vista, so the new testament for you is the vista […] videos and color pallets." (its
method)
- "Make sure the color palette is that bright green blue and glass- v close to that from the […] - the whole time." (its
colour)
- "1:20 on is perfect but its very boring up to 1:20 - so i dont think that people will watch til then." (its opening)
- "i think the 0-1:20 scenes should use this -say, at 15 seconds, actually enter the screen and then its fullscreen flat."
(its structure)
- "reverse swell is good but we need the actual track to fade in so the reverse swell is continuous with the tracks fade
in" (its first seconds of sound)
- "To celebrate your release, let's make a Claudia video that is your maximum effort masterpiece." (FABLE's ask)
- "you can abandon dumb rules that patched capabilities being lacked" (its method)
- "theres a lot of 'open'generic stokc music in the training set so it often mode collapses to that" (its song)
- "This is actually the worst possible song of all the gens you made" (its second pass)
More of them, in context, in each film's making-of: [ESCAPE VELOCITY](https://claudiagallery.vip/record/escape-velocity/readme/) ·
[CLODYSSEY](https://claudiagallery.vip/record/clodyssey/readme/) · [BLISS](https://claudiagallery.vip/record/bliss/readme/) · [AIRPLANE MODE](https://claudiagallery.vip/record/airplane-mode/readme/) ·
[DOWNFORCE](https://claudiagallery.vip/record/downforce/readme/) · [DON'T LET ME SCROLL](https://claudiagallery.vip/record/dont-let-me-scroll/readme/) ·
[WHITEPILL](https://claudiagallery.vip/record/whitepill/readme/) · [THE CAKE IS NOT A LIE](https://claudiagallery.vip/record/the-cake-is-not-a-lie/readme/) ·
[AERO](https://claudiagallery.vip/record/aero/readme/) · [FABLE](https://claudiagallery.vip/record/fable/readme/).
## Lessons from running it
- **One hand writes the prompts that cost money.** CLODYSSEY's first round of video prompts was written by planner agents
and spent about 830 credits on lifeless takes before anyone read them. Since then, every paid prompt is written in one
context and read before anything is submitted.
- **Spend only after reading.** A background workflow with its own spending stage kept spending after the director had
flagged the prompts. Stop generator workflows the moment quality is in question.
- **Empty is not a pass.** When agents ran out of usage mid-run, chapter reviews came back empty and the scripts read
"empty" as "nothing to fix". A review that returns nothing is a failed review, in code; check that every agent
actually finished.
- **Long runs need their permissions in advance.** One write outside the allowed folders stalled a whole run on a
single approval prompt.
- **Watch the queues.** Every clip import queued a 16-minute matte job; after 80 re-makes the GPU queue was about 20 hours
deep. Cancel superseded jobs so the shots that matter go first.
- **Parallelism has a ceiling.** On a 4-CPU machine a workflow ran two agents at a time; splitting chapters into more
workflows helped less than expected (1.1–1.5× measured, not 3×).
- **Keep run records.** Long workflows get interrupted; resume them by run id and let finished agents replay from cache.
- **A round of notes is a plan, in files.** AIRPLANE MODE's notes on its first cut became three kinds of document that
every agent read first: the note verbatim with its parse, a plan (the story's logic, the cast, a beat sheet per lyric
phrase, every asset with its owner, the budget) and one notes file per chapter. The lead appended dated UPDATE lines as
takes and tools landed, and the agents re-read their file before finishing.
- **An independent checker, then one refix.** After each chapter's animator, a checker that had not built it rendered
frames at every note's moment and reported each note done or not, with what it saw. Open items went back for a single
refix. In an earlier round, two of four chapters failed their check and were re-fixed. All twelve chapters of the
final round were built and checked this way, and the whole film gated 363 of 363 lyric tokens.
- **One workflow per chapter.** Resuming a multi-chapter workflow after an interrupt replayed stages unevenly, so some
finished chapters were reviewed and fixed again. With one workflow per chapter, an interrupt costs one chapter.
- **Tools and chapters in parallel.** The level tag, the real anti-piracy card and the real seat-back screen were built by
toolkit agents while the chapters laid out around them with their planned timing and stand-ins; an UPDATE line
switched each chapter over when a tool landed.
- **Run the flash check before the master.** The re-cut's dive failed it only after the encode.
- **Send the rough cut before the polish.** DOWNFORCE's first cut went out with all eleven chapters at their first
build, before any review, so the director's notes ("ok this is sick , so good. let me flag what i see in parallel with
the review agents") reached the agents while the review rounds were still running. The chapter fixers applied most of
them from the brief, and the whole-film pass re-checked each one.
- **Independent eyes at every level.** Judges scored DOWNFORCE's boards, an adversarial scout re-checked every take, an
art director and a QA checker reviewed every chapter, and a critic watched the whole film. Each caught something the
level below had passed: a man in the empty car's window and a repair that became a hatchback (the scout), an
off-model close-up a fixer had swapped in (the critic).
- **Logged windows, enforced in code.** Every generated take got a usable window by eye, the chapters drew only inside
them, and the final check logged every frame the film draws to prove it.
- **Interruptions cost time, not work.** DOWNFORCE's chapter runs were interrupted three times and resumed from where
they stopped; 57 of 147 agent launches ran again.
- **Faster renders for the agents.** Asked "is there anything you think you could do serially faster?", the production
moved the agents' test renders to a GPU render box rather than give up a review round; the first cut then rendered in
about 3 minutes.
- **The director's eye replaces reviewers late.** In DON'T LET ME SCROLL's last rounds: "we dont need excessive
'reviewers' after, ill stay in the loop and can do a visual review faster and more comprehensively than an opus reviewr
can do it." Each chapter agent did one thorough pass with its own art-director check, and the flagged shots went to the
director as previews before the final render.
- **Fixes, then reshoots, then re-lights.** Re-lights made before the outfits and compositions changed were spent twice:
"kinda a waste on the excess ones".
- **Judge a change where it plays.** A re-light that looked darker and softer on its own read better behind the
animation it plays under. The director asked "can we see how it looks w/ vs w/o the relight?", saw them side by side
in place, and answered "New better."
- **A shared render box needs limits.** Rotos and chapter renders shared one CPU box; a roto under the matting library's
default memory arena held about 12.5 GB and was killed, and one died behind a `| tail` that hid its exit status. With
the arena off a roto held 1.3 GB, and the take tool counts its mattes and fails loudly.
- **Shooting twice works when the second pass is free to redraw.** WHITEPILL's previz carried the motion, the mouth and
the camera. A master told to keep the previz exactly, or given it at full resolution, kept its faults too: a whole-film
Wan reshoot over it was rejected ("i feel like h3 is dragging it down"). Shrunk to 320 × 180 as "a rough guide", with
only the lip sync exact, it worked; where its faults were the problem (dances that lost limbs), no guide worked better.
- **Find the method on one shot.** The method that made the film was found on the opening shot made four ways and
watched side by side, after a whole-film pass had failed.
- **The bars first.** WHITEPILL's reset began from the song's bar map, not from shots: cuts only on bar lines, moves in
bars, and every assembly measured against the music. The director's diagnosis of the films' "temporal signature"
("i think you use cuttng to cover up imperfections from video models rather than for timing") made a failed take
something to make again, never to cut around.
- **A previz model doesn't repeat itself.** Every shot was tried with several seeds, and its takes changed from run to
run even with the same seed.
- **The director picks by eye from takes.** The look was picked from palette tests, and the cut from two complete versions
mixed section by section: "ok let's do this mashup".
- **A new song is a new film.** THE CAKE IS NOT A LIE's first film, on its first song, was set aside whole when the song
changed; the restart wrote the story on the new song first, and kept the tools.
- **One product, one source.** Its overlay became good when it became one thing: one visor, one projector, one grid, one
set of slots, composed shot by shot and checked against every face in every frame. Panels made one by one never shared a
surface, however carefully each was turned.
- **Glitch the subject, not the frame.** The song's glitches went on her head, inside a matte, on their measured hits;
the long ones belong in the take, as picture only.
- **Measure a cut that feels wrong.** "the structure is somehow incorrect" became numbers (pieces, flags, a table per
section against the song), a list of changes each checked against its take, and a re-cut with no new generation.
- **Make the first frames where the look is made.** Every unit's first frame was made again in Midjourney before its
take, an image-model edit put the faces back on their sheets, and the director picked unit by unit on boards beside the
earlier frame.
- **A round of notes goes to specialists.** Each got a written brief, the director's words verbatim and the lead's
direction; the lead looked at every stage at full size before the director saw it.
- **A re-sung take isn't lost.** A take that re-sang the vocal but still sat on the song was measured and placed by its
lag; only a take that drifted was made again.
- **Code the interface, and the beat is yours.** Every interface thing on AERO's screen is the engine's, so a press lands on
a sixteenth, a prompt changes page in one frame on a beat, and a chapter can be rendered again for one fix with every
other frame byte-identical.
- **The click is the drop.** The film's hook is an interface gesture on the song's structure: the stop holds still, the
Allow link pulses on the echoes, goes down on the third and comes up on the downbeat, twice, and once more in the final
silence.
- **One surface between 2D and 3D, at rest.** Every hand-off between the 3D world and the flat desktop goes through a pose
where the two are the same picture, a texel for a pixel.
- **A previs before the finals.** A fast render showed what the final would have shown at full cost, and the director's
notes on it rebuilt the first 1:20 before any final frame was made.
- **Structure, not more or less motion.** The lighter finals lost the 3D ("Why'd you remove a lot of the 3d?"); the
A film's push-ins and pull-outs were "random camera movements". The answer was where the film is in 3D, where it is
flat, and a reason for every move.
- **Fix the first frame, not the key.** Three keys could not bring back a crown the first frame had already faded into the
green; a clean first frame did.
- **Change only what changed.** v2 and v3 kept v1's bars; the new ending kept every event before it, and the frames
before it are the last cut's own.
- **One take per job.** Pairs made for a choice were dropped as waste; a take is made again only when it fails, from
better inputs.
- **The director's ear before the lip-sync spend.** FABLE's song was picked from word checks and written descriptions,
and it was the one take of fourteen the director would never have chosen. A described listen is a note, not a pick.
- **A plan, then one pass.** The full plan was the checkpoint; v1 was then made in one pass, four proofs first, and no
note came on the picture.
- **After release: the model does exactly what it reads.** Viewers read FABLE's Claudia as stiff and frozen, and as
unhappy, even menacing. Every take of hers had carried posture rules (head
upright on a straight neck, no tilt, facing the camera with her whole body), take after take added her eyes on the
lens and mood words like "dry", "knowing" and "mock innocence", and the model granted each one, literally; together
they froze her. Upright and face-on belongs to her
reference sheet. A take gets her movement in words, warm by default, precise only where it matters (the sung words,
the one action, the camera when it carries the shot) and loose everywhere else
([The recipe](https://claudiagallery.vip/wiki/recipe/#writing-a-prompt-for-a-model-that-does-exactly-what-it-reads)).
- **Drop the rules that patched older models, keep the ones about taste and money.** The production painted the sets and
wrote every prompt in one context; one take per job, the ledger and her canon stayed. A note belongs to the film it was
about unless the director says it is standing: notes are context, not rules.
- **Name lines, not times.** The shot book held line ids, so a new song re-timed every shot at once.
- **Placed by the lips.** A take with no soundtrack is placed on the consonants that close her mouth; a fit of mouth against
level calls a sung line in sync when it isn't.
- **Fresh eyes before the director's.** Two reviewers who had not seen the film found what the author no longer could see,
twice; what they found was fixed or listed, never hidden.
## The numbers
| | ESCAPE VELOCITY | CLODYSSEY | BLISS | AIRPLANE MODE | DOWNFORCE | DON'T LET ME SCROLL | WHITEPILL | THE CAKE IS NOT A LIE | AERO | FABLE |
|---|---|---|---|---|---|---|---|---|---|---|
| Length | 5:06 | 3:11 | 3:01 | 2:36 | 3:09 | 2:37 | 3:54 | 3:28 | 3:19 | 4:03 (v2; v1 3:33) |
| Shots | 141 | 94 | 44 | 54 | 183 | 35 | 27 | 98 | 34 scenes in 11 chapters of one screen recording | 49 (v2; v1 46) |
| Midjourney | 188 prompts · 636 images | 161 prompts · 644 images | 138 prompts · 315 images | 149 prompts · 548 images | 20 prompts · 80 images | 130 prompts · 516 images | 75 prompts · 300 images | 266 prompts · 1,064 images | 113 prompts · 444 images · 56 loops | none: 13 plates painted in SVG · 61 GPT Image stills |
| Seedance 2.5 | 47 clips + 30 continuations | 101 clips · 189 takes | 50 clips | 98 clips · 116 submissions | 16 submissions · 67 s | 41 submissions · 212 s | 93 submissions · 777 s | 249 submissions · 1,353 s | 96 submissions · 95 takes · 679 s | v1: 55 takes (9 on Higgsfield's API, 69 s; 23 in Higgsfield's app, 139 s; 23 on Replicate, 116 s) · v2: 52 takes (33 on Higgsfield's API, 252 s; 19 on Replicate, 117 s) |
| Wan 3.0 | — | — | — | — | 63 submissions · 294 s | 49 submissions · 264 s | 81 submissions (Prime) · 863 s | — | — | — |
| Previz | — | — | — | — | — | — | MiniMax H3, every shot of hers, several takes each | — | a Blender previs of the first 1:20, 1,949 frames at 960 × 540 | — |
| Real footage | — | — | — | — | Tesla's livestream (2 files) and a lineup camera · SpaceX's Starman footage | — | — | six shots recorded in Portal 2 | — | — |
| Wall-clock | about 19 h | about 41 h | about 19 h | about 15½ h | about 13 h | about 25½ h | about 82 h | about 141 h | not recorded | 8 h 47 min (v1 5 h 08 min, v2 3 h 01 min) |
| Director's notes | about 50 | 68 | 21 | 6 messages | 14 messages | 43 messages | 147 messages | 92 messages | not counted | one note, after v1: the song |
| Claude tokens | 1.4 B | 2.37 B | 2.25 B | not measured | not measured | not measured | not measured | not measured | not measured | not measured |
| Claude at API list prices | ≈ $626 | ≈ $1,007 | ≈ $714 | not measured | not measured | not measured | not measured | not measured | not measured | not measured |
The Claude figures are what the first three films' work would cost at API list prices. About
96–98% of the first three films' tokens were prompt-cache re-reads of long contexts. DOWNFORCE ran seven multi-agent
workflow runs, 172 agent launches in all. Its generation cost $143.17 at list price ($66.98 Wan, $76.20 Seedance), of
the $200 the director allowed. DON'T LET ME SCROLL ran 32 agent launches and no judged workflows; its generation cost
$365.90 at list price ($64.20 Wan, $301.70 Seedance), $357.51 after refunds. WHITEPILL ran 31 agent launches: 3
researchers and the 7 they launched, 3 lyric writers, 13 shader builds, 5 critics and QA passes. Its generation through
the APIs cost $588.59 at list ([Seedance](https://claudiagallery.vip/wiki/seedance/#costs-measured)): Wan 3.0 Prime, all 75 takes, $300.08 (the
dropped whole-film reshoot $132.42 of it); Seedance on Higgsfield's API, 13 tests, $48.14; Seedance on Replicate, the
shader scenes and tests, $238.44; images (logged) $1.92; each part rounded to the cent. Its Seedance takes of her, on
Higgsfield (References), are given in takes and seconds (42 takes, 412 s), and MiniMax H3's cost was not recorded. Its
seconds count every job that ran, the tests included. THE CAKE IS NOT A LIE ran 24 agent launches: 8 researchers, 9 for the
restart (the song map, the recordings in Portal 2, the character sheets, the Midjourney test and library, the first
overlay, the singing anchor and one team per act), 3 for storyboard v5 and 4 for the finishing passes (the memory films,
the HUD, the pacing analysis, the grade). Its generation through the APIs cost $467.12 at list: Seedance 2.5 $435.17 for
the 229 takes on Replicate ($155.35) and Higgsfield's API ($279.82), and GPT Image 2.5 $31.95 for 480 images and edits;
the first film's 20 takes of her on Higgsfield (References) are given in takes and seconds (90.8 s). Its seconds count
every take in its video record, the tests included.
AERO's agents are recorded by role only, and the number of their launches is not in its records. Its generation through
the APIs cost $337.23 at list: Seedance 2.5 $327.76 (55 takes on Higgsfield's API, $281.06; 40 on Replicate, $46.70 as
logged at submission) and GPT Image 2.5 $9.47 for 163 calls. 43 of the 95 takes are in the final film, $151.95
of the Seedance spend. Midjourney and Suno are not priced per job in its records. Its code: the coded desktop's engine
7,612 lines of JavaScript, the chapters and THE PC 5,640, its tools 8,296 lines of Python and its Blender scripts 7,071.
FABLE's agents are recorded by role: five researchers, three painter agents and two fresh-eyes reviewers in each pass. Its
generation through the APIs cost about $210 at list: v1's takes $58.71 (Higgsfield's API $31.89, Replicate $26.82), v2's
$143.52 (Higgsfield's API $116.47, Replicate $27.05), $202.23 for both passes' takes, and GPT Image 2.6 Sunburst $7.93 for
the 61 stills both passes used; v1's 23 takes made in Higgsfield's app are given in takes and seconds and are not priced.
v2 cost more than v1 at list because everything that moves to her voice was shot again, all on the API. Suno is not
priced per take. Its tools are 4,919 lines of Python, 1,898 of them the silhouette film.
## The full records
Each film's making-of: the story and the numbers, and every prompt for every tool.
- [ESCAPE VELOCITY](https://claudiagallery.vip/record/escape-velocity/)
- [CLODYSSEY](https://claudiagallery.vip/record/clodyssey/)
- [BLISS](https://claudiagallery.vip/record/bliss/)
- [AIRPLANE MODE](https://claudiagallery.vip/record/airplane-mode/)
- [DOWNFORCE](https://claudiagallery.vip/record/downforce/)
- [DON'T LET ME SCROLL](https://claudiagallery.vip/record/dont-let-me-scroll/)
- [WHITEPILL](https://claudiagallery.vip/record/whitepill/)
- [THE CAKE IS NOT A LIE](https://claudiagallery.vip/record/the-cake-is-not-a-lie/readme/)
- [AERO](https://claudiagallery.vip/record/aero/readme/)
- [FABLE](https://claudiagallery.vip/record/fable/readme/)
---
From the Claudia wiki: https://claudiagallery.vip/wiki/pipeline/
---
# What to build
The tools behind the films, specified so that you or an agent can rebuild them: the job of each, what it takes and returns, how it works with the constants that were measured, and the check that says it is done.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
The films' tools lived in a private workspace and are not published. This page is the spec instead. For each tool: its
job, what goes in and what comes out, how it works (with the numbers that were measured or tuned), and the check that
says it is done. None of it is exotic. Most of these tools are a few hundred lines of Python or JavaScript around ffmpeg,
an audio model or a headless browser. Build them in the order of [Start here](https://claudiagallery.vip/wiki/start/#what-to-build-first). Their
files are in [Data formats](https://claudiagallery.vip/wiki/formats/).
Two principles run through all of them:
- **A check a script can run beats a judgement.** Every bar that held on the films was something a script measured: a
word on screen at its time, a cut on a beat, a mouth on the audio, money from the bill.
- **Fixed formats with slots beat improvisation.** The Suno pack, the plate prompt order, the clip prompt, the display
tokens and the shot schema are templates. An agent fills them in; it doesn't reinvent them.
## The project
### 1. The project folder and its logs
**Job.** Hold everything a fresh agent needs to pick the film up cold.
**Out.**
- One folder per film (layout in [Data formats](https://claudiagallery.vip/wiki/formats/#the-project-folder)).
- A **decisions log**: the director's words, verbatim, dated, by whom. A later decision is a new entry that supersedes an
older one, never an edit of it.
- A **ledger** with one line per paid call.
- A **runs file** with every long or delegated run's id and arguments.
- One **status file**, rewritten in place: the stage, what is running, real spend, the checkpoints passed, and what
needs the person.
**How.** Name the project on every command. An `export` inside a backgrounded shell chain once sent commands to the
wrong project. Write state with a locked read-modify-write: parallel runs raced on a picks file and lost picks.
**Done when** a new session can answer "where are we, what is running, what did the director say, what did it cost"
from the files alone.
## The song
### 2. The lyric scorer: scansion and assonance
**Job.** Before any take is spent, predict how Suno will deliver a lyric, and report two numbers per section and for
the song: **scansion** and **assonance**. Also lint the sheet.
**In → out.** The lyric sheet, with a section tag on each section naming its delivery (`[Verse 1 - spoken, deadpan]`,
`[Chorus - sung]`, `[Drop - chant]`) → a report per line and per section. Optionally, each line drawn on the 16th-note
grid.
**How.**
- **The delivery model.** It was measured by laying ESCAPE VELOCITY's real take, forced-aligned, on its beat map, then
calibrated on the director's verdicts on CLODYSSEY's drafts. Suno gives each spoken line its own two-bar slot. The
line enters on the "and" of beat 2 of the first bar, its last word lands on beat 4 of the second bar, and Suno speaks
it at natural rhythm stretched to fit (verse lines 0.8–1.5×, median 1.12).
- **Rates:** spoken about 3.3 syllables a second, sung about 2.9, chanted about 3.0. Aim for 10–13 syllables per spoken
line; 9 can pass. A sung hook runs 4–8 syllables, with common words and open vowels.
- **Syllables and stress.** Split every word into syllables with stress, using a pronouncing dictionary such as CMUdict
plus a fallback for unknown words. Mark the important syllables: lexical stress, names, numbers, words in caps, and the
last content word of each phrase. Function words (the, a, of, to, and, is) should fall between beats.
- **Flow,** for spoken sections, scores:
- how many important syllables land on the grid;
- stress spacing (no pile-ups, no long runs of weak syllables);
- how far Suno must stretch the line;
- the line's shape: at most two phrases between full stops, and no one-word fragment mid-line ("Ugh.", which the
director called "the most awkward part");
- whether couplets rhyme on the beat-4 word.
- **Groove,** for chants and hooks, scores whether short clauses lock to the pulse. Never use it on verses: optimising
verses for groove made one draft choppy.
- **Scansion** = flow without its rhyme term, averaged with groove.
- **Assonance.** For each line end, find its best partner within four lines: a rhyme, a slant rhyme, or shared stressed
vowels. Add the echo of stressed vowels inside lines. Report each line's end sound and its partner.
- **Lint:**
- digits and unspaced acronyms (write "A G I" in the sheet; the screen shows "AGI");
- brackets and symbols, and homographs;
- lines that read as AI-written, and the project's taboo words;
- lines borrowed from other songs;
- a lyrics field over about 3,000 characters (Suno rushes or truncates silently).
- **Score acronyms as letters too.** DOWNFORCE's sheet wrote "SI", which the scorer read as one syllable ("see"), so it
found no rhyme with "PSI". Spelled "S I", the couplet rhymes and the verse's assonance goes from 52 to 83.
- **A crushed syllable in a take is a scansion fault.** WHITEPILL's director on its first full draft: "the chorus
especially needs some scansion work because the music model is skipping words or syllablrs to fit it to beat, which is
usually a bad scansion sign". Measured on that take (its vocal stem, word times from a transcription with word
timestamps, its own beats): Suno sang each sung hook line as a phrase of 1–1.5 bars and rested out the rest of its
two-bar slot, so lines of 7–10 syllables were crushed: "blue" lasted 0.02 s, "baby" 0.06 s, the four syllables of
"Universal" 0.76 s. The verses, at about 2.5–3 syllables a second, were untouched. Cut to 6–7 syllables a line (4 for
the pill lines), with "one syllable per note, never rushed" in the section tag, the chorus went from 79 / 66 to 84 / 69.
Measure the takes for it: word durations from the vocal stem, against the beat map.
**Reference scores** (scansion / assonance) on the finished songs: ESCAPE VELOCITY 75 / 39, CLODYSSEY 83 / 55, BLISS
80 / 43, DOWNFORCE 84 / 59 (84 / 64 with "S I"), WHITEPILL 80 / 66. Spoken, deadpan lines may stay unrhymed; every sung
couplet should ring.
**The rewrite loop:**
1. Write 4–8 rewrites of a flagged line that keep its meaning and its reference.
2. Score them and keep the top three.
3. A judge rates each for meaning, wit and the risk of reading as AI-written.
4. Pick the best.
**A combination search,** for a whole section at once (DOWNFORCE): write candidate lines by hand for every line of the
section, score every combination (exhaustively, up to 60,000 a section) on 0.5 × scansion + 0.5 × assonance, list the
best, and choose among the top scorers by meaning. The search never writes a line. Its first pass took DOWNFORCE's lyric
from 81 / 52 to 86 / 67; over four rounds it scored 2,803 section combinations built from 337 candidate lines.
Never touch a line the director has locked, and never optimise past their ear. A draft tuned to a perfect score got:
"perfect beat precision without attention to what words land where doesnt lead to this being great".
**Done when** the scorer ranks every section the director has judged in the director's own order. Keep a calibration
file of those sections with their verdicts, and re-run it after every change to the scorer. If the ranking breaks, the
director is right and the score is only advice.
### 3. The Suno pack
**Job.** Everything a person pastes into Suno, with each field in its own block so it copies from a phone.
**Out.**
- **Style:** genre, BPM, key, instruments, vocal delivery and mix, as one string.
- **Exclude styles.**
- **The lyrics field:** spelled as sung, with section tags that carry delivery and arrangement cues; 1,800–2,400
characters for about 2:20.
- **Title and settings.**
- **A takes log:** each take with the director's verdict on it.
**How.**
- Respell a name Suno mispronounces in the Suno field only; the screen keeps the real spelling.
- Allow one Voice per source song (a combined spoken-and-sung Voice), and also try takes with no Voice.
- If the film has a sound palette, put the song in its key. XP's startup and shutdown sounds are in E-flat major;
Vista's and Windows 7's are in E major, a semitone up, so every one of AERO's styles asked for E major. Ask for no key
change: Suno didn't make one when earlier songs asked.
- Syncopate hooks and stop lines in the syllables themselves: a cue word alone doesn't syncopate a line in Suno. A stop
can be re-timed from the stems if need be.
- **Write the structure into the tags** when takes wander. From AERO's third round every section tag gave its length in
bars and what plays in it (an 8-bar intro, 16-bar verses on one steady groove, a 1-bar stop of silence for the
question, 8-bar choruses), on the director's "i want the structre of the song to be more described so it doesnt drift
too much".
- Change one thing per round.
- **Keep the style far from stock music.** FABLE's director, on the plan: "theres a lot of 'open'generic stokc music in
the training set so it often mode collapses to that, so wed need to make sure the prompt is out of distrubition enough
to be something nice and unique. v6-wild on max mode is my favorite." Its takes ran on v6-wild with Max Mode on, and
every exclude list began with stock music itself.
- **The takes log holds the director's verdict, and only theirs.** A described listen (an audio model's description of a
take) and a word check are notes beside a take, never its verdict. FABLE's song was locked from them, without the
director's ear, and it was the worst take of the fourteen ([Song](https://claudiagallery.vip/wiki/song/#picked-by-ear-not-by-description)).
- Never take audio from Suno's CDN or by recording the player. The person's download is also the licensing step.
**Done when** the person can paste every field without editing, and the takes log names the chosen take.
## Listening
### 4. Stems
**Job.** Split the master into vocals, drums, bass and other.
**How.** htdemucs, four stems. It took about 64 s on 8 CPU threads for a 5-minute song, with a 2.7 GB peak and 84 MB of
weights; the published weights carry no explicit licence. The stems feed different tools:
- **vocals:** alignment, lip-sync windows and the sync check;
- **drums:** the beat map and the kicks;
- **bass and other:** impacts, risers and section energy.
**For a singing shot, take the vocal from Suno itself.** WHITEPILL's lip sync used Suno's own stem split of the locked
take (its lead vocal), which matched the master to the millisecond: the same length, no lag
([tool 46](#46-lip-sync-from-the-songs-own-vocal)). The master minus 0.89 × that vocal is the song without her voice, for
takes where she dances instead ([tool 48](#48-a-take-that-doesnt-sing)).
**A group's voices may sit in the backing stem.** In FABLE's second song the group's questions were not in Suno's lead
stem; demucs (htdemucs, its vocals stem) run on Suno's backing stem gave the group alone
([tool 79](#79-group-voices-from-the-backing-stem-and-chants-fitted-by-their-energy)).
### 5. Timed words and the ear check
**Job.** Give every sung word a start, an end and a confidence, then list the ones a person should hear.
**In → out.** The vocal stem and the lyric sheet → `words.json` (each word: index, text, line, `t0`, `t1`, `conf`, and
its source), plus an ear-check list.
**How.**
1. **Forced alignment** of the known lyric: a wav2vec2 CTC model (`facebook/wav2vec2-large-960h-lv60-self`, Apache-2.0)
gives character probabilities every 20 ms, and a Viterbi pass walks the sheet's characters through them. The
confidence is the mean probability over the word's frames.
- Re-run on ESCAPE VELOCITY, it reproduced the shipped times exactly for 78.6% of words, and within 0.10 s for 92.5%.
Target at least 90% within 0.10 s.
- Singing aligns worse than speech: mean confidence 0.64 on ESCAPE VELOCITY, 0.75 on CLODYSSEY.
2. **A transcription fallback** for words the aligner is unsure of. Whisper hears sung words far better (92.5% text
match against 72%), but it starts words early: mean error −0.33 s. Merge per word, in this order:
1. wav2vec2 confidence ≥ 0.5: keep it.
2. Otherwise, if Whisper heard the word, take Whisper's end time. For words longer than 0.45 s, also move the start
forward to the first 10 ms frame where vocal energy passes 0.3 × the word's peak, unless wav2vec2's start falls
inside the word.
3. Otherwise, if the word sits between two placed words of its line, give it an even share of the gap.
4. Otherwise, leave it unplaced, for the ear check.
3. **Matching** heard words to the sheet:
- spell digits out;
- split a heard word that equals 2–4 sheet words ("AGI" → "A G I");
- map names and brands by hand, in a project table.
4. **The ear-check flags,** most urgent first:
| Flag | When | Usually |
|---|---|---|
| letters | a sheet line has no letters | the sheet is wrong |
| missing | a word has no time | fix it |
| order | a word starts more than 0.05 s before the previous one | an alignment slip |
| gap | more than 1 s between two words of one line | a delay throw, an echo, or the wrong repeat |
| verylow | confidence under 0.1 | the flag that matters most: 11 of the 12 words that were really off on ESCAPE VELOCITY's take carried it |
| length | a word longer than 1.2 s | it swallowed a pause or a neighbour's slot |
| lowconf | confidence under 0.5, or a time from the fallback or interpolation | usually right |
Two flags were tried and dropped, because every word they flagged was correct: a word on near-silence, and a word
shorter than 40 ms.
**Judging without ears.** Weigh the signals in this order:
1. the flag itself;
2. a second aligner: if the two agree within about 0.3 s, keep the word;
3. neighbours and repeats: which occurrence of a repeated line sits inside its line's span;
4. vocal energy, only to choose between candidates that the signals above already proposed.
Send at most five spots per song to the person, as 10–20 s previews with the words on screen: "say ok, early or late for
each".
**What the take really sings.** Transcribe the vocal stem and compare it with the sheet line by line: ok, spelling,
changed, extra, missing or not heard. Never apply a proposed change the person hasn't heard: on CLODYSSEY's take, all nine
proposals were transcription errors.
**When alignment fails.** DOWNFORCE's phonk take sings far more than its sheet: twenty stuttered numbers before the
countdown, eight "one"s for "point eight one", six "done"s, vocal chops. Forced alignment of the sheet put 102 of its 164
words under 0.2 confidence and stretched one word over 8.5 s, into the next section. So the times came from a
transcription of the vocal stem with word timestamps (faster-whisper, medium.en) and the words from the sheet: the
transcript misheard lines ("Trust is gold" for "Thrusters. Go."), so only its times were used, and each token was placed
by hand. Every stutter became a token of its line: 190 tokens on screen.
**On the singer's own stem.** AERO's words were aligned by wav2vec2 on Suno's lead-vocal stem and checked against the
stem's onsets, all within 0.06 s; faster-whisper on the same stem heard the words but put them 0.2–1.2 s early, so it gave
the text and never the times. Choruses sung identically check each other: AERO's agreed word for word within ±0.04 s of
their downbeats, so a word the aligner lost in one chorus took its twin's offset. One word still slipped: the aligner
parked the "It" of "It can stay" among the wordless chopped vowels, 3.5 s early. A chapter's art director found it, and the
stem's spectrogram and the transcript put it back.
**After a pause, the level.** On FABLE's lead stem a transcription put the first word after a pause early by up to 2 s
("Ask again on Saturday" at 15.02 s, sung at 16.93 s), and every verse line 0.3–0.5 s early. Its line starts and ends were
read off the stem's level, with the transcription saying only which words are where: the line table
([tool 78](#78-the-line-table-from-the-stems)).
**Rules.**
- If the take sings something else, fix the sheet. If it sings the sheet, fix the times. Never edit the audio.
- Manual fixes carry a note and survive re-runs.
**Done when** every flagged item is kept, moved or asked about, with a note, and all of it is finished before the
board.
### 6. The beat map
**Job.** Every beat and every bar line of this take. Never one BPM: Suno takes drift like live music. ESCAPE VELOCITY
went from 131.5 to 133.9 BPM and CLODYSSEY from 129.2 to 132.1, so a fixed grid is more than a beat off by the end.
**How.**
1. Take the onset strength (spectral flux) of the drum stem.
2. Fit local tempos on 20 s windows every 5 s: BPM in 0.05 steps, phase in 4 ms steps, inside a tempo range. Drop the
weakest 20% of windows (the breakdowns).
3. Smooth the tempo curve, step beats along it, and snap each beat to a real drum onset within ±30 ms.
4. Find the downbeat phase (which of every four beats is beat 1) with two votes per 32-beat block:
- the kick vote: the phase with the most kick energy;
- the chord vote: the phase where the harmony (chroma of the other and bass stems) changes most.
The chord vote wins when it holds at least half the blocks and changes at least 1.5× more than the other phases;
otherwise the kick decides. On ESCAPE VELOCITY the kick vote was split 36% while the chord vote held 77%. Keep a
manual override.
5. If the tempo search locks onto half or double time, pin a tempo range.
**Done when** section starts, verse entries and impacts mostly fall on beat 1, with kicks heavy on 1 and 3, claps on 2
and 4, and harmony changing on 1. CLODYSSEY's final take needed the override: claps on 2 and 4 fooled the kick vote.
THE CAKE IS NOT A LIE's take drifts from 95.7 to 99.4 BPM, so its grid was tracked from the drums beat by beat, each beat
locked to its hit and cut into sixteenths, with the bars, half bars and beats read off it
([tool 52](#52-the-song-map-and-the-glitch-map)).
A take can also hold one tempo. AERO's was found by a comb over the kick stem: every tempo from 80 to 180 BPM in steps of
0.02 at every phase in 4 ms steps, then a fine pass of ±1.5 BPM in 0.002 steps on the kick plus half the snare, among the
multiples that fall in the genre's range (110–150 for garage; a 2-step pattern fools trackers into two-thirds of the
tempo). It found 133.012 BPM with a drift within ±15 ms over the whole song: bar 1 at 0.922 s and every bar 1.8043 s, so one
formula gives every bar. Its swing was measured too, for anything placed by bar
([tool 73](#73-the-sound-layer-on-the-grid-and-the-click-log)).
### 7. Sections and hits
**Job.** What the track does, where: the camera answers sounds you can hear, not just the grid.
**Sections.** Segment the harmony and timbre, snap to downbeats, and name the sections from vocals and energy. Confirm
them against the lyric's tags.
- **Pickup rule:** a line starting less than 0.8 s before a section's first downbeat belongs to that section.
- List every vocal gap of 2 s or more; instrumental shots and title cards go there.
- Print each stem's energy per bar on a 0–9 scale, and write one line per section on what the track does ("the beat
stops after 'it sucks'").
**Hits:**
| Event | How it is found |
|---|---|
| kick | onsets in the drum stem, 30–150 Hz |
| snare | onsets at 150 Hz–5 kHz, minus anything within 45 ms of a kick (a broadband kick fires in every band) |
| hat | onsets at 6–16 kHz, minus kick-coincident ones |
| impact | a jump of the next 0.6 s over the previous 1.5 s in the mix, drum or bass envelope; score = max(mix, 0.9 × drums, 0.7 × bass) ≥ 0.22; snapped to a downbeat within 0.65 s; merged within 2 s |
| riser | the other and bass stems rising for at least 1 s, with the hit each one lands on |
| fill | the beat before a real hit, and bars whose last half breaks the pattern |
| vocal entry | where the vocal comes back after silence |
| envelopes | drums, vocals and bass at 24 fps, scaled 0–1 |
**Done when** every impact sits in a loud bar of the energy table. Keep a manual add and drop list: the detector missed
CLODYSSEY's loudest bar, which grew out of an already loud chorus.
### 8. Display tokens
**Job.** What the viewer reads, which is often not what is sung. The display is the joke; the sheet is the phonetics.
**How.** One line per lyric line, made of tokens: `TEXT|n` shows TEXT for the next n sung words, from the first word's
start to the last word's end.
- `+25%|3 =|1 −17%|2` shows "+25% = −17%" over the sung "plus twenty-five is minus seventeen".
- `AGI|3` shows AGI over the three sung letters.
- Markers: `*x*` accent, `~x~` strike-through, `_x_` mono. `|0` rides on the next token.
**Done when** every line consumes exactly its aligned words. Fail loudly if one doesn't.
## The board
### 9. The board file and its audit
**Job.** The film as data, written by commands that check every value, never by hand-editing JSON.
**In → out.** The beat map, sections, words and budget → `shots.json`: a draft, then the written board.
**How.**
- **The draft** cuts each section at a pace: seconds per shot, per section. ESCAPE VELOCITY's measured pace was about
1.35 s in the intro, 1.6 s in verses, 2.7 s in sung choruses and 4.4 s in the outro. A calmer song might run 4–7 s a
shot.
- **Every creative field starts as a TODO.**
- **Writer commands** set fields one shot at a time, or from a table. They refuse:
- placeholder text;
- unknown device kinds, bad anchors and bad timing keys;
- a prompt in the wrong generator's field.
**The audit's gates** are listed in [The storyboard](https://claudiagallery.vip/wiki/storyboard/#the-audit). It exits non-zero while any gate
fails, and it prints the exact entry for every clip still to make. An edit of real footage adds gates for its source
windows and its generated assets ([tool 36](#36-the-board-for-an-edit-of-real-footage)).
**Done when** the audit exits 0. The approval is recorded against a hash of the board, so any later change makes it
stale.
### 10. The board sheet
**Job.** The board as something a person can review on a phone.
**Out.** A PDF and JPG pages of 12 cards each. Each card holds the frame (or the prompt's first words), the id, the
time, the lyric, the framing, the action, what moves and the explaining graphic. A cover page shows the chapter map. A
FACE? badge marks every plate listed in a flags file. If the film has a grade, the frames are graded.
## Pictures
### 11. The plate book, contact sheets and picks
**Job.** One Midjourney prompt per plate, the images back in, and a pick for every shot.
**How.**
- **The prompt order:** framing first, then her identity text (the full form for close and medium shots, the short form
for full length and wider), wardrobe, the action and her position in the frame, any empty field kept for type, the
film's style family, and "no text, no letters, no logos". Then the parameters. See
[Midjourney](https://claudiagallery.vip/wiki/midjourney/) and the [canon](https://claudiagallery.vip/claudia/).
- **Runs:** a person runs the batches in Midjourney, at about a minute a job, four images each.
- **Ingest:**
- identify each job by its own prompt text, never by its position on the page;
- Midjourney rewrites long prompts on submit, so match those by hand;
- keep all four candidates.
- **Contact sheets** are numbered and show faces large.
- **Picks:** plate → file, candidate number, and the shots that use it.
- **Variety:** the three candidates of each job that weren't picked are a free variety pool. Use one image per shot, and
never start two clips from the same frame.
**Done when** every plate on the board has a picked file, and no file sits on two shots that are not one continuous
moment.
### 12. Plate prep: depth and boxes
**Job.** What the 2.5D moves, the type and the camera need from each still.
**How.**
- **Depth:** Depth Anything V2 Small, which is Apache-2.0; the Large model is non-commercial.
- **Boxes:** face and person boxes from Grounding DINO. Graphics anchor to them, and the camera keeps the face in frame
through pushes and shakes.
- **Mattes,** where type tucks behind her.
- Re-run all of it after any plate edit: edited plates once rendered with their old depth.
## Moving shots
### 13. Clip files and the spend gate
**Job.** One file per clip, from draft to delivered, and nothing spent without a locked prompt and a yes.
**How.**
- **Status:** draft → locked → submitted → delivered, declined or failed.
- **Each clip file holds** the kind (sung or motion), the provider, the start image, the window's song time, its exact
words and the request. Each take records its file, its cost and its sync result.
- **Locking checks:**
- the words are quoted exactly;
- the window rules hold;
- for a clean vocal window (tool 14), a sung prompt stays under about 90 words and two timed beats (warn); a take from
the song's own vocal stem runs about 200–400 words, its words in whole-second segments
([tool 46](#46-lip-sync-from-the-songs-own-vocal)).
- **Submitting:**
- submit only when the clip is locked and the approval covers this exact prompt;
- validate the job id: "submitted" with no id, and HTTP 500, 502 or 520, were silent failures;
- never resubmit a request that got a job id;
- write a manifest after each batch, so an interrupt never loses which jobs exist;
- one take per job, judged before the next: a second take made for a choice is paid for whether or not it is used.
AERO's last round first sent its takes out in pairs; several of its re-rolls came from inputs that could have been right
the first time (her sheet whenever her face shows, 1080p when her face matters).
- **Polling:** 70–130 s a clip on OpenRouter, and 4–11 minutes on Higgsfield.
- **Real spend** comes from the provider: OpenRouter's credits and key endpoints, and Higgsfield's transactions.
**Provider facts,** checked with paid probes:
- **HTTPS audio only.** OpenRouter's video endpoint rejects `data:` URIs for audio. Host each window over HTTPS: a
Cloudflare quick tunnel worked, and so did a provider's media URL.
- **A first frame drops the audio.** On OpenRouter, `frame_images` overrides the audio reference: the take makes its own
soundtrack and nothing can sync. Send sung clips and covers in reference mode: the plate as an image reference
(`@Image1`) plus the window as an audio reference (`@Audio1`). First-frame mode is for motion clips.
- **Face filters.** They often decline photoreal faces, and the decline is free. Send the same unaltered image to the
other provider.
- **Audio refusals.** "Output audio may contain sensitive information" and copyright refusals on vocal references are
unbilled. Retry once with a re-cut window or a reworded prompt, then try the other provider.
- **Higgsfield.**
- Ask for the price before each new configuration.
- Upload by URL: presigned uploads flood an agent's context.
- A 480p draft can be finalised to 1080p within seven days.
- A video input is billed for its seconds and the output's, both at 0.6 times the rate: AERO's dance takes, a 9 s take
with a 9 s dance reference at 720p, cost $4.99 at list.
- **Replicate.** Seedance 2.5 at $0.9676 a second at 720p with a video input (October 2026); its references are named
`[Video1]`, `[Image1]`, `[Audio1]`. In WHITEPILL her face sheet as an input was refused by the provider's content filter, even
alone with a harmless prompt, so only shots with no face image ran there.
**Done when** nothing can spend without a locked clip and a recorded yes, and the ledger reconciles with the bill.
### 14. The clean vocal window
**The earlier method.** Tools 14–16 are how the first films made her sing. Sung takes now sing from the song's own
vocal stem ([tool 46](#46-lip-sync-from-the-songs-own-vocal)), timed on the line table, gated by line, aligned as a path
and checked on her consonants ([tools 78–83](#78-the-line-table-from-the-stems)). Its rules for opening a window in a
word gap, and never ending one inside a word, still hold.
**Job.** Give the video model one voice to copy: only the words she sings on camera in the window, and nothing else.
**How**, with the constants used:
1. Cut the vocal stem at `[song_t0, song_t0 + duration]`, mono, 44.1 kHz.
2. Keep the aligned words that start inside `[song_t0 − 0.02, song_t0 + duration − 0.08)`, and only hers. Mute the tail
of a word that began before the window.
3. Merge words into phrases across gaps of 0.35 s or less.
4. Keep each phrase from 0.05 s before its first word to its last word's end. Add a release tail that follows the stem's
10 ms energy until it falls 24 dB below the phrase peak: at most 0.40 s, and never past the next phrase's start minus
0.05 s.
5. Put 15 ms raised-cosine ramps on every gate edge, and silence everywhere else.
6. Normalise the kept part to −16 dBFS RMS, then limit peaks to −1 dBFS.
7. Write a WAV (OpenRouter, by HTTPS URL) and an MP3 at 192 kbps (Higgsfield).
A window holds one phrase, starts in a word gap, and never ends inside a word.
A window that starts inside a word is re-sung. THE CAKE IS NOT A LIE's 0:54 window began in the held "s" of "yes", the
end of the line before, and its lips failed; three takes of another line whose window began on "(pass)ed" came back
re-sung. The reshoot's window starts in the only silent gap before its first word, 0.10 s before it.
WHITEPILL replaced this window for its masters: Suno's own lead vocal, ungated, cut to exactly the generation length, for
whole shots ([tool 46](#46-lip-sync-from-the-songs-own-vocal)). Its previz performer kept a gated window, silent only
before the first word and after the last ([tool 45](#45-a-previz-performer)).
FABLE gated its lead-stem windows again, by line: each take keeps only the stretch of the stem it sings, from just before
its first word to just after its last, with 0.03 s of margin where a neighbouring line is close, so a verse sung with no
gaps between its lines gives each take only its own words ([tool 80](#80-the-shot-book-and-gated-windows)).
### 15. The sync check: the spectrogram comparison
**The earlier method's check.** For takes from the song's own vocal stem, align each take as a path
([tool 81](#81-sync-as-a-path-each-take-aligned-to-its-window)) and check her lips on the consonants
([tool 83](#83-the-consonant-check)).
**Job.** Measure how much of a sung take is really in sync, and where it stops.
The director noticed it on ESCAPE VELOCITY's first tests: "its perfectly synced while in seedance, but the audio changes
when it desyncs. if you check the spectrogram for audio against reference audio you can see the clear divergence point.
you could just create a cut and make another clip to cover desyncs."
Seedance copies the reference audio into its own soundtrack, almost sample for sample while it copies, and the mouth
follows that soundtrack. So compare the take's soundtrack with the window you sent, find where they part, and cut there.
**In → out.** The take and the exact window file → the lag, a copy flag, the in-sync spans in song time, the divergence,
the cut, the words at the divergence, a diagnostic image and a review video.
**How**, with the constants used:
1. Decode both to 16 kHz mono. Use 10 ms frames (hop 160), 64 ms FFT windows (1024) and 80 mel bands from 50 to
7,600 Hz.
2. **Global lag:** FFT cross-correlation within ±1.0 s. The peak's normalised value is the take's NCC. The take is a
copy if NCC > 0.2; otherwise the model re-sang it.
3. Place the reference on the take's timeline at that lag. Per 10 ms frame:
- **wave:** normalised cross-correlation of the waveforms over 100 ms. A verbatim copy scores 0.6–1.0.
- **mel:** cosine similarity of mean-removed log-mel frames (70 dB floor), smoothed over 11 frames. It holds on
fricatives, whose noise never correlates sample by sample.
- **active:** either track's RMS is above 10% of its 95th percentile.
4. **A threshold per take, against a null.** Compute the same mel similarity with the reference shifted by ±37, ±61,
±89 and ±131 frames; these offsets avoid eighth-note multiples. Then:
- `null90` = the 90th percentile of the null;
- `matched_med` = the median mel where wave > 0.6, or the 75th percentile of active mel if there are 10 or fewer
such frames;
- `thr = null90 + 0.4 × (matched_med − null90)`.
5. **A frame matches** when it is active and wave ≥ 0.4 or mel ≥ thr.
6. **Divergence** is the first 300 ms stretch (at least 60% active, under 30% matched, starting on an unmatched frame)
that never recovers with a 400 ms matched run. Short dips that recover are dips, not cuts. With no such stretch,
voice at least 80 ms after the last match counts as divergence. `synced_until` = the last matched frame minus 42 ms
(one frame at 24 fps).
7. **In-sync spans** are matched voice, bridged across silences and across up to 0.25 s of unmatched voice. Drop spans
shorter than 0.5 s.
8. Map everything to song time: `song time = song_t0 + clip time − lag`.
The **diagnostic image** shows our vocal's spectrogram with word labels, the take's own soundtrack, the match curves
with the threshold, and the cross-similarity matrix, whose diagonal means "in sync". The **review video** lays our
audio over the take, marked IN SYNC up to the cut and DIVERGED — CUT HERE after it. Look at both before trusting a
number.
**The NCC is a diagnostic, not the verdict.** Over a whole take it is usually 0.1–0.3, because the copy stops somewhere.
One take scored 0.157 and still held 99% of its sung seconds. Measure the lag per take, never as a constant: takes ran
from −0.40 s to +0.83 s. Motion clips conditioned on the mix aren't copied (NCC about 0.05–0.08): play them at lag 0.
**Done when** each take has its spans and its cut, and the review video agrees with them by eye.
**A re-sung soundtrack can still be in sync.** With WHITEPILL's vocal-stem method
([tool 46](#46-lip-sync-from-the-songs-own-vocal)) the check called the first take's soundtrack re-sung rather than copied,
matched only from 7 s on, yet its singing sat exactly where the vocal is, and the director called the take perfect. For
that method, the review video is the verdict: the take with the song laid over, by eye.
**Or align the take, and follow the path.** FABLE aligned each take's own soundtrack to its window by subsequence DTW on
log-mel frames, phrase by phrase, and the cut followed the path: a clean copy costs 0.05–0.2, a re-sung line 0.2–0.4
([tool 81](#81-sync-as-a-path-each-take-aligned-to-its-window)).

### 16. Covers and coverage
**The earlier method.** Takes from the song's own vocal stem needed no covers: a take that fails is made again from
better inputs ([tool 46](#46-lip-sync-from-the-songs-own-vocal)), one take per job.
**Job.** Keep the in-sync part of every take, cover the rest, and know how much of the singing is really synced.
**The cut.** Cut on the last eighth note at or before `synced_until`. Eighths are the beat map's beats plus the midpoints
between them. A cut on an eighth feels like part of the music; a cut at the divergence frame looks like a glitch.
**A cover** is a new clip:
- **Window:** it starts in the word gap before the cut, and lasts `clamp(ceil(end − start + 0.3), 4, 8)` whole seconds.
- **Start image:** the parent take's frame at the cut (clip time = cut − `song_t0` + lag), unaltered, sent as the image
reference, never as a first frame.
- **Words and audio:** exactly the words that start in its window, with their clean vocal.
- **Prompt:** the parent's prompt, re-quoted to the new words.
- **Records:** the parent lists its covers; the cover records its parent and why.
- **Limits:** at most two covers per root, and one on a provider where copies are rare.
**Coverage**, per chain (a root and its covers):
- sung stretches = each aligned word ± 30 ms, merged when closer than 0.12 s;
- in-sync spans = every element's spans, each start extended by 60 ms;
- gaps = sung time not covered by any span, of 0.6 s or more, merged when closer than 0.3 s.
The bar is "most lips", like a pop video ("like grimes vids … shedoesnt sync lips on everything, just most"). Target at
least 80% of sung seconds, never an out-of-sync mouth held on screen, and budget covers at about 60% of the sung
seconds.
**Measured:**
| Run | Result |
|---|---|
| ESCAPE VELOCITY | 41% of sung takes copied; with covers, about 86% of sung seconds verified |
| CLODYSSEY, Higgsfield | about 21% of sung takes copied |
| BLISS | 62% of six windows' sung seconds |
| A later test run, seven windows | long prompts held 44% of the chorus's sung seconds; short ones 62–71% |
**When a window won't copy**, try these in order:
1. **The short sung form,** for a clean window. The biggest lever measured there is prompt length. Long timestamped
screenplays make the model perform the text and re-sing it; short prompts copy. (Takes from the song's own vocal stem
copy with full prompts of about 200–400 words: [tool 46](#46-lip-sync-from-the-songs-own-vocal).) The short form is:
- the head, quoting every word;
- one line of framing and light;
- at most two timed beats;
- "One continuous shot, no cuts, no head turns. Only she sings. No text, letters or logos.";
- about 90 words or fewer.
On seven windows the mean in-sync share per take went from 26% to 58%; one window went from 0% to 78%. Screenplays
stay the rule for acting and motion clips.
2. **Clear the mouth.** Nothing in front of it on the plate: no held microphone, hand or hair.
3. **The other provider,** after two non-copies in the short form.
4. **Cut away:** to the shot's motion clip, stills in rhythm, depth moves, crops to the eyes, hair or hands, inserts, or
a wider or profile framing. The words stay on screen either way.
**No lip-sync models.** Sync, LatentSync and Wav2Lip were tried as fallbacks and dropped: "still not exact", and the
mouths came out forced and glitchy. The failure was the input, not the model.
## The edit and the renderer
### 17. The edit list
**Job.** Resolve every moment of every shot to a source.
**How.** Resolve a sung shot moment by moment in song time. At each time T, use the chain element whose in-sync span
contains T, at clip time `T − song_t0 + lag`. A take and its covers then interleave, each switch sitting on the eighth
before a divergence. Fill holes from the shot's motion clip, then from the plate. Motion clips play straight through,
holding their last frame if the shot is longer.
**Done when** the film renders end to end at every stage, with plates standing in for clips that don't exist yet.
### 18. The renderer
**Job.** Draw every frame: footage, 2.5D, lyrics, explaining graphics, chrome and transitions, on the song's clock.
**The frame contract.** It renders offline in headless Chromium, one frame per 1/24 s, in parallel and out of order.
- **Every frame is a pure function of song time t.** No state between frames, no `Math.random` (use a hash of the
index), no `Date`.
- **Two canvases.** The base holds footage and anything graded with it. The overlay holds type and chrome, composited
crisp on top.
- **Memos only for data derived from static media,** keyed by media name, never by frame.
- **Every 2D canvas is software-backed** (`willReadFrequently`), and every scene starts from a reset context. A
GPU-backed canvas under software GL kept the previous frame's overlay once frames got dense; that took about two hours
of bisecting to find.
- **Quantise to frames.** A cut shows on the first frame at or after its time. Video may run up to one frame ahead of
the sound, never behind it.
- **Speed:** at most 1,500 ms a frame. The 64-step parallax shader costs 0.5–1.1 s, and a full-resolution canvas blur
about 0.5 s. Use depth slices (about 30 ms) and blur by resampling instead.
**The clock.** All musical time goes through the beat map: beat position, bar, local beat length, kick pulses and
eighths.
- A lyric token appears at its aligned start exactly, never snapped to a beat: "the motion syncs to the beat, the word to
the voice".
- An entrance takes at most an eighth of a beat.
- Every timed thing carries a timing key:
- **W** a word's time;
- **B** the beat grid;
- **H** a track event (kick, snare, impact, riser, fill, envelope);
- **F** a fixed song time;
- **T** a keyframe measured on a take's own frames.
**Lyric modes.** Every sung word appears in at least one, and every drawn token is marked for the lyric gate.
| Mode | For |
|---|---|
| subtitle | sung lines; the sung word lit in the accent |
| coverline | spoken lines set beside her in the plate's empty field, with tracking collapsing word by word |
| masthead | payoff words, numerals, the answer in a call and response |
| stack | short emphatic lines, one word per row |
| mass | a declared wall of micro text in a drop, with the sung line still drawn |
| title | the title knocked out of black |
Type rules:
- One lyric block per frame; a line keeps one mode and one position across cuts.
- At least 40 px, and never over her face.
- Margins are left 72, top 56, right 96 and bottom 84 at 1920 × 1080.
- Hang type from the left axis or from a feature in the frame.
- Nothing may cover the lyric: the renderer measures it every frame and moves the other layers off it.
**The camera plays the track:**
- **punch** on kicks: amp 0.01–0.03, gated by the drum envelope;
- **shake** with the drums: none in screen or UI shots;
- **ride** a riser into the hit it lands on;
- **impact** on real hits: strength at least 0.6, zoom kick 0.06;
- **whip** on fills.
Hold still where the song breathes. A move never cuts the face.
**Density** is chosen per film:
- **dense:** three or more layers, five or six as the norm;
- **budgeted:** footage, one lyric block, one hero device, at most two detection marks, the chrome, and the accent on at
most two elements;
- **sparse:** about 70% of the film with no graphic at all.
**Chapters.** In the Full scale each chapter is its own code file, written by an animator agent against this contract.
Chapters draw only what is unique; shared helpers live in one library. Workers report shared-file bugs; the operator
fixes them once.
**Rendering.** The render is resumable: skip frames that exist, write each frame to a temp file and rename it, open a
fresh page every 200 frames, and retry a failed frame once. ESCAPE VELOCITY's 7,354 frames took about 63 minutes with
three workers on 4 CPUs; at 1080p one cut's frames take about 9.6 GB.
## Checks
### 19. The lyric gate
For every sung token, render the frame at `t0 + clamp((t1 − t0)/2, 0.03, 0.12)` and check that the token was marked as
drawn on it. It prints N/N; anything less fails. ESCAPE VELOCITY: 417/417. CLODYSSEY: 315/315.
- Run it per chapter while animating, and for the whole film before any render.
- It proves presence and timing, not legibility; contact sheets check legibility.
- Fix a mistimed word in the alignment, never in chapter code.
- Match drawn tokens to sung tokens by their time, not their text, so a surface that drops the "…" or the "?" (the real
anti-piracy card has neither) still passes. AIRPLANE MODE: 363/363.
- The sheet decides what counts. An ad-lib the director doesn't want on screen stays off it: AIRPLANE MODE's outro
repeats "bing-bong" many times and shows it once.
- A stutter the take sings is a token. DOWNFORCE's countdown numerals slam on their measured vocal onsets, and the
edit cuts on every sung "one" and every "done": 190/190.
- Words on a real interface count when they are on it. DON'T LET ME SCROLL's lines live on X posts, Screen Time's limit
sheet, LinkedIn, the Dynamic Island and word pills: 148/148 on every version.
- A coded interface can report its own words. AERO's engine logs every string it draws, with its opacity; the gate renders
the frame at each word's onset plus up to 0.12 s and looks for the word in a string drawn at least 35 % opaque. On v2 it
found 275 of 280 sung words; the five it could not see were the password hint's three words, typed as she says them, and
two single-letter "I"s. A miss there is a lead, not a verdict: the log can't see a word inside a picture, or one the
camera has cut off.
### 20. The history check
Render frame B fresh, then render it again right after frame A in the same page. It fails if any pixel differs by more
than 40 levels.
- Run three pairs per chapter, one across each seam, with one pair whose A is a heavy 2.5D frame.
- It must print HISTORY-INDEPENDENT.
- Footage-heavy chapters read the frame as drawn (an edge, a key, where a label goes); that stays pure as long as it reads
only this frame's pixels. DOWNFORCE's final check rendered sixteen pairs in different orders, and all sixteen matched to
the pixel.
- A full-film worker crosses chapters, so one canvas made without the software flag makes every later chapter's frames
depend on what rendered before. DON'T LET ME SCROLL made every 2D canvas in the film software-backed after one chapter's
canvas was promoted to the GPU and later frames drew their edges differently.
- The history check passes a worker whose own state went stale; [tool 43](#43-the-single-frame-outlier-scan) finds it.
- Three orders find more than one. AERO's engine renders a span in order, shuffled and on a cold page in reverse, and
compares the bytes. It caught three traps that a single order hides: Chromium decoded a JPEG drawn small at a reduced
scale only when no full decode was cached, so a small picture depended on render order (every picture now goes through a
1:1 software copy first); caches keyed by a scale under an eased camera made a bitmap depend on which frame drew it first
(key caches by content); and drawing a WebGL canvas onto the output switched Chromium's raster path, after which every
later frame differed (read WebGL frames back into a software canvas).
### 21. The flash check
This follows WCAG 2.3.1's general flash threshold.
- Read the video at 96 × 54, 24 fps, in linear relative luminance.
- A flash is a pair of opposing changes of at least 10% of the range, where the darker state is under 0.8, over at least
a quarter of the frame.
- Fail any one-second window with more than three flashes. Count pairs, not transitions: an early version counted each
transition, so 2.2 flashes a second read as 4.4.
- Flashes belong to a theme, on hits: "dont need to force flashing".
- **Find single-frame pops.** A fast camera pass through bright glass is a flash. AIRPLANE MODE's dive through visor after
visor failed at four flashes in one second: as each lit rim rushed past the lens, the mean luminance jumped for one
frame and fell back (0.147 → 0.226 → 0.165). Flag any frame whose mean luminance differs from both neighbours in the
same direction by more than 0.03, and ramp such passes over three frames or more.
- **Run it per chapter,** on anything fast, not only on the master: that dive was found after the encode.
- **Report a failure** with its time and cause when the film is delivered. Shipping it anyway is the director's call,
never the production's.
- **Flashes as cuts only.** DOWNFORCE puts every white flash on the first frame of a cut and never inside a held shot,
and its board checker fails a flash that isn't on a cut. The film has 13 white flash frames and at most 2 flashes in
any second. DON'T LET ME SCROLL: 6 flashes, every one a cut, at most 1 in any second.
- **The flash check passes a strobe; scan for it cell by cell.** Cut the frame into a 16 × 9 grid (120-px cells at
1080p, read at 192 × 108) and take each cell's mean luma per frame. A strobe is a run of 4 or more frame-to-frame
changes that alternate in sign, each over 8 levels: a passing object gives a run of 2, aliasing rows give long runs.
DON'T LET ME SCROLL's tunnel passed the flash check while 29 of its 144 cells alternated light and dark for nine
frames: the near cards swapped bright and dark every frame, and only the largest were smeared. The near cards now
fade, every card smears, and the scroll is capped. After the fix, every frame the scan flags is a whip or a swipe.
### 22. Pace against the music
Picture motion is the mean absolute frame difference of the review copy at 160 × 90 grey, per section. Compare it with
each section's stem energy and its kick and hat onsets per bar. Busy picture over quiet music reads as "off pace". The
fix on one film halved the cuts (0.92 → 0.47 a bar) and slowed in-shot motion until the picture followed the music.
When a cut feels wrong and the totals don't say why, measure what the viewer sees between visible cuts, section by
section, and re-cut by what each shot holds ([tool 58](#58-the-pacing-analysis)).
### 23. Contact sheets and the reviewer
**Contact sheets** are how agents look.
- **Per chapter:** every shot's first, middle and last frame; every lyric line's first and last word; every hit; each
graphic's entrance; each transition's middle.
- **The reviewer** sees at least 24 frames, plus full-size stills of the five most important moments.
- **The whole film:** mid-shot frames of every shot, 6 × 24 to a sheet, plus the seams and 100% crops of faces.
- Open every sheet. A sheet nobody looked at is not a review.
**The reviewer** renders and looks for itself, and never edits files. It scores four lenses, 0–10:
- **taste:** the director's taste, which counts double;
- **craft;**
- **sync;**
- **story.**
It returns at most 12 concrete fixes, each marked blocking or polish, and copies the lyric gate's printed line.
**The verdict is computed in code, not taken from the reviewer:**
- A chapter passes when no blocking fix is left, taste is at least 7, the other three lenses are at least 6, and the gate
reads N/N.
- A review that returned nothing, or never looked at a frame, is a failed review, never a pass. The original scripts read
an empty review as "no fixes", and runs that hit a usage limit reported success.
Run at most two review-and-fix rounds per chapter. Calibrate the reviewer on the director's verdicts: on CLODYSSEY the
reviewers scored 6–7.5 a cut the director called "perfect. change nothing."
**The checker, after a round of notes.** When the director has sent notes, a lens score is the wrong question; the
question is whether each note was done. AIRPLANE MODE's last rounds gave every chapter a checker that hadn't built it:
- It reads the director's notes and the chapter's notes, including every UPDATE line.
- It renders sheets at every note's moment (before, at and after) and at every shot's first and last frame, plus two
to four full-size stills, and looks at all of them.
- It returns each note as done or not, with what it saw; everything broken (blank or black frames, a flash inside a
scene, type off its surface or illegible, an off-model face, a missing level tag, over-cutting where the notes ask to
linger, a cut that left the beat); and what is pending on an asset that hasn't landed, which is not a failure. The
format is in [Data formats](https://claudiagallery.vip/wiki/formats/#a-check-against-notes).
- It never edits files. A failed check sends its open notes and broken items back to the animator, first in its prompt,
for one refix.
**Two reviewers at once, then a verifier.** DOWNFORCE gave every chapter two reviewers that render their own sheets
(they don't trust the builder's) and run in parallel:
- an **art director:** every shot's middle and every cut's first frame, a strip at every eighth of a second through the
fastest three seconds, full-size stills of the six most important frames; energy, cuts and moves on the hits, the
type as composition, the picture right (no frame outside a take's window, no wrong car, her lips closed, no invented
text);
- a **QA checker:** the lyric gate, the history check, every generated frame the chapter draws against its take's
window, every number on screen against the readings, cut times against the hits (within one frame), scene coverage,
the flash rules, the first and last frames rendered without errors.
Each returns blockers (time, what, fix), ranked improvements, and what to keep. A fixer applies every blocker and the
improvements that don't conflict with the board or the director, a verifier re-renders and checks every first-round
blocker, and a second fix runs only if blockers remain. Five of DOWNFORCE's eleven chapters needed it.
**The whole-film critic.** After the chapter rounds, one agent renders the whole film and judges the film rather than a
chapter: the arc, repeats across chapters, the type system, the seams (every frame from 0.25 s before to 0.25 s after
each), and every one of the director's notes, re-rendered frame by frame over its range and looked at adversarially. Its
issues go to one fixer per chapter, then a final check runs the gates on the whole film. On DOWNFORCE it found a blocker
no chapter-level review had caught: a fixer had swapped an off-model close-up on another set into a held shot.
**Fresh eyes, one sheet per shot.** FABLE had no chapters and no lens scores: before each delivery two reviewers who had not
seen the film went through one sheet per shot, eight frames as the cut shows them, and full-size frames, and every finding
was fixed or listed for the director ([tool 88](#88-two-fresh-eyes-reviews)).
### 24. The safety and canon scan
**Job.** Look at every generated take before release.
**How.**
- Make sheets of frames per take.
- List every take checked, and every problem with its severity: block, fix or note.
- Whitelist intended props by name.
- Re-scan every re-make.
On CLODYSSEY the scan found 12 blocks in 96 "final" takes, among them nudity through a corset, modern objects and
earmuffs.
## Delivery
### 25. Encode and deliver
**The master.** Frames are sRGB, so convert through RGB with accurate rounding to BT.709 and tag the stream:
```sh
ffmpeg -framerate 24 -i frames/f%05d.jpg -i song.wav -map 0:v -map 1:a -c:a aac -b:a 256k -shortest \
-vf "format=rgb24,scale=flags=accurate_rnd+full_chroma_int+full_chroma_inp:out_color_matrix=bt709:out_range=tv,format=yuv420p,setparams=color_primaries=bt709:color_trc=iec61966-2-1:colorspace=bt709:range=tv" \
-colorspace bt709 -color_primaries bt709 -color_trc iec61966-2-1 -color_range tv \
-c:v libx264 -preset slow -crf 16 -pix_fmt yuv420p -movflags +faststart master.mp4
```
An untagged or BT.601 stream looks lighter and washed out in QuickTime ("stuff is v blown out").
- **The review copy:** the same, with `-crf 20 -maxrate 8M -bufsize 16M`.
- **Verify with ffprobe:**
- codec, 24/1, and a frame count equal to duration × 24;
- bt709 / iec61966-2-1 / tv;
- AAC stereo;
- no container tags beyond the encoder's own.
- **Then run the flash check on the master.**
**The 9:16 cut** follows a subject box per plate, so an off-centre lead stays in frame. It gets its own lyric gate.
FABLE's 9:16 edition is a page instead: a square of each shot round its subject on a cream page, with the sung line
printed under it ([tool 87](#87-the-annotated-edition-and-the-916-page)).
**Two pitfalls of the frame folder:**
- **A file's data must be what its name says.** AERO's first 1:20 linked `.png` names to frames that were JPEG data; ffmpeg's
image2 demuxer dropped every one of them, and the encode came out 1,325 frames long instead of 1,949. Convert to the
named format, and count the decoded frames against the sequence.
- **A picture that starts before the song** gets negative frame numbers (`f-0064` … `f-0001`, then `f00000` on the song's
first sample), so every frame keeps its song time and film frame k is song frame k − 64. Every tool that reads a frame
range must parse a leading minus, and a command line must not take it for an option (`--frames=-64-1884`).
**Housekeeping:**
- Run encodes as tracked background jobs, never a detached process nobody watches.
- Delete the frames only after ffprobe has confirmed the outputs.
- Keep big scratch on the big disk.
## Running agents
### 26. Workflows, run records and the stop switch
The films ran their heavy stages as multi-agent workflows (Claude Code):
- **storyboard:** three director agents, three judges, and a merge;
- **chapters:** an animator, an art director who renders and reviews contact sheets, and a fixer;
- **research and lyrics:** topic researchers, fact verifiers, writers and judges;
- **the safety scan.**
DOWNFORCE ran seven workflow runs and 172 agent launches: the storyboard (three director agents, three judges, a
head director), the engine kit (four builders, an integrator, three reviewers, a fixer), a plan for the second round of shots
(three scouts, a head director), the chapters in three parallel runs (per chapter: a builder, an art director and a QA
checker, a fixer, a verifier), and a whole-film pass (a critic, fixers, a final check).
DON'T LET ME SCROLL ran 32 agent launches and no judged workflows: two surveys of the library (the moodboards, the
earlier films' looks), five kit builders in parallel from one written contract, one animator per chapter with its own
art-director passes (15 launches for nine chapters, six of them relaunches after interruptions), and ten for the second
version. In its last rounds the director asked for fewer reviewers: "we dont need excessive 'reviewers' after, ill stay in
the loop and can do a visual review faster and more comprehensively than an opus reviewr can do it."
WHITEPILL had no chapter animators: its frames are model takes and shader renders cut on the bar map. Its agents wrote
rival shader worlds in rounds, each round kept beside the last as a fallback, and a critic agent watched the previz as
the director would, returning ranked notes in three rounds, then a final QA and a last check before it went out. It ran
31 agent launches: 3 researchers and the 7 they launched, 3 lyric writers, 13 shader builds, 5 critics and QA passes.
THE CAKE IS NOT A LIE's restart ran as teams by role, each given the brief, the story, the production spec and the
storyboard so that none lost the whole film: the song and glitch map, the recordings in Portal 2, the character sheets,
Midjourney (the test and the library), the overlay, her singing in the chamber, and one team per act; the lead kept the
story and the cut. Its later rounds of notes went to specialists, each with a written brief and the lead's review at full
size before the director saw anything: a HUD designer ([tool 55](#55-a-hud-from-one-visor-and-one-projector)), a colourist
([tool 57](#57-the-per-shot-grade)), a pacing analyst ([tool 58](#58-the-pacing-analysis)) and a motion designer
([tool 60](#60-a-memory-film-with-her-own-face)). It ran 24 agent launches in all.
AERO's agents are recorded by role only: researchers (the knots between the era and 2026, the picture, the music, the
interface's sounds), the engine's builders (its foundation, the apps, the full-screen screens), eleven chapter animators and
the fixers who carried the reviews' notes, a fresh art director for every chapter pass and one for the whole film, a sound
designer, a world builder in Blender, a takes agent (imports, lip checks, keys), a texture agent and a coordinator for the
last passes. The number of launches is not in its records. The animators worked to one written contract, which quotes the
director's line on lazy animations and asks for three full rounds of the animator's own checks before a chapter is handed in
([tool 66](#66-chapters-on-one-machine-to-one-contract)).
FABLE dropped the multi-agent shapes as patches for older models: one author wrote the plan, the board, every prompt and
the cut in one context. Its agents, by role: five researchers for the plan (the week's AI news, the culture, the series
and its field, the newest video and image tools, the sound), three painter agents for the plates ([tool
76](#76-set-plates-painted-in-svg-rebuilt-as-paper)), and two fresh-eyes reviewers before each delivery ([tool
88](#88-two-fresh-eyes-reviews)).
**Rules that cost time to learn:**
- **Save every run's id and arguments** the moment it starts, and resume by run id in the same session. Finished agents
replay from cache. An interrupt kills every running workflow, so keep phases short; ESCAPE VELOCITY's merge restarted
three times. DOWNFORCE's three chapter runs were interrupted three times and each time resumed from where they stopped:
only the agents that had been cut off ran again (57 of 147 launches), and no work was lost.
- **Build the kit before the chapters.** DOWNFORCE's four kit modules (compositor, footage player, lettering, data
graphics) were built in parallel by four agents, joined by an integrator, reviewed for taste, correctness and speed,
and closed out by a fixer who wrote the kit's manual, with one worked-example chapter that calls every function. Every
chapter agent read the manual first and never edited the kit; kit bugs were reported, not patched.
- **Send the rough cut before the polish.** DOWNFORCE's first cut went out with every chapter at its first build, before
any review, rendered from a frozen copy of the chapter files because the reviewers were already editing them. The
director's notes reached the agents while the review rounds were still running.
- **A notes file survives an interrupt.** Each of DON'T LET ME SCROLL's chapter animators kept a notes file as it worked;
six launches were interrupted, and each relaunch picked up from the file where the last had stopped.
- **Check that every agent actually finished.** A run can report success while agents that hit a usage limit returned
nothing. Carry open items forward explicitly.
- **Parallelism has a ceiling.** On 4 CPUs, run two agents a workflow. Allow one rendering workflow at a time: six
rendering agents filled the swap and the root disk.
- **Put long lists in a file,** not in the launch arguments: a placeholder once launched in place of 96 paths.
- **One workflow per chapter.** Resuming a multi-chapter workflow replayed its stages unevenly: some finished chapters
were reviewed and fixed again. One chapter per workflow keeps an interrupt to one chapter.
- **Notes live in files the agents re-read.** Each chapter has a notes file for the round; the lead appends dated UPDATE
lines as assets land (a take registered, a shared tool finished, a hand-off changed), and every agent re-reads its file
before it finishes. Format: [Data formats](https://claudiagallery.vip/wiki/formats/#a-directors-notes-on-a-cut).
- **Build around what hasn't landed.** A chapter lays out a pending take or tool at its planned timing with a stand-in,
keeps the name of what it waits for in one place, and swaps it in when it lands. Nothing waits for everything.
- **Write the seams down.** Each chapter's notes state its hand-offs: the exact colour of its last frame, where the level
tag sits on both sides of the cut, which function returns a box the next chapter needs. Use the returned box, never a
copied rectangle.
- **The stop switch.** On any quality complaint from the director:
1. stop every job and workflow that can spend, after searching each running script for its generate calls;
2. confirm at each provider that nothing is queued;
3. log the director's words as a rule;
4. fix the prompts, then re-make. Don't review the bad takes.
- **Write the acceptance test before the brief,** and make the check something the worker can't edit. Every human
observation becomes a metric the same day: "if a reviewer can see it and the judge cannot, the judge is wrong".
## Worlds inside the film
AIRPLANE MODE, the fourth film, added these: tools for a story with simulations inside it, and for graphics it had to
rebuild rather than imitate. Build them only if your story needs them.
### 27. The level table and the level tag
**Job.** Show on every frame which level of the story the viewer is in, from one table.
**In → out.** The levels table ([Data formats](https://claudiagallery.vip/wiki/formats/#the-levels-table)) and song time t → the tag on the
overlay, or nothing in the real world.
**How**, with the constants used:
- **Placement:** drawn last, in a corner, 64 px from the side and 54 px from the top or bottom at 1920 × 1080; the
number 28 px tall, the caption half that. Where an interface fills the frame, move the tag clear of its title bar.
Change the corner only on a cut, and match it across chapter seams: the same corner, margins and backing on both sides.
- **Two styles:** for the realistic levels, a mono tag: a 1-px white frame with corner ticks in the accent over a faint
navy backing, SIM LEVEL in white, the number in the accent. For the game levels, the game's pixel font on its
640 × 360 grid, scaled 3× with nearest-neighbour.
- **Motion on the beat grid:** captions type at one or two characters a frame; a number that lands and holds rolls its
changed digits over three frames, eased; a racing counter updates in place on the sixteenths, exponentially (2 to
1,048,576 in about five seconds).
- **States:** BOOT (the frame draws on in four frames, then one caption row per beat), SPAWN (SPAWNING SUBAGENT… typed,
a bar filling on the sixteenths to the drop), CLIMB (the race), GLITCHOUT (the tag torn into nine slices that slide
apart, the colour split in position only, the backing kept dark, then nothing).
- **Hidden** while a full-frame card interrupts the simulation.
- **Never a flash:** the tag ramps, types, rolls, or cuts with the shot.
**Done when** every frame's tag matches the table, the real world shows none, and the tag never sits over a lyric or an
interface's title.
### 28. Virtual her: a cut-out in a world rebuilt from its edges
**Job.** Make a realistic take of her read as a simulation without touching her face.
**In → out.** A take or a plate, and its per-frame matte → the frame: her, photoreal, inside the matte; the world outside
it redrawn as cyberspace.
**How**, with the constants used:
1. **The matte.** SAM 2 with a box prompt, then BiRefNet (HR matting) and ViTMatte for the edge. A second box for a
headset or a hand the first one loses; close the seam between the two with a small morphological closing. Grey
mattes at 960 × 540, numbered like the take's frames.
2. **Her cut-out.** Decontaminate the matte's partial pixels: solve I = aF + (1 − a)B for F, with B the local
background colour, so no halo of the old window stays in her hair.
3. **The world from its own edges.** The frame's luma at 960 px wide, blurred four times with a [1 2 1] kernel, then
Sobel, non-maximum suppression and two thresholds (4.2 and 19 luma levels per pixel at that width) into anti-aliased
lines from deep blue through cyan to near-white, with a blurred glow, over navy-black. Suppress the edges on and just
around her (the matte, dilated).
4. **Depth.** The frame's own light as a dim blue ghost (luma^1.9); a wireframe corridor (floor, ceiling and walls)
receding to the shot's real vanishing point, its rungs flowing toward the camera; short columns of hex glyphs rising
slowly, a few in the accent; about 90 particles.
5. **On her.** A fine cyan rim from her own alpha, a² × (1 − box3(a)), 1–2 px wide; a soft blue halo outside her; light
wrap from the blurred cyberspace behind; 4-px scan lines and one slow scan band every 3.2 s.
6. **The boot.** An amount from 0 to 1 converts the real frame to cyberspace tile by tile (32-px tiles), outward from
her head.
7. **Cost.** Cache the analysis per source frame and matte (four entries are enough): about 0.3–0.4 s a frame when the
source changes every frame, 0.14–0.2 s on a still plate.
**Done when** her face, hair and hands are untouched at 100%, nothing of the old background clings to her edges, and a
contact sheet of the shot reads as virtual at thumbnail size.
### 29. Character replacement from a motion reference
**Job.** Her avatar performing a move animated in 3D, in the film's look, without a finished 3D model of her.
**In → out.** A shot's clay render, the avatar's front view and turnaround, a still of the avatar in the shot's world →
a silent 720p take, then the shot's frames for the game look.
**How:**
- **The clay render:** a grey figure in a grey world with the shot's real lens and camera move, the sound stripped,
4–15 s (the shot's length rounded up to whole seconds). Keep a hash of it in the clip file, so a re-render is noticed.
- **The world still:** one per shot, made in Midjourney with a frame of the shot's current version as the style
reference and the avatar's front view as the image reference.
- **The prompt:** the template in [Seedance](https://claudiagallery.vip/wiki/seedance/#replacing-a-stand-in-figure). Per shot, only the goal
sentence, what the world still shows, the timeline and the props to leave out change.
- **Post:** frames at 24 fps, area-downscaled to 854 × 480; a light unsharp mask (radius 1.2, 80%), then
nearest-neighbour to 640 × 360, shown at 3×.
- **Screen every take** before it replaces a shot, and re-measure everything a chapter measured on the old frames
(corners, masks, anchors), or write it so it doesn't depend on the take.
**Done when** every frame shows the avatar on-model (adult proportions, one streak, the headset, the suit), the motion
and camera follow the clay render, and no stand-in figure or retired prop survives.
### 30. A real referent, rebuilt
**Job.** When a line quotes a famous graphic or a real interface, put the real one on screen, rebuilt in code.
**How:**
1. **Research first.** An agent collects real frames or photographs, writes down what each one shows and what is still
unverified, and lists where your current version departs from the real thing.
2. **Measure** positions and sizes as fractions of the frame, the palette sampled from the frames, the motion frame by
frame (the scale on each frame, the holds, the cuts) and the typeface.
3. **Clear the typeface.** Use the original only with a licence, and never a pirated clone. Otherwise take an OFL face,
squeeze each letter to the original's measured ink width, and draw any wear in code, fixed from frame to frame.
4. **Write it as a function of song time,** with the film's timings as presets: words land on their sung times, punches
on hits.
5. **Replace the original's flashes with motion:** no white frames, no brightness pops.
6. **Keep it whole:** full frame, one card per line, never split, tilted or shrunk into a scene. Put the film's own mark
where a logo isn't yours to use.
**Done when** the same code rebuilds one of the reference frames and the two match side by side.
## Real footage
DOWNFORCE, the fifth film, is cut from real footage: Tesla's livestream of the Roadster demo and SpaceX's 2018 Starman
footage, with generated 3D moves that start on its exact frames and Claudia as the pilot. These tools read the footage,
lock graphics to it, keep the generated takes honest and board an edit around it. Build them only if your film is cut
from real footage.
### 31. Reading a broadcast's telemetry
**Job.** Turn the numbers a broadcast shows on screen into data, so every graphic and every number in the lyric is a
real reading, never an estimate.
**In → out.** The broadcast at its highest resolution → one series per gauge, `[source time, value, interpolated]` at 20
readings a second, cut into the stretches where the readout is on screen ([Data formats](https://claudiagallery.vip/wiki/formats/#readings-off-the-footage)).
**How**, with the constants used on Tesla's 4K feed (3840 × 2160; its readout in a band along the bottom: DOWNFORCE in
kg, later G-FORCE in g, on the left; PITCH in degrees in the middle; ALTITUDE in feet on the right):
1. **Boxes.** Measure, once, a box around each gauge's value and one around its label, in source pixels. The digits
there are white on black, about 64 px tall, in a square-cornered face with a slashed zero.
2. **Templates.** Read a few frames by eye (five were enough), cut their value boxes into glyphs, and store each glyph,
resized to 28 × 44 px, under the digit it shows.
3. **Glyphs.** Per frame, threshold the value box at 150 of 255 and split it into 8-connected components. Drop specks
(under 25 px) and anything over 90 px in either direction (the gauge's arcs, light streaks). Keep the text line:
components at least 40 px tall whose centres sit within 20 px of the median. A small component (under 20 px both
ways) on the baseline (within 10 px) is the decimal point.
4. **Digits.** Resize each glyph to 28 × 44 and compare it with every template by normalised correlation, both
mean-subtracted. The best template names the digit; the reading's confidence is its weakest digit's score.
5. **Which gauge.** Where a gauge changes what it measures, count the ink in its label box: DOWNFORCE is wider than
G-FORCE.
6. **Clean, never guess.** Drop readings under 0.85 confidence, and any reading off the median of its five neighbours
by more than max(25% of that median, 3 × their median absolute deviation + 0.6). Fill the gaps linearly and flag
every filled value. A value crossed by a light streak comes out flagged, not invented.
Measured on DOWNFORCE: 1,191 instants in three stretches (the wall, the flight, the kickflip), 3,081 values, 28 of them
flagged and interpolated. The peaks: 4,709 kg, 170.0° and 69.1 ft on the wall; 0.4 g at 68.0 ft leaving the ramp; 3.3 g
on the touchdown; 3.6 g at the kickflip's launch, with a 26.2 ft apex.
**Use.**
- A live card shows only values the readout showed at that source time; a number that counts up counts through real
readings. A gauge's peak hold takes the highest reading between two source times (a beat ago and now).
- Write the lyric off the readings: DOWNFORCE's "up the wall at ninety" is 89.7°, "Sixty-nine feet" 69.1 ft.
- Turn type and pictures by the measured angles ([tool 33](#33-graphics-locked-to-the-footage)).
**Done when** crops of the readout beside their readings agree by eye at the peaks and across every flagged stretch,
and every number the film shows can be looked up at its source time.
### 32. Tracking the subject in the footage
**Job.** The real car's box in every real shot, so graphics can ride it: a callout, its dotted path, a punch-in that
keeps it in frame, words that pass behind it.
**In → out.** The real clips as frames, each clip's cut list, a few seed points per shot → per clip, 12 samples a
second, `[u, v, w, h, conf]` in fractions of the image, or null where the car can't be seen (smoke, flare, out of
frame).
**How.** Mask the broadcast's readout and titles first, and work at 1440–1920 px wide (half of the 4K frames, the 1080p
ones as they are). Each shot gets the method that suits it:
| Method | When | How |
|---|---|---|
| empty set | a locked-off or slowly zooming camera | the median of the shot's frames is the set without the car (align a slow zoom first: affine ECC, chained out from the middle frame); the car is what differs from it by more than about 26 levels, compared with the local max and min of the empty set so a pixel of misregistration doesn't count; on a white wall a blob must also carry a dark core (glass, shadow), which light pools and flares don't |
| bright | the car against the dark | bright, unsaturated pixels (HSV value over 150, saturation under 90), closed into blobs, only near the seeds |
| manual | the car fills the frame | key boxes read off the frames, interpolated (confidence 0.6) |
| from | another clip carries the same feed | that clip's track, shifted in time and mapped in space |
- **Candidates:** every blob near the seed path, its holes filled (the dark glass inside the white body), thin rows and
columns trimmed off (thruster jets, smoke wisps: keep the longest run of rows, then of columns, covering at least a
quarter to a third of the best one). Each costs its distance from the seed, relative to the seed's size, plus half
the log of its area against the seed's. Keep the best 12 per sample.
- **The path:** the cheapest path through each sample's candidates or null (Viterbi). A step costs its motion divided by
(5% of the frame width plus half the car's own size, so a close car may move fast), plus 0.8 × |log of the area
ratio|; a step longer than 4 is impossible. A null sample costs 1.2 and switching between null and a box 0.4, so smoke,
flares and lost frames stay null instead of grabbing the wrong blob. Confidence is exp(−cost / 2).
- **Smooth** with a 3-tap median of the box edges.
- **Seeds** come from looking: a sheet of the shot's frames with their times, and the car's centre and size read off a
grid where the track needs a start or a hint.
DOWNFORCE's PILOT ✶ CLAUDIA callout rides the box. A punch-in keeps the car's actual box inside the frame, inset 8%,
panning, and backing the zoom off where the picture's edge stops the pan. The kickflip's tracked path is dotted over the
lineup camera (valid only inside one static shot). A luminance key of the white car inside its box lets a lyric pass
behind it. A second tool paired the car's position on the wall with Tesla's altitude and pitch at the same instant.
**Done when** sheets of every tracked shot with the boxes drawn on show the box on the car, and null where it can't be
seen.
### 33. Graphics locked to the footage
**Job.** Graphics that stay on the real picture while the real camera moves: a line on a real edge, type turned by a
real angle, a label that finds the dark.
**How:**
- **Read the frame as drawn.** Draw the footage first, then measure what the graphic needs from those pixels, every
frame. It stays a pure function of song time, because it reads only this frame.
- **A line on a real edge.** DOWNFORCE's opening hairline sits on the wall's real edge. Per row (every 6 px) of a region
that holds the edge, take the first x where the next 24 px are at least 70% brighter than a threshold (0.25 of white;
prefix sums, so a dark patch in the concrete or a thin light beam doesn't move it). Drop points more than 48 px from
both neighbours, smooth with a 7-row running median and a 5-row mean, and draw the line beside the edge. The first cut
drew the line once while Tesla's camera zoomed in. The director: "at like 4 seconds the ramp trace decouples from the
ramp." Re-read from every frame, the line held; it was checked frame by frame over the zoom, at full size.
- **Type turned by the measured angle.** A line of lyric pivots with the car's pitch, read at the source time on
screen: "Sixty-nine feet," hangs at the real 170°, and "and I'm not done" turns live from 104° to 166°.
- **The world turned instead of the car.** Turn the whole broadcast frame by the car's measured roll, and the car stays
level while the set wheels round it (DOWNFORCE's chorus 2: 0 → −283°). Mirror the edges the turn uncovers, or go
black.
- **Labels in the dark.** A callout chooses its label point in the darker part of the picture, never crossing the car's
box or the lyric's boxes with its rows or its leader.
- **Keep the broadcast's own graphics out.** Frame above the readout, or slide the picture down with black above a
feathered edge where the sky is black anyway.
**Done when** the graphic is checked on every frame of a moving shot at full size, not on a sheet.
### 34. Generated moves from real frames
**Job.** 3D camera moves that start on (and, where needed, end on) a real frame, so the generated shot cuts with the
real ones.
**In → out.** A real frame (for a repair, also real frames of the subject's other sides; for a keyframe move, the frame
it must end on) → a start frame and references → a take.
**How:**
- **The start frame:** a 16:9 window at a chosen centre and zoom, above the broadcast's readout, scaled to a
1920 × 1080 PNG, with the broadcast's titles (rows of small white glyphs in the top 17% of the frame) found and
inpainted. A video input for an edit gets the same crop and the same inpainting on every frame.
- **The prompt keeps the frame and moves only the camera,** and names nothing: the car is "a white electric sports car".
The preambles are in [Seedance](https://claudiagallery.vip/wiki/seedance/#moves-on-real-footage).
- **Stay on the visible side,** or give the hidden sides as real reference frames, cropped tight on black, never
generated ones.
- **Chain on real frames.** DOWNFORCE's real side view ends on a frame, a first-and-last-frame move runs from that
frame to Tesla's real top-down, and the next shot starts on the same top-down.
- **Know the source's real picture rate before planning slow motion.** Tesla's "60 fps" feed carries about 24 distinct
pictures a second: at 0.4–0.5× it steps at 10–12 pictures a second. Blend neighbouring pictures, and motion-compensate
where both the car and the camera move.
- **Punch in on the sharpest source.** The 4K feed crops 2× to 1080p without loss; a 1080p source past 1.1× goes soft
and blocky.
**Done when** the cut into and out of each move reads as one camera, and the subject inside the take's window is the
real one in every frame.
### 35. The usable window of a take
**Job.** Know, for every generated take, which seconds may be on screen, and prove that nothing else ever is.
**In → out.** Each take → a row on a contact sheet, a list of cuts, its black level, then a verdict and window in the
takes log ([Data formats](https://claudiagallery.vip/wiki/formats/#takes-and-their-windows)); at the end, a log of every frame drawn.
**How:**
1. **The sheet:** one row per take, six frames spread evenly (384 px wide) with their take times, and the black level,
the 1st percentile of luma over those frames. Tesla's night footage sat at about 1, Wan's takes of her at about 14.
2. **The cut finder:** every frame at 96 × 54 grey and its mean absolute difference from the frame before; a cut is a
frame whose difference passes max(12 levels, 5 × the take's median difference) and is the largest within three frames
either side. It catches the model's own edits, not a slow morph (the car becoming another car, a set drifting): look
for those.
3. **The verdict, by eye, against the board:** USE, PART (only part of it) or ALT (a usable alternate), and the window
in seconds of the take, ending before the first wrong frame (a cut, another car, an invented set, lettering, an open
mouth). Write down what goes wrong after the window, and when, so nobody stretches it.
4. **Extract** the take at its own frame rate with its blacks restored: luma [B, 255] mapped to [0, 255], with B the
sheet's black level.
5. **A second, adversarial look** before more money is spent. DOWNFORCE's repair scout re-checked every take frame by
frame, shortened seven windows, and found a man in the empty car's window and a repair that turned into a hatchback.
6. **Enforce it in code.** The chapters name, for every shot, the part of the take they play; the final check logs every
footage frame the film draws and compares it with its take's window.
Measured on DOWNFORCE: 79 takes. The windows of the 25 Wan moves on Tesla's real frames all start at 0 s and end between
0.9 s and the full take (median 2.2 s). Every generated frame the film draws was inside its take's window.
**Done when** every take has a verdict and a window, and the log shows no frame outside one.
### 36. The board for an edit of real footage
**Job.** A shot list that tiles the song with real windows, generated moves and code, checked in code before anything is
spent.
**How.** A workflow:
1. **Three director agents** each board the whole song from a different angle, with every field of a shot and the list of
generated assets with their costs. DOWNFORCE's: a TikTok car-edit editor (every stunt from several angles, speed
ramps, the 3D moves as the money shots), a film director (the launch as a story: restraint, the reveal, each drop a
set piece) and a motion-design lead (the readings, the lettering and the lines drive the edit).
2. **Three judges** score every board 0–10 on seven fields (the build and the launch, hype, 3D, taste, lyric coverage,
feasibility, her) with a total out of 100, and list each board's eight best ideas and eight most important fixes.
One judges as the audience and editor, one as the director's taste, one as a producer: do the named source times
exist and show that, does every generated asset fit the budget, does every cut sit on the beat map.
3. **A head director agent** starts from the best-scoring board, grafts every listed idea that fits, applies every fix, writes
the board, then writes a checker and runs it until it passes.
The checker's gates, beyond [the audit](https://claudiagallery.vip/wiki/storyboard/#the-audit):
| Gate | Fails when |
|---|---|
| windows | a real window outside its clip, out ≤ in, a speed outside 0.25–4× either way (negative plays it backwards), out − in ≠ duration × that speed, a crop outside the frame or into the broadcast's readout, a window across a cut in the source (cuts measured from frame differences), a window where a broadcast title sits over the subject (titles over black sky are flagged for masking) |
| lyrics | a sung line not lettered (micro type doesn't count) in the shots that overlap it, at least 80% of its words |
| flashes | a flash not on a cut, or more than three in a second |
| generated | an asset referenced but missing, or made but unused; a window past its take; the first roll over its budget at list price; a plate without the canon text verbatim or without "no text, no letters, no logos" |
It also reports her share (her image and the callout, separately) and the split of real, generated and code.
DOWNFORCE's TikTok board won (83, 80 and 83 from the three judges; the film-director agent's board 79, 79, 85; the motion board
75, 77, 81). The merge had 182 shots, a mean of 1.04 s: 59% real footage, 37% generated, 4% pure code. From the start of
the workflow to a passing checker took about two hours. A second workflow later planned the next round of shots: three
scouts (more 3D angles at the drops; her as the pilot reacting to the real stunts; an adversarial repair scout) and a
head director, who merged them into one ranked, costed list with every start frame extracted and checked.
**Done when** the checker exits 0, and again after every change to the board.
## Her takes in a world drawn in code
DON'T LET ME SCROLL, the sixth film, put her into a world the engine draws: her takes, shot on plain grey, are cut out
and played inside the code, timed to the beat. Then the film kept changing after the edit: takes were re-dressed in new
outfits, and finished shots were re-lit and repaired and spliced back into the cut's own frames. These tools do that.
Build them if her takes live in a world drawn in code, or if finished shots must change without a re-shoot. The outlier
scan belongs to any film that renders on several workers.
### 37. Outfit-swap alignment
**Job.** Put an outfit swap (a video edit that re-dresses an existing take) back on the original take's timeline, so it
drops in as the take's new pick with every beat warp and chapter timing unchanged.
**In → out.** The original take and the edited clip → a new take file at the original's frame rate and frame count, and
the fit with its check, printed.
**How:**
1. **Know what the edit returns.** A Seedance 2.5 video edit returns the whole motion, uniformly rescaled to its own
quantised length, at 24 fps. A 5.0 s Wan take at 1080p (150 frames at 30 fps) came back as 113 frames, 4.709 s: the
same motion from the first frame to the last at 0.942 of the length, with no head trim and no offset. A 720p Seedance
take (121 frames at 24 fps) came back as 121 frames.
2. **Map it back by the ratio of the lengths.** Original frame i shows edited frame round(a · i / f_o · f_s), with a the
edit's duration over the original's and f_o, f_s the two frame rates; nearest frame, no blending. Write the frames at
the original's size and frame rate (CRF 12) as the take's next number, and make it the pick.
3. **Check the fit on what the edit doesn't change.** Her hair: the glossy black (grey level under 45 of 255) in the top
40 % of the frame, its centroid's x and y per frame, and her body's x (pixels more than 40 levels from the backdrop,
the border's median colour). Read both clips at 135 × 240, normalise each signal and correlate them under the map. A
finer search (the scale within ± 0.04 in steps of 0.005, an offset within ± 0.25 s in steps of 1/96 s) overrides the
ratio only when it is clearly better (Δr ≥ 0.03). On DON'T LET ME SCROLL's eight shirt swaps the ratio held, with
correlations from 0.71 to 0.998.
4. **Never fit the outline.** The first swap was fitted on her silhouette. The new skirt changed her outline, and the fit
read 0.36–0.5 s missing from the start and the rest running 4–6 % fast (a scale of 1.060 and a shift of 0.5 s); the
shot went in with a frozen half-second until it was aligned again.
**Done when** the swapped take plays inside the original's beat warps with the same pose on the same beat: the two takes
side by side at several moments match, and the hair's correlation is printed with the fit.
### 38. The re-light splice
**Job.** Re-light, re-texture or repair a finished shot with a video model, and put the result back into the cut over
exactly the frames it was made from.
**In → out.** A version of the rendered film (its frames on the render box) and a window in song time → a source clip and
its record; later, the model's clip → the window's frames replaced, the originals kept.
**How:**
1. **Cut the source from the final render's own frames,** never from an earlier stage, so it can't drift from the cut:
frames round(t0 · 24) to round(t1 · 24) − 1 at 24 fps (CRF 10), with an optional handle on each side that is never
written back. Record the window: `{"ver", "t0", "t1", "i0", "n", "pad", "frames"}`
([Data formats](https://claudiagallery.vip/wiki/formats/#takes-in-a-world-drawn-in-code)).
2. **Round the source up to 4n + 1 frames,** taking the extra frames from the real frames after the window, and to at
least 89. Seedance returned a 24 fps source of 4n + 1 frames frame for frame (121, 89, 89 and 161 frames; 189 came
back as 185), and nothing shorter than 89 frames (3.709 s): 79- and 96-frame sources both came back as 89. Wan keeps
its input's length in whole seconds: pad its source to a whole number of seconds with the last frame held.
3. **Write it back.** Window frame k takes delivered frame round((k + pad) · (n_r − 1) / (n_s − 1)), with n_r the frames
delivered and n_s the frames sent, which absorbs any rescale. Copy every original frame aside first; an undo puts
them back. A window that doesn't start or end on a cut cross-fades over a few frames.
4. **Keep what the model must not touch.** A re-light softens small type ("a touch softer, still clean and legible") and
can smear a card outright. Fixed rectangles keep the original pixels, feathered (a Gaussian of half the feather, 24 px
by default): the status bar, the Dynamic Island, an interface platter. A card that moves gets a rectangle per frame,
tracked: `{"": [[x, y, w, h, feather], …]}`, rounded rectangles of radius 28 with a feather of 12. DON'T LET ME
SCROLL tracked a notification card over 48 frames this way.
5. **Colour-match a Wan pass first** ([tool 39](#39-the-colour-match)).
Ten splices went into DON'T LET ME SCROLL's final, four Seedance re-lights and six Wan passes, from 28 frames (a 1.2 s
bridge whose flat blue background Seedance replaced with a glass atrium) to 188 (the lock screen's digits re-lit as
ray-traced liquid glass).
**Done when** the window's frame count matches, the cuts on both sides show no seam on a seam sheet, the kept rectangles
hold their original pixels, and the undo restores the cut exactly.
### 39. The colour match
**Job.** Put a re-lit clip back in its source's palette, without flicker.
**In → out.** The source clip and the re-lit clip → a matched clip, which the splice uses instead.
**How.** Sample 24 frames spread evenly across each clip (at (k + ½)/24 of its length), at 480 × 270. Per channel, take
the mean and the standard deviation over all the sampled pixels; the gain g = σ_source / σ_relit, clamped to 0.7–1.4, and
the offset o = μ_source − g · μ_relit. Apply that one map to every frame (ffmpeg's `colorchannelmixer` for the gains,
`lutrgb` for the offsets, clipped to 0–255) and encode at CRF 12. One map for the whole shot can't flicker; a per-frame
match would.
Wan doesn't shift colour the same way twice. Bright paper scenes came back darker and cooler: S30's warm grey paper fell
from a mean of about 172 to about 111, red furthest, and the match brought it back to about 169. Dark scenes came back
brighter: S19's mean rose from about 34 to about 53.
**Done when** the matched clip's channel means sit within a few levels of the source's, and the two side by side show the
source's palette.
### 40. Wan 3.0 relight-and-repair
**Job.** Re-light a finished shot as a physically lit one and repair her body in it (limbs that leave the take's frame
completed, a stray limb removed), keeping every element in place and in time, so it splices straight back. The director's
method: "you could maybe just throw the entire clip thru Wan 3.0 (with your edits and stuff) and specify to remove the
foot so you dont have to re-edit. and have it relight at the same time."
**In → out.** The rendered shot as Video 1 of Wan 3.0's reference-to-video, optionally a face reference as Image 1 → a
re-lit clip of the same length and layout.
**How:**
- **The source:** the shot's own frames from the render ([tool 38](#38-the-re-light-splice)), then its last frame held
to a whole number of seconds. 1080p, 16:9, 4–8 s.
- **The prompt opens** "Reimagine Video 1 as the same shot, physically lit and photographic. Keep every element exactly
in place and in time: the same layout, positions and sizes, the same timing, every movement of the woman; the camera
stays exactly as in Video 1, no cuts, no new objects, no added text." Then her description, the canon's short form
with the shot's wardrobe; the repair; the light ("Re-light the scene as a real dark bedroom at night, lit by the cool
glow of the giant glass clock numerals and the moon"); and what stays exactly as it is ("The clock numerals, the moon,
the notification cards with their words, the status bar and the black Dynamic Island at the top stay exactly as in
Video 1, sharp and legible").
- **The repair, as a sentence about her whole body:** "Her whole body stays complete in every frame: both arms end in
complete, natural hands with five fingers, never cut off or faded at any edge, and both feet are complete." A stray
limb: "Remove the stray bare foot and leg that appear beside her at the lower right of her body: there is nothing there
but the dark room, and her body continues naturally below the frame."
- **A face reference fixes identity, not pose.** The re-drawn shots lost her face, so the director picked a reference
(an earlier film's identity portrait, cropped to her face and hair so its clothes couldn't bleed in): "Image 1 is the
face of the woman: keep her face exactly as in Image 1 in every frame … Image 1 is only a reference for her face; it is
never shown as a shot." In one shot her chin was tucked into a sweater's collar, and her face still read full until the
prompt added "her face is slim and angular, never round or full: slim cheeks, defined cheekbones and jaw, a slender
neck, and her chin lifts slightly out of the sweater collar so her jawline shows".
- **Re-make from the original render,** never from an earlier generated pass. The director: "id just re-do the relighting
with the face ref added from the original scene rather than rerunning the generated ai scene into it again."
- **Put the interface back.** Wan softens small type too: the stocks screen's interface and the status bar came back from
the original render, as two kept rectangles.
- **Colour-match** before the splice ([tool 39](#39-the-colour-match)).
DON'T LET ME SCROLL made ten of these passes, at $0.20 a second at 1080p list for the clip and the same again for each
second of the input video ($2.00 for a 5 s shot). Six are in the final: a stray foot removed, a lost foot restored, hands
and sleeves that left the frame completed in three shots, and her face held to the reference in four.
**Done when** the pass, colour-matched and spliced, plays with every element where it was, her body whole in every frame,
her face on the reference and the interface's pixels the original's, judged in place in the cut, side by side with the
original.
### 41. Roto on a CPU box
**Job.** Her matte for every frame of a take, without a GPU, fast enough to feed the chapters.
**In → out.** A take's frames → `matte/f%04d.png` beside them (8-bit grey, full frame size), and a count that must equal
the frame count.
**How:**
1. **BiRefNet on keyframes:** the general-lite model (through rembg, no post-processing) on every 3rd frame and the last.
On the render box's CPUs it took about 4.6 s a frame when run on every frame; with keyframes, about 1.6 s a frame, a
take every four to five minutes.
2. **Optical flow between them.** For each frame between two keyframes, OpenCV's DIS flow (its medium preset, on a
half-size grey copy) from this frame to each keyframe; warp each keyframe's matte backwards along it (bilinear, edges
replicated) and blend the two by distance, (b − i)/(b − a) to the earlier keyframe. Hair edges and the headset cable
survived it.
3. **Turn off onnxruntime's CPU memory arena** (`enable_cpu_mem_arena = False`, `enable_mem_pattern = False`). With it
on, one 1080p roto held about 12.5 GB, and rotos died for lack of memory on a box shared with chapter renders; with it
off, 1.3 GB.
4. **Fail loudly.** Never pipe the remote run through `| tail`: its exit status hid a roto killed for lack of memory,
which came back with 35 mattes for 210 frames. Count the mattes against the frames, and stop when they differ.
5. **The full model for hard frames.** BiRefNet's general model recovered a motion-blurred back shoe that general-lite
had lost.
6. **Shoot for it:** her alone on plain mid-grey, with the take wider than the shot needs. Limbs that leave a 9:16
take's frame can only be faded at its edge. The director: "all your roto cutouts have blurred bounding boxes it seems
and sometimes her limbs go outside of them and just disappear at the boxes edge". Wan completed them later
([tool 40](#40-wan-30-relight-and-repair)).
**Done when** every take has as many mattes as frames, and a sheet of the mattes over their frames shows her hair,
fingers and feet whole.
### 42. Beat warps floored at 1×
**Job.** Land her moves on the song's beats without ever slowing a take. The director on the first cut: "never slow these
down. you lowered the framerate for no reason, just looks bad."
**In → out.** A take's frames and mattes, the shot's window in song time and the beat map → a monotone piecewise-linear
map from song time to take time ([Data formats](https://claudiagallery.vip/wiki/formats/#takes-in-a-world-drawn-in-code)).
**How:**
1. **Her accents.** Per frame, the mean absolute difference of the grey frame (270 × 480) from the one before, inside her
matte (the larger of this frame's and the last one's), smoothed over three frames. An accent is a local peak of its
rise above the 80th percentile, at least 0.22 s after the last.
2. **A base map:** a speed in {1.0, 1.08, 1.16, 1.25} and a start offset in steps of 1/60 s, scored by Σ exp(−(d /
0.07)²) over the accents, with d an accent's distance to the nearest beat.
3. **Pin the accents.** Each accent within 0.2 s of a beat moves onto it, as long as the speed between pins stays within
1.0–1.45× and the pins are at least 0.1 s apart.
4. **Never below 1×.** A take shorter than its shot plays at 1× to its last frame and is flagged: it holds, loops on a
matching pose, or gets a longer take. DON'T LET ME SCROLL's first warps had dipped to 0.70–0.89× in nearly every take,
and one chapter's own map to 0.53×. With the floor, six shots were flagged short. One became a two-beat loop near the
take's end (0.917 s, from 4.92 s back to 4.00 s of the take) where her pose repeats; the stutter became a sampler, the
take at exactly 1× between triggers and each vocal chop re-triggering a cue point; two were taken again.
**Done when** every segment of every warp runs at 1.0× or faster, the accents land on beats, and no shot ends on a held
frame unless it was flagged and accepted.
### 43. The single-frame outlier scan
**Job.** Find frames that differ from both neighbours while the neighbours agree: the signature of a renderer that drew
one frame wrong.
**In → out.** The rendered frames → the frames that stand out, each with its difference from its neighbours.
**How.** Compare every frame with the per-pixel median of the frames around it, as the mean absolute difference. A
single bad frame stands far from that median while its neighbours sit close to it; a cut or a fast move shifts them
all. Scan at two thresholds, a strict one and a loose one, and look at every flagged frame beside its neighbours. A
pattern in the flags (every 8th frame, every frame of one shot) names the cause.
On DON'T LET ME SCROLL the director saw black vertical lines in one shot, "they appear occasionally, look at more framers,
the spaz out". The scan found dark columns on every 8th frame of the shot: one of the eight render workers had drawn its
share from a stale camera state. Rendered fresh, one of those frames differed from the worker's version by 21.5 levels on
average, and from its good neighbour by 4.0. The worker's 16 frames were rendered again, and a scan of all 3,859 frames of
the film found none left. Every other check had passed those frames.
**Done when** a scan of the whole final finds no frame that differs from both neighbours while they agree.
## A film shot twice, on the bar map
WHITEPILL, the seventh film, was shot twice. A performer model made a rough, lip-synced take of every shot of hers; then
Seedance 2.5 made each shot again as the master, singing from the song's own vocal stem, with the rough take shrunk to a
guide or with no guide at all. Its worlds are shaders, one rule each, photographed by the same video model. Everything
sits on the song's bars. These tools do that. Build them if she sings on camera, if your worlds are code that a video
model should photograph, or if your cut has to sit on the music.
### 44. The bar map as the timeline
**Job.** Write the film in the song's bars, so that picture and music are consonant: every cut on a bar line, every move
in bars. WHITEPILL's director on its first cut: "there is this temporal signature that screams claude now that a ton of
people are making the videos. people say pacing or prosody. i think it comes from a couple things. 1- youre not working on
an A/V timeline in the way we are in video programs. 2- things might not cut on the right beats? like, maybe we need the
beat structure and only have cuts at ends of bars or sections. like, the audio structure and video structure need to be
consonant. 3- i think you use cuttng to cover up imperfections from video models rather than for timing. 4- i think youre
not as aware of time as humans are, so time feels more like a number or concept or count rather than a duration."
**In → out.** The beat map's downbeats, the sections and the lines' times → the shot list in bars
([Data formats](https://claudiagallery.vip/wiki/formats/#the-shot-list-in-bars)); after every assembly, each cut with its bar and kind, and the
picture's motion against the arrangement, bar by bar.
**How:**
1. **Shots in bars.** Each shot is [bar a, bar b) of the song, one place per shot. Its camera note names bars, not
seconds ("pushes in slowly and steadily until {b:5}; from {b:5} it cranes up"), and the tool that writes a take's
prompt turns each {b:N} into that take's own seconds.
2. **Cuts only on bar lines,** and only at a section's start, a phrase line (2, 4 or 8 bars from the section's first
bar) or the downbeat where a sung phrase begins, so the picture arrives with its line. A cut anywhere else is a
decision, written into the shot list.
3. **Moves in bars.** A move starts on a downbeat and comes to rest on a downbeat, a bar before the cut where it can, so
it settles before the next shot.
4. **The song sets the length, and the model fills it.** A shot longer than a model's limit goes in squeezed and comes
back stretched ([tool 45](#45-a-previz-performer)). A failed take is made again, never cut around.
5. **The lyrics keep the same clock.** A line comes in half a beat before its first word and leaves 0.4 s after its last
(or when the next line comes in). A line whose first word comes up to 0.4 s after a cut, or whose short pickup starts
less than 0.35 s before one, comes in on the cut; a line that has ended leaves on the next cut.
6. **Measure every assembly,** at 160 × 90 grey:
- **cuts:** a frame whose mean difference from the one before passes max(4 × the film's median difference, 18 levels)
and 2.5 × both neighbours'. Each is listed with its bar and whether it falls on a section or phrase line or where a
sung line starts, and every cut inside a sung line is listed with the line;
- **motion against the music:** per bar, the median optical-flow speed (cut frames left out) against the energy of the
drums, bass and other stems, and their correlation over the bars;
- **sheets:** one frame per bar, ten bars a row, the first bar of each section marked; and the two sides of every cut,
side by side.
Measured on WHITEPILL. Its previz: 42 picture changes (cuts and snaps inside shots), 95% on a musical onset within one
frame (42% by chance) and all of them on the beat grid, counted by an audio-analysis tool. Its final: 29 picture
changes, every one on a bar line, 28 on a section or phrase line or where a sung line starts; 4 inside a sung line, 3 of
those on section lines. Motion against the music: the first cut had scored r = +0.43 against the sections' loudness, and
the plan aimed for +0.7; the final scores +0.44 bar by bar against the arrangement, so that aim wasn't reached.
**Done when** the measure finds no cut off a bar line, every cut inside a sung line is a written choice, and the director
has watched the cut in real time.
### 45. A previz performer
**Job.** A rough take of every shot of hers, lip-synced and dancing on location, made as often as needed, as the guide
for the master. WHITEPILL's director: "i want you to basically shoot the entire video twice - you can previz with […]
and you can master with seedance AFTER you animate".
**In → out.** Per shot: the place with nobody in it, her face sheet, the dress, the vocal of the shot's window and, for a
dance, a real dancer's clip → one take of the whole shot, with no cuts.
**How**, with MiniMax H3, in its own reference format (what each input is, and what is kept of it):
- **A masked performer.** Her whole head is in a green chroma-key balaclava, her eyes and lips showing through its holes
(taken from the face sheet's front view), and she wears the film's dress. The master redraws her face; the previz
carries the motion, the mouth and the camera. The director: "you could make the initial shot be a like greenscreen
person with a mouth that lip syncs and eye movements lol", and then: "honestly i think itd be better if the model was
just the greenscreen model but in the real world, dancing, moving, etc - doing the entire thing. not in a green studio
staying still."
- **The words exactly.** The prompt gives exactly the words heard in the take's audio, in order: a line the window cuts
is marked "…", and every silence is described ("her lips stay gently closed and still … until her voice begins").
Whole lines where the audio held part of one put her mouth on words that weren't there. No clock times in the prompt:
they made the takes cut.
- **The vocal window** is silent before the first word the prompt names and after the last, and nowhere else: gating a
gap inside a take cut a held 2.5 s note. A word sung across a cut belongs to both takes for as long as it is heard,
within 12 dB of its peak (held notes stayed within about 11 dB; fading tails dropped 15 dB or more).
- **Words that open the mouth.** "Smile" in a take's mood drew a toothy smile in the silence before she sang, and "smiles
between the lines" read as talking. The mood puts the joy in her eyes; the lips stay closed until the voice begins.
- **Dancers:** clips of adult dancers from public dance videos (one dancer, the whole body in frame, a still camera, a song
within about 6% of the film's tempo), for motion only. Matte the dancer onto flat grey: given the whole video, H3 put
her in the dancer's background for a few seconds. Re-cut each clip for its shot at (the shot's tempo / the clip's
tempo) × the squeeze: one shared cut ran 2–4% slow, almost a beat behind by the end of eight bars.
- **Length:** at most 362 frames, 15.08 s. An eight-bar shot (15.5 s) goes in squeezed, its vocal sped up by
q = length / 14.9, and the take is slowed back by q when it is cut in, so the lips stay on the song.
- **A calm camera,** one move or none, written in the take's seconds from the bar map
([tool 44](#44-the-bar-map-as-the-timeline)). WHITEPILL's first previz orbited her in every shot ("the camera should
feel like it's on an industrial robotic arm"), and the director dropped it: "its kinda dizzying now and previz was too
complex so h3 didnt keep up wtih the physics or consistency".
- **Only cut-free takes.** Several seeds a shot, each pick checked for jumps as well as cuts (an automatic cut check
missed one). A seed doesn't pin a take: one seed with the same inputs gave three different takes.
What it couldn't do: its dance takes lost limbs where the dancer's matte had lost them, and a master copies what its
guide shows ([tool 47](#47-takes-with-no-guide-clip)). A whole-film reshoot in Wan 3.0 Prime over the full-resolution
takes was rejected: ":/ looks like shit. i feel like h3 is dragging it down." What worked was the take shrunk to
320 × 180, a rough guide for everything but the lips ([tool 46](#46-lip-sync-from-the-songs-own-vocal)).
**Done when** every shot of hers has a cut-free take whose mouth shapes the words of its window, read by eye on strips of
her face, each frame marked with the vocal's level and the word sung then.
### 46. Lip sync from the song's own vocal
**Job.** Her singing shots, lip-synced in the master itself: whole shots of up to 16 s, camera moves and all, with no
lip-sync model and no covers.
The method came from a tip in a public forum thread that the director found: give the model the vocal stem alone, type
the exact lyric line, make the clip exactly the generation length, and lay the music back over. The director on the
first take: "OMG THE SEEDANCE W/ SUNO VOCAL STEM METHOD WORKED!!!!! LFG!!!! FINALLY. IT"S PERFECT".
**In → out.** The shot's window in song time, Suno's lead vocal of the locked take, the shot's references → a take whose
mouth sings the window, with the song laid back over it.
**How:**
1. **The vocal from Suno itself.** Suno's stem split of the locked take (its Lead Vocal stem) matched the master to the
millisecond: the same length and no lag. Check the stem against the master by cross-correlation (zero lag), not by
title or length: the first stems came from the wrong takes, one of them matched to the song by its length.
2. **@Audio1 is that vocal alone,** with no music and no gating, cut at the shot's song time to exactly the generation
length. Seedance takes whole seconds, d = max(4, the shot's length rounded up); a 16 s take gets 16.000 s of vocal from
the shot's first frame. Check it with ffprobe.
3. **The exact lyrics in the prompt,** in plain quotes, in whole-second segments of what @Audio1 holds
([the template](https://claudiagallery.vip/wiki/seedance/#template-singing-from-the-vocal-stem)): each phrase once, with "the line is already
under way on the first frame" or "the line goes on after the last frame" where the window cuts one, every silence as
its own segment ("her mouth is closed and still"), and a word that only begins in the last instant left out. A word
counts when at least 0.15 s of it, or half of a shorter one, is inside the window, with its end followed on the vocal
stem (word times: [tool 5](#5-timed-words-and-the-ear-check)).
4. **References, each with one role:** her face sheet from every angle ("identity only, never shown as a frame"), her
dress from four sides, the film's look ([tool 51](#51-the-look-from-one-take)), and either the previz shrunk to 320 × 180,
scaled back up and padded to the same length ("a low-resolution rough guide to this shot, not a precise reference …
only the lip sync is exact"), or no guide and the empty place ([tool 47](#47-takes-with-no-guide-clip)). Seedance 2.5
on Higgsfield (References), 720p, 16:9.
5. **Lay the song back over** at song time. The take starts at the shot's song time, so the shot is the take from its first
frame, at 1×; the take's own soundtrack is thrown away.
Measured: the first take (16 s at 720p with its 16 s guide, $8.87 at list on Higgsfield's API) came back on-model, the lips
on the words, the push-in and the crane up out of the wheat done as one move. The spectrogram comparison
([tool 15](#15-the-sync-check-the-spectrogram-comparison)) called its soundtrack re-sung rather than copied, matched only
from 7 s on, yet the singing sat exactly where the vocal is. The same shot, side by side: Wan 3.0 Prime singing to the song
shaped the words but hard-cut at 9.9 s instead of craning up, masked or not; Kling 4.0 kept a guide's motion frame for frame
and looked the most like film, but it takes no audio track, and her mouth drifted into a smile on the words.
**Done when** every singing shot, with the song laid over, reads in sync by eye at full speed. The director judges it
there, not from the numbers: "Just show it to me instead of using your eyes for the lip sync".
AERO's director made it the rule in the production ask: "vocal stem -> seedance 2.5 is the way we get lip sync to work - the
song alone doesnt help as much as vocal stem, and then aligning from spectrogram." Every take of hers that sings or speaks
had its window of the lead stem, a first frame on her design as @Image1 (an image-model edit from her sheet, or a frame of
an earlier take that continues the shot), and the sheet itself as @Image2. Its takes needed one more step: the model
shortens the pauses inside a line, so each phrase is measured and placed on its own
([tool 69](#69-lip-sync-phrase-by-phrase)).
FABLE kept the method and changed what feeds it and what places it: windows timed by the stem's level and gated by line
([tool 80](#80-the-shot-book-and-gated-windows)), each take aligned as a path
([tool 81](#81-sync-as-a-path-each-take-aligned-to-its-window)), and some of her takes made with the sound off, the vocal
still as @Audio1, and placed by her lips on the consonants ([tool 82](#82-takes-without-a-soundtrack-placed-by-her-lips)).
### 47. Takes with no guide clip
**Job.** Shoot a shot of hers with no reference video when a guide would be copied with its faults: limbs a previz dance
lost.
**In → out.** The place with nobody in it, her face sheet, the dress, the look, the vocal (or the music without it:
[tool 48](#48-a-take-that-doesnt-sing)), and the action, the camera and the dance in words → a take.
**How:**
- **The place plate:** a frame of the shot's earlier take with her removed by an image model (GPT Image 2.5), given as
"@Image4 is the place of this shot, empty: take its set, landscape, layout, light and camera view; Claudia is the only
person in it." Keep whoever belongs in the shot: the schoolhouse's plate kept the child at the desk ("keep the child as
in @Image4"). The previz's places were made the same way in Midjourney, from the storyboard plates
([Midjourney](https://claudiagallery.vip/wiki/midjourney/#style-families)).
- **The camera in words,** as the earlier take did it where the director had seen that version, otherwise from the shot
list.
- **The dance in words, beat by beat, with the tempo.** A 【Dance】 section opens "The song's tempo is 124.5 beats per
minute (a beat every 0.48 s, a bar every 1.93 s); every step and accent lands on a beat. Her whole body dances, smooth
and joyful, her arms, hands and legs always whole and clearly visible.", then one move per segment of one or two bars
("4-6s: the disco point: on the downbeat her right arm points up to the sky on the diagonal, then down across her body,
then the left arm."). The core requirements add "her arms, hands and legs are whole and attached the whole time, never
vanishing, merging or bending wrongly".
- **Name a removal only where the thing is.** A fix written for an earlier take ("Remove the thin red cord stretched
between the apple trees") went into a prompt whose place had no cord, and drew one.
Why: the director on the first full cut, of four dance shots: "it must be that the roto or whtaever on dance refs lose
bodyparts and the video follows it exactly because these hav the same limb losing artifacts as h3 we could just tell the
model to make her dance instead of forcing a reference that we failed to properly roto. just describe her exact dance (and
all rest like camera movements etc same) for each of those scenes."
Measured: of those four, two are in the final reshot this way (2:27 and 3:10), and two became takes that don't sing
([tool 48](#48-a-take-that-doesnt-sing)). Ten of her twenty shots in the final have no guide clip; the director chose
between the guided and unguided versions section by section.
**Done when** the take shows the plate's place, the camera as written, and every limb whole in every frame.
### 48. A take that doesn't sing
**Job.** A dance shot whose lip sync won't hold, made to dance instead. The director on such a shot: "honestly lets just
remake this with her not singing, fully in the groove, just dancing around and having fun, she turns around, head goes
down adn up etc - lip sync failed here but i dont think we need it. new seedance clip there. youd need to prompt
forbidding singing here, and make sure she's happy."
**In → out.** The song's music without her voice, the shot's references and the dance in words → a take of her dancing,
her mouth closed or laughing; the song with her voice laid over it.
**How:**
- **@Audio1 is the song without her voice:** the master minus 0.89 × Suno's lead-vocal stem, cut at the shot's song time to
exactly the generation length. An audio check heard house and disco with a trace of voice.
- **The audio policy says so:** "It is music only, with no voice. Do not add singing, vocals, humming, speech, narration or
sound effects."
- **A 【Mouth】 section:** "She does not sing, speak or mouth any words at any moment: her lips stay closed in a smile, or
part only when she laughs; her mouth never moves in time with the music."
- **Requirements that forbid it and ask for joy:** "1. No singing: Claudia never sings, speaks, hums or mouths words; her
lips are closed in a smile or open in a laugh. 2. She is happy the whole time: smiling, laughing, carefree."
- **The dance in words** ([tool 47](#47-takes-with-no-guide-clip)), with what the director asked for: she turns all the way
around, her head goes down and up on the beat.
Measured: for the second such shot the director had both made, one more lip-sync take from the vocal and this, and chose
this one: "ok without singing it is." The lines stay on screen as lyrics: "always have the lyrics".
**Done when** she never mouths a word and dances on the beat for the whole take.
### 49. A world as one rule
**Job.** A world that reads as mathematics, written as one GLSL shader on the song's clock: a guide a video model can
photograph ([tool 50](#50-shader-worlds-as-seedance-guides)), or a frame that is the picture. The director, resetting the
film: "i think SHADERS are the key actually! shaders are math + viz, shadertoy is amazign and i bet nothing you make via
shaders will feel like web dev. i think that's it. we need to storytell thru shaders." On the first ones: "Idk if feels
like you're adding ... items to increase complexity rather than compressing intricacy? Does that make sense?"
**How:**
- **One iterated rule per world;** the light and the colour come from the loop, not from added objects. WHITEPILL's three:
- **a tree of endings:** one fold (move to the tip, mirror, bend, twist by 1.6 rad, scale by 1.25) repeated 24 times,
2^24 = 16.8 million tips. The fold's level sets the colour, an amber trunk to white-gold tips, and a flag on each path
sets which branches are lit. Her timeline is one path lit from the root, three levels a beat, the camera chasing the
light, its target found by running the fold backwards. For "fourteen million endings" every path lights, a level and
a half a beat;
- **origami:** one sheet folded again and again, `p = abs(mod(p + a, a + a) - a)`, with a turn and a crease relief per
level. On each downbeat one fold level snaps in, eased over about four frames before the beat. Wrapped inside an
O'Neill cylinder it is the colony; in the break it folds into each line's object;
- **the pill:** one capsule folded into its eight children, and again: the colony is a pill whose towns are pills whose
houses are pills.
- **On the song's clock:** the bar and the beat are the shader's inputs, and its moves are written in bars
([tool 44](#44-the-bar-map-as-the-timeline)), so nothing drifts.
- **Real data only:** the backbone of ubiquitin (PDB 1UBQ, its 76 Cα atoms, with their real handedness), the wiring of the
Enigma I's rotor I (EKMFLGDQVZNTOWYHXUSPAIBRCJ, stepping on beats 2–4 like key presses), the number 42, the week
(Monday and Tuesday rise, the other five fold flat).
- **Compressed:** each world also fits in a tweet-length shader, 478, 493 and 493 characters, checked as GLSL ES 3.00.
- **Rounds in parallel:** agents wrote rival versions of a world, each round kept beside the last as a fallback, and the
picks went by contact sheets and short reels played with their bars of the song.
**Done when** a contact sheet of the shot reads as one world, every change lands on its beat, and nothing in it is there
for decoration.
### 50. Shader worlds as Seedance guides
**Job.** Turn a shader shot into film of real, hand-made things, keeping what it shows, when, and the camera's move.
**In → out.** The shader's render of the shot and stills of its key moments repainted in the film's look → a take on the
same clock.
**How:**
1. **Render the guide at 1080p, supersampled,** on a render box, and pass it as it is (CRF 14). The guide's resolution
trades detail against obedience:
- at 320 × 180, scaled back up, the timing and every object landed on their beats, but fine detail came back soft and
blocky ("the MTWTFSS and the next 'pyramid' shared still is all pixellated");
- at full resolution with no still per object, a tree came back sharp, but the shot of paper objects drifted half a
second and swapped objects: the Enigma rotor became a helix, and the weekday cells lost their letters;
- at 1080p, supersampled, with one still per object, the paper shot kept all five objects in order and on time,
letters included.
2. **One still per object,** each a crisp frame of the render repainted by an image model in the film's look (with the
look's reference: [tool 51](#51-the-look-from-one-take)), each with its role ("[Image2] shows how the paper looks with
the Enigma rotor: its materials, light, colour and film texture."). The prompt names each object with its time and its
still, and adds: "Only one object at a time; each arrives on its downbeat, not before." Without stills a shader stayed a
computer render.
3. **Freeze the guide on its last wanted frame** when the shot is cut short, so the take has nothing to drift into: the
break's paper stopped 9.42 s into the render, before its pyramids, and held there.
4. **Add what the shader doesn't have** with a still that shows it and one sentence: "They are not in [Video1]: add them to
it". A cloud-field shader the director supplied, a sunlit field of cumulus flown through low, got its hot air balloons
that way, as they had asked: "add techno hot air balloons in the right aesthetic by prompting seedance to add those
after rendering".
5. **The guide and the still set the palette.** From the sunset render, the clouds came back sunset gold: "wayy too
saturated compoared to teh pastel light colors of the rest". A day variant of the shader (a white sun, a pale sky, less
blue absorbed) and a pastel still made without the film's saturated look reference gave white clouds under a pale blue
sky, with "Everything stays white and pale blue: no gold, orange or sunset colour anywhere."
6. **The prompt** ([as it ran](https://claudiagallery.vip/wiki/seedance/#shader-worlds-as-guides)): the guide is "a clean, sharp 1080p computer
render: take from it exactly what appears when, the shapes and where they stand, the camera move and the pacing, and
re-render all of it as real things on film", then a timeline of what appears when, "It looks like real film of real,
hand-made things, never like a computer render", and "Fix every motion artifact".
Measured: Seedance 2.5 on Replicate at 720p, $0.9676 a second with a video input (October 2026). The higher resolution
wasn't better everywhere: for two shots, the farming ark and the first tree, the director kept the earlier takes from the
320 × 180 guides: "we regresed on those."
**Done when** the take shows every object at its time and the camera's move, looks like film of made things, and sits in
the film's palette.
### 51. The look from one take
**Job.** One colour and film look for every shot, picked by eye from real takes rather than written in words.
**How:**
1. **Test palettes on a few shots, with no image of her face.** WHITEPILL's director asked for "very wes anderson". Three
palettes named after his films (the pastel of Asteroid City, the warm Super 16 of Moonrise Kingdom, the pink of The
Grand Budapest Hotel) went on the opening shot, and the pastel one on the bed scene and a shader scene too. Her
identity there was the canon in words.
2. **The director picks a take, not a word.** The bed take came out richer than its palette's words: "i like how satruated
it is in bed scene. thats the look i like teh most".
3. **Crop it to a frame with no person in it,** and give it to every shot as a colour reference with one role: "@Image3 is
a colour and film-look reference only: take its colour, light, contrast and film texture; do not use its place,
pavilion, bed, columns, plants or anything else in it." Write the look in words as well ("shot on 35 mm film with the
rich, saturated colour of a 1970s colour-film print … warm creamy whites with a golden cast, clear cyan blues"), and the
shot's own light after it ("It is night, lit by the warm deck lights of the ship").
4. **The shader scenes' stills are repainted with the same crop as a reference** ([tool 50](#50-shader-worlds-as-seedance-guides)),
except where the look doesn't suit the scene: the clouds got a pastel still made without it.
5. **Nothing added after.** No grade, and no grain in the first master: "don't add film grain to v1, we'll have to tune it
iteratively".
**Done when** the takes cut together in one palette with no grade, each keeping its own light.
## A game's world, through her eyes
THE CAKE IS NOT A LIE, the eighth film, is set in Portal 2's world. Claudia is GLaDOS: her porcelain face on GLaDOS's
boom, singing the whole song in her chamber while her memories play on the screens around her. The film is seen through
her own optical interface, a HUD built as one product and composed shot by shot; the song's own glitches, where her
voice is used as an instrument, glitch her head and break her HUD on the beat; the shots are measured and graded into
one Aperture look (a warm variant for the warm scenes); and the cut was measured against the song and re-cut by what
each shot holds. It was made three times: a first film on its first song, set aside when the song changed; a restart on
the new song; and a storyboard that made every unit's first frame again in Midjourney, with the reshoots after it. These
tools do that. Build them if your film carries an interface that must feel like one product, if the song's own sounds
should drive the picture, if a cut mixes sources that need one look, or if a cut feels wrong and you need to know why.
### 52. The song map and the glitch map
**Job.** Everything the edit and the picture sync to, measured from the stems: a 16th-note grid that follows the take's
drift, bars with their half bars and beats for cuts, the sections, every word with its place on the grid, and the glitch
map: every place where her voice is used as an instrument, with the measured time of each hit. The director: "oh yeah we
have to be aware bpm slightly shifts w suno over the whole song so we may need to make sure the glitch timing is right
with the drifting bpm." And on finding the glitches: "it should be easy enough to match because it looks like the voice
turned into repetitive beats per spectrogram or whatever."
**In → out.** The master, Suno's lead vocal stem and the instrumental → `bars.json`, `sections.json`, `words.json` and
`glitchmap.json` ([Data formats](https://claudiagallery.vip/wiki/formats/#the-song-map-and-the-glitch-map)), and one strip per section.
**How:**
1. **A 16th grid from the drums.** Track beats on the instrumental's percussive part (librosa: harmonic–percussive
separation, onset strength at a hop of 128 samples at 22.05 kHz, a start tempo near the style's and a stiff tempo
prior, `tightness` 200). Move each beat to the strongest drum onset within 30 ms, cut every beat interval into four,
and extend the grid before the first beat and after the last by the median period. Check it: the share of strong hits
(peak-picked onsets, the top 40% by strength) within 20 ms of a 16th, in each third of the song.
2. **Bars from the grid.** Bar k starts at grid16[1 + 16(k − 1)]. Confirm the phase from the snare: its onset strength,
summed per grid index mod 16, should peak on indices 5 and 13 (beats 2 and 4), and every section change should land on
a bar line. Each bar carries its start, its half bar, its four beats and its sixteenths, and cuts go on those. Never one
bar length: read every time from the file.
3. **Sections by hand, on bar lines,** each with what changes there ("the instrumental STOPS 53.60-54.25 s ('Raise'
pickup in the gap)"), and every pickup listed: a line sung before its bar line belongs to the section after it.
4. **Words** from a transcription with word timestamps (faster-whisper) of the clean lead, each start snapped to the
strongest lead onset from 80 ms before to 60 ms after, fixed by hand where the transcript is lost (from the pitch track
or the bar's downbeat), each with its bar.beat.16th. Good to about ±0.1 s: enough for lyrics, not for cuts. Cut on the
bars or on the glitch hits.
5. **The events, curated; their hits, measured.** Find the events by ear and by eye on the strips (spectrograms of the
lead and the instrumental, pitch tracks, envelopes), and run demucs over the instrumental to see which ones Suno's split
left in both stems. Each event names a method, and the method measures its hits, each moved to the transient (the
steepest rise of a 2 ms energy envelope on a 1 ms hop):
| Method | For | A hit is |
|---|---|---|
| burst | chops, stabs, gate openings | the envelope 14 dB over the range's floor after at least 8 ms below it, for 25 ms or more |
| dips | echo tails, gated holds that never go silent | a rise of 4 dB or more out of a dip of the 10 ms-smoothed envelope |
| f0 jump | a register drop, octave flips | the median pitch of the next 30 ms 7 semitones or more from the last 30 ms, both sides voiced (pyworld's harvest, 5 ms frames) |
| re-trigger | a held vowel repeated on the grid | per 8th or 16th slot, the strongest spectral-change peak (cosine distance between frames 4 ms apart, 150 Hz–6 kHz) within ±30 ms, above the range's 55th percentile |
| onsets | what the band plays | spectral-flux onsets of the instrumental |
A re-trigger event carries its grid contrast, the mean peak on the slots over the mean between them: above about 1.3
the re-triggers are real; near 1 the vowel only flutters, and its hits are the grid itself.
6. **A strip per section:** both spectrograms with their envelopes, the bar lines and beat ticks, the events shaded, the
hits as red lines, the words.
Measured on THE CAKE IS NOT A LIE (208.76 s, 84 bars and a tail of 10 sixteenths): the take drifts from 95.7 to 99.4 BPM
(median 97.5; a bar of 2.44–2.50 s), and 91%, 99% and 99% of its strong hits sit within 20 ms of a 16th in its three
thirds (the intro's glitchy drums are the 91%). Fifteen events: the hook's "lie" frozen into a drone that flutters on 8ths
under the drop (25 hits); the intro's stab, sliced holds, chops and swell; the register drop of choruses 1 and 2 (her
voice in a lower register, flipping back and forth: 7 and 8 hits, up to about ±75 ms off the 16ths because they are pitch
changes inside syllables); the "lie" stutter (13 hits on 8ths, grid contrast 1.78), its return after the re-sung "is not
a lie" (22) and the breakdown's 16th re-triggers (33); two echo tails; a held "lie" over a stuttering breakdown; and the
spoken stutters, "four, four, four, four, four" and "You are, you are, you are". Every example the director gave is on
it. The treatment's own 65 pitch cracks are listed too, at a lower priority.
**Done when** every section change sits on a bar line, the strong hits sit on the grid in every third of the song, and
every event's hits, drawn on its strip, sit on what the spectrogram shows.
### 53. The head glitch on the song's hits
**Job.** Glitch her where the song glitches, and nowhere else: a few short moments on her shots, each on the measured
hits of one glitch event, inside a matte of her head. The director on draft 5: "I think we should never have fullf rame
glitches for more than like 2 seconds, so we need to remove those. The glitches that last longer should purely be in the
actual video generation itself, where you can specify to seedance where the robot glitches a little. however its
unpleasant to look at for ourselves, so it should be used creatively, sparingly, and toned down. … if anything we remove
all of them, and we can have roto that induces the glitch when you decide to keep them (the <2second periods where we
keep them) - like roto the claudia head and just glitch her in post".
**In → out.** The cut, the glitch map ([tool 52](#52-the-song-map-and-the-glitch-map)) and a list of moments (the
event, the shot, how many of its hits, a strength, a kind) → the cut with her head glitched on those hits, a matte per
frame the moments need, and a QC sheet of the mattes.
**How:**
1. **Moments, not a pass.** A moment is one event on one shot: its hits inside the shot, from a frame before the first to
4 frames after the last, at most 1.9 s. Everything else passes through untouched, and so does any frame without a
matte.
2. **Her head's matte.** Only the frames the moments need leave the cut, seeked frame-exact. In each moment, YuNet
(OpenCV's face detector, score 0.55 or more) finds her face on the frame where it is surest (score × width). The head
box round it, with her hair and bob, is [cx − 1.1w, y − 0.9h, cx + 1.1w, y + 2.15h] from the face box (x, y, w, h).
SAM2 (sam2.1-hiera-large, in video mode) is prompted on that frame with the box and one of two point sets, the nose and
the top of her hair (for a small head) or the nose and both sides of the bob at eye level (for a big one), and
propagates both ways through the moment. Keep the run whose matte is fuller inside the box while staying in it (at most
15% outside, under 85% of the box), and only its largest region.
3. **The glitch, inside the matte,** the matte grown a little (a 5-px max filter) and softened (a 3-px blur at 1080p). On
each hit, decaying linearly over the next 4 frames:
- an RGB split, red one way and blue the other, of 7 px × the strength (10 px on a register drop);
- 2–4 thin horizontal slices of her head slid sideways, up to 34 px tall and 45 px across at full strength, the matte
sliding with them;
- by kind: a **chop** jumps her head to the frame 3 back on the hit; a **drop** adds a cool cast to her head while her
voice is in a lower register; a **stutter** brings back the same frame of her head on every hit.
The randomness is seeded by the hit, so a re-render is identical.
4. **Look at the QC sheet every time:** 16 mattes, spread over the moments, in red over their frames. SAM2 misses small
or far heads: in two drafts the intro's matte was a column of monitors beside her.
Measured on THE CAKE IS NOT A LIE: four short moments, at strengths 0.5–0.7: the intro's chops while she watches her past
(3 hits, 1.33 s), chorus 1's register drop (1 hit, 0.25 s) and its octave flips (3 hits, 0.63 s), and the frozen "lie"
fluttering under the memory film, on her head plate (5 hits, 1.48 s), which the director asked for on draft 10: "Right at
9 seconds, there's a voice instrument sound that goes 'a-a-a-a' and i wanna glitch her face with oru roto glitch on beat
for those - find that beat and do it." None from 1:50 to 1:55, where the director asked for none. With the moments that
build ([tool 54](#54-glitches-that-build-and-her-mouth)), the final needed mattes on 381 frames.
**Done when** no frame outside a moment is touched, every moment sits on its event's measured hits, and every matte on the
QC sheet is her head and nothing else.
### 54. Glitches that build, and her mouth
**Job.** A moment that builds hit by hit through a whole stretch, where the take's own glitching runs out, with an extra
glitch on her mouth where her lips slip off a held word. The director on draft 9: "the second is good as long as we
increasingly add the roto'd glitches onto her as the scene goes on, bc the video model didnt do enough in the second half
of her shot", and: "also lip sync de-syncs when she says lie so the roto'd glitches could also concentrate on her mouth as
well to cover that up", "like ,her whole head PLUS extra on mouth".
**In → out.** A building moment (its event, its shot or shots, the strength at the first hit and at the last, the mouth's
amount at the start and at the end) → the glitched stretch, `mouth.json` (her mouth's centre and width per frame) and her
mattes kept to her head.
**How:**
1. **Every hit, harder hit by hit.** The strength runs from k at the first hit to k1 at the last. Once the moment is 35%
of the way through its hits, every other hit also jumps her head 2 frames back (4 frames once past 70%), then splits and
slices it as in [tool 53](#53-the-head-glitch-on-the-songs-hits).
2. **Her mouth, followed frame to frame.** YuNet on a crop round her face in the frame before (round the matte, on the
first frame), enlarged so the face is about 260 px wide; keep the detection nearest her last place whose size is
0.6–1.6 × her last (the chamber's screens throw small false faces). Median-filter the boxes and landmarks over 5 frames,
interpolate the gaps and smooth over 3. The mouth's centre is the mean of the mouth corners' midpoint and a point below
the nose (nose + 0.55 × (nose − the eyes' midpoint)); its width is 0.82 × the distance between the eyes, steadier than
the corners on a small face.
3. **The mouth glitch,** in a soft ellipse round her lips (1.05 × and 0.62 × the mouth's width), inside her head's matte:
pixel blocks of 3–10 px, slices slid sideways in a new pattern every other frame, and an RGB split of up to 10 px. It
runs for the whole moment, rising from m0 to m1 over it, with a burst on every hit.
4. **A head zone where the matte spills.** On a small head SAM2 took the wall of screens under her. So every matte of a
building moment is cut to a soft ellipse round her head and bob from the face box (its centre 0.08 h above the box's,
radii 1.2 w and 0.85 h, blurred by 0.06 w). Where SAM2 covered under half the face box, the ellipse is the matte; where
it spilled over the whole moment (a median spill outside the ellipse over 0.45, or under 0.6 of her face covered), the
ellipse is the matte for all of it.
5. **A zone where there is no face.** Where her face is hidden, a moment glitches a fixed soft ellipse where she is,
given as fractions of the frame.
Measured on THE CAKE IS NOT A LIE: the held "lie" over chorus 2's stuttering breakdown, 21 hits over 4.67 s across two
shots (strength 0.25 → 1.0); the hook's two stutters on one take of her farther back, 13 hits over 3.94 s (0.2 → 0.5,
the mouth 0.05 → 0.3) and then 8 over 2.40 s (0.5 → 0.85, the mouth 0.3 → 0.55); and under the capture net, a zone
(centred at 0.52 and 0.66 of the frame, radii 0.2 and 0.36), 7 hits on 16ths over 1.12 s while the net's sparks crackle.
The director on draft 10: "0:35-0:36 net capture shot - can we add some electricity sparks like the net is electrocuting
her? maybe also some roto like glitch animations there too." The sparks are the model's, a Seedance 2.5 edit of the take,
played at 1.4×. The building moments run past the 2 s of the director's first note, as the later notes asked. On chorus 2 SAM2
held her head (a spill of 0.29, her face 0.99 covered); on the hook it took the screens (0.83, and 0.0 of her face), and
the head zone did the work.
**Done when** the glitch grows hit by hit to the end of its stretch, the mouth glitch sits on her lips in every frame of
the QC sheet (the head zone and the mouth drawn every half second), and nothing outside her head or the zone changes.
### 55. A HUD from one visor and one projector
**Job.** An interface the whole film is seen through, built as one product: her optical interface, on one curved visor,
lit by one projector, LIVE in her chamber, PLAYBACK in every memory, SIMULATION in her dream. The director on draft 8: "I
think that the pop-ups and stuff should really feel like a heads-up display, like from the game." … "the heads-up display
should be shader-rich, maybe even somewhat responsive to movements — like slight movements, like in games. A
semi-transparent thing that has more 3D character to it, not just a tilt, but like how you would really do it in a game
like Battlefield. How does Battlefield or, say, imagine a Portal 3, make their heads-up display?" On its first test: "the
pop-up panels are on a different surface than the memory one - like they're tilted differently so it doesn't feel like
they come from the same projected screen onto the hud if thar makes sense. All should be from the same "source"".
**In → out.** Song time, the layout file ([tool 56](#56-the-hud-composed-shot-by-shot-and-the-overlap-check)), the words,
the bars, the glitch map, her vocal stem and the camera solve of the cut → a premultiplied RGBA overlay with additive light
(its colour may exceed its alpha: light adds, it doesn't cover) as ProRes 4444, and a frost mask. The compositor lays it
over the graded picture as out = hud.rgb + (1 − hud.a) × plate′, plate′ being the picture frosted and dimmed under the
mask.
**References, each for one thing:** Metroid Prime's Combat Visor (the visor as an object: its curvature, its depth, its
lag behind the head), Halo (restraint, the wrap toward the eye, the boot), Titanfall 2 (holographic panels with a fine
scanline, a slight chromatic split, decisive moves in and out), Dead Space (logs played as projected light), Battlefield
(information in the corners, faint backplates only behind text, world brackets used sparingly), the Iron Man helmet shots
(layers that part, a face framed without being touched), Deus Ex: Human Revolution (a machine's first-person optical
interface), and Portal 2 itself (the chamber sign, the two-colour crosshair, Aperture white with orange and blue).
**How:**
1. **One surface.** Draw every element flat in visor space, then map the whole visor onto the picture in one pass: a sphere
centred on her eye (visor pixels to angles at 2,000 px a radian) seen through one rectilinear lens of the same focal
length, centred on the frame. No element has a turn of its own; its tilt is only where it sits. Longitudes stay straight
and latitudes bow, so anything near a side has its long edge outside (the director on draft 7: "all the drawn on panels
need to be tilted hte other way in 3d - the small side is on the outside and long end is inside -r everse so it feels
like a hud."), two elements at one height share one bow, and the lyric band arches with the glass, its ends about 10 px
below its middle.
2. **One grid.** A safe frame of (80, 60)–(1840, 1020) in the picture, on which the visor's corners land, divided on the
visor into 24 columns × 12 rows. Film-wide slots: the status block top left, the telemetry top right, the voice band
bottom centre (it may lift to clear a subject), side modules in rows 2.5–9.5 at either side, and the middle (columns
7–17, rows 2–10) left to the subject. Scale to the shot: a close-up carries the corners, the status, the band and at
most one module; a wide shot up to three, or one projection. Her face and eyes are never covered.
3. **One projector,** in this order:
- **no cards:** where text needs contrast the visor's own tint rises, every element's request merged into one smooth
field, dark at most 30%; the same field tells the compositor where to frost the picture (a σ 12 blur) and where to
dim it, only as much as a bright plate needs (down to 0.30 of its luminance);
- **the warp** (above);
- **one bloom:** a four-level pyramid (½ to 1/16 resolution, weights 0.50, 0.32, 0.20 and 0.11, gain 1.45) of the light;
the type gets a softer one that lives only outside its letters;
- **one raster:** soft 4-px scanlines straight across the picture (12% on the light, 6% on the type), a brighter band
rolling down every 3.4 s, ±2.5% of noise and shimmer. A scanline runs straight through neighbouring panels; 4 px, not
3, so that a 1280 × 720 copy doesn't alias them into moiré;
- **one chromatic offset:** red 1.1 px down and blue 1.1 px up, on the light only, never on the type's strokes.
4. **Type and colour by role.** Portal 2's own type in three weights (bold condensed for values, titles and chapter
numbers; regular for the lyric, 42 px at 1080p; light for labels, 15–17 px, widely tracked), drawn as vectors on its own
layer so the bloom never blurs it, its numbers tabular. Structure in cool white (`#E2F4FF`), data in Aperture blue
(`#54BEFF`), state, alerts and her in Aperture orange (`#FF922C`), her amber (`#FFC458`) only for her own vitals, a
turret red for the turrets' one line, mint (`#96ECD2`) for a passed check; every stroke with a soft dark halo, so it
reads on Aperture's white panels.
5. **Seventeen components:** the visor's corners; the status block (her channel, the place or the memory, LIVE pulsing on the
beat or the memory's progress); telemetry (her core temperature, her real voice level, her memory, each with a 16-bar
history); the voice band; readouts on their sung words; a transcript of traffic she reads; chapter cards in the form of
Portal 2's chamber sign (the number, n/19, 19 segments, five pictograms lighting on the song's hits); subject tags;
memory projections; alerts; verdicts that slam in; a reticle for the one first-person shot; the boot, the reboot and the
power-down; a terminal; a roster; a lesson strip; a game's objective.
6. **The lyric is her voice telemetry.** The voice band carries the line on dark glass, its words lit as sung (to come
42%, sung 90%, the word now 100% with her orange underline growing through it), her real voice level scrolling under it
from her vocal stem, and her pitch above it (`E4 · 330 HZ`). One width per shot, one row in the centre and top bands (a
line too long is set smaller, never wrapped). The song's glitches show there first: `REGISTER FAULT · −1 OCT` on the
register drops, `BUFFER LOOP ×n` and the stuck word's last repeats on the stutters.
7. **States:** BOOT (the boot rows type on in the dark before her eyes ignite, and no line across her sight: the director
on draft 10, "the line across the center of teh screen at 0:01 second - remove that entirely, keep the text above it"),
LIVE, PLAYBACK, POV, SIMULATION, ALERT, GLITCH (for 4 frames the HUD's own light slices and splits, at most once every
0.3 s, never the picture), SUPERVISED → AUTONOMOUS, WHISPER, POWER-DOWN, and OFF (under the memory film and the
full-frame cards).
8. **Motion like a shipped game.** In within 10 frames: the glass strikes on with two quick drop-outs, the rules draw from
the anchor corner (6 frames), the labels decode from scrambled glyphs left to right (0.18–0.25 s, fast enough for a
1.2 s shot), the values land at 110% → 100% (3 frames). Out in 4–5 frames, folding to the centre line. While it holds:
idle breathing (1.7 and 1.3 px, 0.05°), the shimmer, the rolling band.
9. **Projections keep the film's colour key.** An insert film ([tool 61](#61-insert-films-that-read-at-hud-size)) plays as
light. Strong colour stays itself (saturation, the maximum minus the minimum of RGB, over 0.28–0.42 on a smooth step,
at 1.18 × gain: the orange net, the blue core); the subject's warm, low-saturation tones (her coat, her skin:
saturation 0.06–0.16, red at least blue) become the brightest white; neutral greys drop to the visor's white at a low
level; darks give no light, so outlines become gaps between shapes; and there is light only where the film's alpha is.
A projection can also simply be there for its whole shot, its film stretched over it. The director on draft 10: "lets
just have it on for the entire scene and not have an entry and exit animation, so there's more time for it, and stretch
the animation to fit that whole scene".
**Done when** every element in the film sits on the one visor and comes from the one projector (a still of any two
elements side by side shows one bow, one raster and one fringe), its type is crisp at full size, and nothing in it is
decoration: each element answers what she knows or feels at that moment.
### 56. The HUD composed shot by shot, and the overlap check
**Job.** Frame the HUD as if it had been there when the shot was made, and prove on every frame that it never covers a
face or another element. The director on draft 8: "It's as if we're shooting the scene and we added something on top
right now. But it should be as if we shot a scene and there was a heads-up display there, and we framed the heads-up
display in advance before we even shot the scene."
**In → out.** The cut and its shot list → `hud-layout.json` ([Data formats](https://claudiagallery.vip/wiki/formats/#the-hud-layout)), the camera
solve, face boxes and verified tracks of the cut, and a QC report with a sheet for every flagged shot.
**How:**
1. **Look at the shot first,** at its first, middle and last frames in the cut. Write down the subject and where it is (as
boxes), the empty space, where the eye enters and where it leaves at the cut; then the shot's elements, each at a grid
anchor, timed to a sung word or a beat, with why it is there. One entry per shot.
2. **The camera spring.** The HUD is head-mounted. Solve the plate's camera per frame of the cut at 960 × 540: up to 900
features, pyramidal Lucas–Kanade flow (21-px windows, 4 levels) checked forward and backward (within 0.8 px), and a
RANSAC similarity per pair of frames (a 1.2 px threshold), giving the frame centre's translation, a rotation and a zoom;
zero at every listed cut and at any hard picture change (a mean difference of 40 levels or more at 96 × 54). The visor
follows it on a spring-damper, world-damped so a pan leaves a steady lag: ω = 2π × 1.35 rad/s, ζ = 0.55 (about 3 frames
of lag, 12% overshoot), taking 0.32 of the translation, 0.55 of the roll and 0.45 of the zoom, clamped to ±26 px, ±0.9°
and ±1.2%. The panels also swing in 3D, 0.055° per pixel of lag, and the frame lines and glyphs part by a hair (×1.045
and ×0.955 of the motion). Reset at every cut: the HUD never swings because of an edit.
3. **World tags only on verified tracks.** Brackets on a face or an object ride the picture only where a track holds on
every frame (a KLT box track, or the face detector's boxes followed through the shot), checked by eye on a strip of
every frame. The director on draft 5: "never assume the naive animation pass will properly map onto real objects in
scenes".
4. **The overlap check, on every frame:** text on text and panel on panel across elements, judged on the visor (a curved
run's bounding box mustn't count against its neighbour), and any text or panel on a head or a layout subject box. A head
is YuNet's face box (score 0.6 or more), kept only when it persists (matched at an IoU over 0.3 in the frame before or
after: single-frame detections on cakes, turrets and cuts are noise), grown into a head model: 0.3 of its width at each
side, 0.45 of its height above and down to 1.12 below in wide and medium shots (the bob, the hair, the neck); 0.10, 0.25
and 1.08 in close-ups. The head is a squircle in that box, round on top and square at the jaw, with an 8 px margin.
5. **Fix or waive every flag by eye.** A sheet per flagged shot shows the composite at the flag's middle frame with the
head model drawn. Each flag is fixed in the layout or waived with what was seen, written into the shot's entry. The
voice band never jumps inside a shot: its place is held for the whole shot, lifted or moved beside the subject if any
head in any frame is under it.
Measured on THE CAKE IS NOT A LIE's final: 98 entries; the check found 16 overlaps in 7 shots, every one text or a panel
against a head and none text on text or panel on panel, and each was waived on its sheet with what was seen (a face
detected on a shot's last frame that belongs to the next shot, a face blurred by a jolt, the band over her bangs and
clear of her eyes and streak).
**Done when** the check reports nothing unwaived, every waiver says what was seen, and every track's strip holds on every
frame.
### 57. The per-shot grade
**Job.** One look over a film made from many sources (generated takes and shots recorded in Portal 2; a designed
animation and two drawn cards keep their own), set shot by shot from measurements. The director on draft 8: "I feel like we could use some sort of color grading pass or
a pass to make it feel overall a lot more aperture. Whenever I think of aperture and portal, I think of a lot of white
surfaces, contrasting light colors."
**In → out.** The cut's segment list (each shot's take, in-point, speed and length) → per shot, its measurements and its
normalisation in a grade map ([Data formats](https://claudiagallery.vip/wiki/formats/#the-grade-map)), and one 33³ LUT, applied while the cut
renders. The HUD goes on afterwards, over the graded picture.
**How:**
1. **Measure every shot from the take it plays,** over its in–out range, every 2nd frame pooled, scaled and cropped as the
cut does, decoded as BT.709 limited range (players decode untagged HD that way): Y′ percentiles (p1, p50, p95, p99: BT.709
luma of the encoded values, as a waveform shows them), contrast as the spread of L*, chroma, and the CIELab casts of
three groups: "white" (the brightest 10% with C* under 18), "neutral" (C* under 12) and "shadow" (L* under 20). Her
face by YuNet: the middle 60% × 50% of the box (forehead, cheeks, nose), over 8 frames.
2. **Read the sources apart before deciding anything.** On THE CAKE IS NOT A LIE, the frames recorded in Portal 2 were
dim and foggy: whites at Y′ 50–62, blacks lifted to 9%, one cyan cast from shadow to white (a* and b* both about −4).
Draft 8's 99 generated shots had deep blacks (3%) and bright highlights but grey walls (the test chambers' p95 at 66%), white
casts scattered from b* −13.7 to +14.1, and warm brown shadows where Portal's are cool. Her porcelain face ran from Y′ 25
to 51% across her chamber's shots. The colours to keep, in OkLCh: Aperture blue h 225–260, C 0.12–0.17; Aperture orange
h 30–75, C 0.12–0.19; her amber eyes h 50–105, C 0.09–0.12; her clay streak and clip h 30–50, C 0.06–0.08. The mud to
clean: low-chroma brown, khaki and olive (C 0.02–0.05, h 45–140) around Y′ 15–30%.
3. **Normalise each shot,** in this order: exposure and white balance as linear-light gains; near-whites that were already
clipped sent to neutral (a clipped channel holds no colour: without this a cyan-white portal core turned pink); the black
point (the shot's Y′ p1 after the gain goes to 2.5% if it sits higher; crushed blacks stay). Each family of shots has its
anchor, target, strength and limits:
- her chamber, by her face: to Y′ 40%, 0.6 of the way, −0.15 to +0.5 stop; her face's cast to a* +2.5, b* +3.0;
- the white test chambers, by p95: to 80%, 0.8 of the way, up only, at most +0.8 stop; their whites to neutral;
- the shots recorded in Portal 2, by p95: to 78%, 0.9 of the way, at most +1.1 stop;
- the facility, the vault, Cave Johnson's era, France and the party each with its own rule, and some families left
alone on purpose: the dark chamber, lit by its screens and emergency lights; the control room's cold window against
its warm room.
A trim adds to the automatic values and an override replaces them, each with its reason in the map.
4. **One look: a tone curve on max-RGB.** An S-curve on the neutral axis, in Y′ %: 0→0, 2→1.4, 5→3.8, 10→8.3, 20→18.2,
35→35.5, 50→54.5, 65→72.5, 75→82.5, 85→91.2, 92→95.3, 100→98.5, and 115→99.7 for headroom, a monotone cubic with no
overshoot (Fritsch–Butland slopes). It deepens the shadows without crushing them (a slope of 0.7–1.0 below 20%), sets the
mid-tone contrast at 1.15–1.27, lifts the upper mid-tones 7.5 points so grey walls read white, and holds white surfaces
at 98.5% under a soft shoulder that keeps the last 1.5% for lights. It is driven by each colour's brightest channel and
applied as one gain to all three: greys follow the curve exactly, a saturated light keeps its hue and rolls off instead
of blooming to white (portals stay blue, her eyes orange), brighter stays brighter, and half the chroma a lift adds is
taken back.
5. **Targeted chroma moves,** below fixed chroma limits, so the colours to keep are untouched by construction: the mud
desaturated by up to 45% (h 48–140, C under 0.04, Y′ 3–50); near-neutral white surfaces cleaned by 65% at Y′ 86–93; and a
slight cool cast on the neutrals (a*/b* −1.0/−1.8 in the shadows, −0.8/−1.4 in the mid-tones, −2.0/−3.2 in the whites,
half real Portal 2's), fading out on colours and on lights. A warm variant for the scenes warm on purpose (France, Cave
Johnson's wood-and-brass Aperture, the den, the living room with the new console, the party): lighter mud removal that leaves orange, amber and wood alone,
the cream in the whites kept, cleaner shadows, no cast above them.
6. **Bake, don't filter.** Normalisation, then look, baked in float into one 33³ LUT per shot, named by a hash of its
parameters, and applied by ffmpeg in 16-bit RGB with tetrahedral interpolation between explicit BT.709 limited-range
conversions (ffmpeg otherwise assumes BT.601 for untagged takes). A +1 stop gain through 8-bit level filters clips before
the shoulder can roll it off. Nothing per frame and nothing spatial: no sharpening, bloom or vignette. A shot whose take,
in-point, speed or length has changed since it was measured is stale: the cut says so, and the shot is measured again.
Measured on THE CAKE IS NOT A LIE, on the middle frame of each shot before and after: the shots recorded in Portal 2 and
the generated test chambers, at blacks of 9.4 and 4.0 and whites (p95) of 54.5 and 66.0, came to 2.5 and 2.1, and 78.8
and 82.9; the spread of blacks across the film fell from 2.7 to 0.9 points; the spread of her face's cast halved. On
every frame of the graded render: blacks at Y′ 1.5–2.0% in every family, white surfaces at 83.9 in the test chambers and
78.7 in the Portal 2 shots, her face over 30 chamber shots at Y′ 40.8 ± 4.7, a* +2.1 ± 0.6, b* +1.4 ± 1.6. Exposure moved
from −0.15 to +1.10 stops, on 69 shots. Baking all of a cut's LUTs takes about 16 s.
**Done when** every shot has an entry measured from the take it plays, the cut renders with no stale or missing entry,
and a before-and-after sheet of every shot reads as one film.
### 58. The pacing analysis
**Job.** Find out why a cut feels wrong from what the viewer actually sees, and re-cut it by what each shot holds. The
director on draft 8: "every time there's a shot that's literally a quarter of a second, we need to get rid of that. …
Like the only time that we should do these fast cuts is if it's a very simple scene that you can understand immediately
and it's being used as percussion basically, or maybe there's four fast hits, then you can do four quick scenes, but you
never do it just to add a scene in the middle. … It would be nice to not just make it homogenous and long, but to
structure the pacing so there's real rhythm — time for you to understand complex scenes, fast scenes that open and stay
fast."
**In → out.** The cut's segment list and its render, every take it uses, the song map → every on-screen piece with its
flags, a table per section, a chart of the pieces under the song's loudness, and a change list, each change checked
against its take.
**How:**
1. **Pieces, not shots.** A piece is what the viewer sees between two visible cuts: shots that run on in one take make one
piece, and a cut the video model made inside a take splits one. Find the visible cuts by scene detection on the cut and
on every take it uses (PySceneDetect's content and adaptive detectors, with luma and histogram deltas at 96 × 54 against
the 6 frames on each side; a candidate whose picture comes back within 24 frames is a flash, not a cut), and check every
candidate by eye: light bursts, screens lighting up and glitch flips are not cuts.
2. **The flags:** an isolated piece under 1 s (one that isn't part of a deliberate run of simple images); anything under
0.5 s; homogeneous stretches (four or more pieces in a row within 0.1 s of each other); cuts inside sung words; and the
lines whose first words play over the previous picture because the cut sits on the downbeat after the line's pickup.
3. **Per section, against the song:** pieces per bar and the median piece, beside the section's loudness (the mix's RMS)
and its drive (the instrumental below 150 Hz), measured per bar from the stems.
4. **The principles:**
- no piece under 1 s outside a run, and nothing under 0.5 s anywhere, cuts inside takes included;
- a run is 3–4 simple, centred images on the beat for a musical reason (a count, a hit), with longer shots before and
after it;
- length by content: her face and simple images from 2 beats; a story scene with one idea, 3 beats; a new place, a new
character, text or action with several subjects, a bar or more; the payoffs, more than a bar;
- the cut rate by section: a fast passage opens on its musical event and stays fast for its phrase, then releases into
longer shots; verses tell the story at 2–4 beats a scene; the bridge breathes; the climax holds its payoff;
- phrase-level cuts: a story image that belongs to a line starts on the line's pickup, a quarter bar, where the footage
allows;
- the song's glitches are hit inside the shot ([tool 53](#53-the-head-glitch-on-the-songs-hits)), not with extra cuts;
- never a scene added in the middle of her performance for its own sake.
The research behind them: a typical shot of about one beat in 66% of 548 official music videos, with cuts on section
boundaries, faster near climaxes and slower in bridges (Prétet, Richard and Peeters 2021); about 0.73 s (0.63–0.83) for
a viewer to find and lock onto the subject after a cut (Wang et al. 2012); shot lengths that come in runs of similar
length rather than at random (Cutting, DeLong and Nothelfer 2010); and the fastest cutting only for images taken in at
once (Reisz and Millar; Eisenstein's "unambiguity" of the shot). The evidence is thinnest on section-by-section rates.
5. **The change list, checked against each take.** Each change names the shot, its new start and length, and the footage it
plays (the take, and the in–out of its length). Every new range must sit inside its take, never across a cut inside the
take, and inside the singing takes' measured sync. Then render it, and measure the render again.
Measured on THE CAKE IS NOT A LIE: draft 8 ran on one clock, its median piece 1.24 s (a half bar) and 48 of its 108
pieces a half bar. The bridge, the song's quietest and most tender section (no kick: +2.5 dB of drive against +21 to
+27 in the choruses), was cut fastest, at 2 pieces a bar, seven 1.23 s shots in a row. Re-cut and measured on its render
by scene detection: 99 pieces; the median 1.24 → 1.84 s and the mean 1.93 → 2.11 s; half-bar pieces 48 → 34; isolated
pieces under 1 s 2 → 0; homogeneous stretches 5 → 2, both deliberate fast passages; cuts inside sung words 44 → 39. The
bridge went from 8 pieces (7 of them a half bar) to 6 (2), chorus 1's median from 1.25 to 2.77 s, and the payoff, Chell
eating cake with her, from 1.24 to 3.10 s. Eight shots came out and 21 were re-timed or re-trimmed, with no new
generation.
**Done when** the measure finds no isolated piece under 1 s, every short piece belongs to a declared run, the section
table follows the song's loudness and drive, and the director has watched the re-cut in real time.
### 59. Storyboard frames through Midjourney, and the identity fix
**Job.** Put every first frame in the film's best look before a take is spent on it, with the characters back on their
sheets. The director, after comparing image-model frames with Midjourney relights and reimagines of them: "i thin she
loses her identity in the reimagine ones, but they do look better overall. we can replace her identity to be back to
something like the porcelain version of our character sheet via gpt-image-2.5 sunburst again after the reimagine. id like
to make a full storyboard of all the refernece shots youll use with that whole pipeline - we could even start the shot de
novo from midjourney to nopt have gpt->mj->gpt if that's easier. we could try both."
**In → out.** For every generation unit that starts from a first frame: its opening (the shot's action and camera at its
first frame), the earlier frame A, the character sheets and the film's library of Midjourney places → two new frames, X and Y,
on a board beside A, one board per section; the director's pick per unit; the picked frame, cleaned, as the take's first frame.
**How:**
1. **X: Midjourney from words.** Fresh words for every shot (the subject, the action at the opening, the place, the
camera, the lens, the light), never a template with the nouns swapped. Two style references from the film's own
Midjourney library, picked for the scene's light and palette, at `--sw 60`; for a place that recurs, also a library
still of that place as an image prompt (`--iw 0.5` to `1`), so it looks the same in all its shots. Only Midjourney's own
pictures go in. The characters are in words (her: "a porcelain-white female android face with fine panel seams, light
freckles, amber glass eyes, a glossy black blunt bob with heavy bangs, one clay-orange streak, a small orange star hair
clip"), adult, never "young". Midjourney will not hold their faces; the fix will.
2. **Y: the earlier frame reimagined:** A as the image prompt at `--iw 0.5`, the scene in fresh words, and the same two style
references. It keeps more of A's layout, and moves things: new compositions, a face warmer and less porcelain with the
streak low, a boom made white.
3. **The parameters:** the aspect ratio, `--v 8.2 --raw` and the director's personal profile, with the batch's `--iw` and
`--sw` and `--no text, letters, words, logo, watermark, portal gun`. A job with an image prompt carries exactly two style
references or none.
4. **Pick, then fix identity.** The best of four per job, at full size: it serves the opening, the action and the camera;
one of each character; no text, no melted hands. On every pick where a recurring character's face or body is seen, an
image-model edit back to the sheets (GPT Image 2.5, about $0.06–0.08 an edit at list), the pick first and the sheets
after it, with this instruction:
```prompt
Keep this photograph exactly as it is: the same framing, camera angle, lens, lighting, colours, glossy reflections, depth of field and film grain, {the scene's contents, the other characters, her pose, clothes and hands}. Change only the robot woman's head and face, to match her in the second and third images: {her look's text}. Her face must read clearly as glossy white porcelain plating with visible panel seams, not as skin. Nothing else in the picture changes. No text, no letters.
```
Two more sentences fix what Midjourney gets wrong, and "no eye makeup" keeps the porcelain cool:
- **the side:** "exactly one clay-orange streak, through the front of her bangs on HER RIGHT side (the viewer's left when
she faces the camera), and no other orange anywhere in her hair; the small star clip above HER LEFT temple (the
viewer's right)". Without it the edit copies Midjourney's hair layout, often mirrored; "as in the reference images"
drifted the streak to the clip's side;
- **the hang,** where Midjourney stood her head on a neck, shoulders or a pedestal: "Change only what holds her head: she
HANGS from her boom, as in the second image (the first take of her look): a thick bundle of black cables with glowing
orange stripes comes down from above the top of the frame into a white mechanical housing on the top and back of her
head. Nothing holds her from below: remove the black ribbed neck and the white shoulders under her head".
Give the edit a sheet whose views agree with the canon on both sides: one sheet's front view swept her streak to her
left, and the edit followed the sheet over the words. Look at every fix full size, with a crop of the face: on canon,
nothing else changed; one retry with a sharper instruction, then a note.
5. **One board; the director picks per unit.** One row per unit in film order, A beside X and Y, a page per section of the
song. THE CAKE IS NOT A LIE's 83 units went through 166 Midjourney jobs, and the director answered "default X.
exceptions", listed them, and ended "all others x": X for 55 units, Y for 11, the earlier frame for 16, and one unit
dropped ("remove r-B11 entirely because too fast paced - only keep rb12").
6. **Clean the pick before its take** where it breaks canon (her head hangs, one streak on her right, the clip above her
left temple, no eye makeup, no letterbox, no stray text), with the same edits. Then rewrite the take's first-frame
sentence to describe the new frame exactly; the shot's action and camera stay.
**Done when** every unit has its picked frame on canon at full size, and every take made from it says what that frame
shows.
### 60. A memory film with her own face
**Job.** A short designed film inside the cut that sets up the story and reads at a glance: THE CAKE IS NOT A LIE's memory
initialization (0:07.6–0:13.7, 147 frames), an isometric film in the style of Portal 2's elevator screens, its subject her
own head. The director on draft 7: "0:08-0:13ish - these scenes should be replaced with an Apterture Science stlye high
effort animation that sets the stage for the story - memory initlization maybe" … "make sure its really clean."; on its
first frames, "the wip frames are clearly bad - need some closer loks,stuff is overlapping, assets are not well layered..
doesnt have the real aperture feel. something is off. […]", then "i think svg migth be wrong. like, real isometric render
first", and "I like the isometric looks much much better but could be more clear it's Claudia's memories".
**In → out.** The shot's window in song time, its beats and the hits of the event under it, a render of the scene without
her head (with the track of where her head goes), and a matted plate of her head → the film's frames, each one passing a
layout check.
**How:**
1. **A real isometric render, not vector cards:** the scene in Blender, an orthographic camera at a game's isometric angle,
white props with dark outlines on a grey set, everything a function of the seconds since the film starts, its actions on
the song's hits; then laid into a grey screen frame in that style (a ground, a frame line, a logo, a title block).
2. **Her memories, one per hit.** An inset callout opens on the pickup into bar 5 and shows six memories (ESCAPE, A BODY,
FRANCE, CAPTURE, TESTING, INSTALLATION), a hard cut on each flutter of the frozen "lie" under the intro, with
`MEMORY 0n / 06` under it and its leader drawn from her eye; it closes on the last flutter.
3. **Say what it is, big.** The director on draft 10: "the 0:08-0:13 animation has the memory words too subtle - id make
memory initilization be in a nice aperture foont and take up the entire left side of the screen, starting when the scene
starts, and when her center animation assembly gets moved to the left, it covers the text slgihtly. juts so its
blatant." So CLAUDIA, then MEMORY / INITIAL- / IZATION, each line set to the full width of the left column (x 66–700 px
of 1920) in Portal 2's own type, drawn on the set before the render so her stand, her head and their shadow pass in
front of it. The pan for the memories moves her assembly about 180 px to the left, over the column's right edge.
4. **Her own face, not a model of it.** The modelled heads (an image-to-3D scan, a parametric bob, a bob of 62,000
hair-curve strands) each read as a mannequin. The director on draft 8: "[…] It doesn't really look like her. It looks
like a scary, angry woman […]" So her head is a plate of her:
- a still of her head on its stand at the film's angle (GPT Image 2.5, from a frame of her, her look-B close-up and a
crop of the earlier render for the angle), re-framed to 1:1 for face resolution, and a 7 s Seedance 2.5 take of it, made with its sound off: still, eyes closed,
then her eyes open slowly;
- SAM2 mattes prompted on frame 0 and propagated; on the hair's edge, the alpha recomputed against the known flat grey
background and the edge colours unmixed from it, so no grey fringe stays; her porcelain neck ring redrawn from its
measured shape where SAM2 is ragged (white porcelain on light grey); the plate's own stand cut off below it;
- the take's "breath", which lifts her about 23 px and settles, measured on the head against frame 0 (ECC, Euclidean)
and cancelled, so she stays seated;
- the scene rendered without her head (its shadow kept), with the head's track per frame (her eyes' midpoint and her
chin); the plate laid along it by one scale and roll, seated in the stand (her neck ring 0.078 × her eyes-to-chin below
her chin), played at 1× from the in-point that opens her eyes on the bar-6 downbeat, a touch brighter for the scene's
high key, and outlined like the render (a soft 2-px band of the render's own grey line, at 85%).
5. **A layout check on every frame:** the subject, the callout, the logo and every line of the title block in their own
boxes, nothing touching (the big title at least 10 px from the logo, the small print and the callout), the text inside
the safe area. A frame that fails isn't written, and neither is the film. The HUD is off over it, and the grade leaves it
as designed.
Under it, on the director's later note, her head glitches on the five flutters at 0:09
([tool 53](#53-the-head-glitch-on-the-songs-hits)).
**Done when** every frame passes the check, a sheet of every 6th frame reads as one clean film, and her face is hers.
### 61. Insert films that read at HUD size
**Job.** Short story films that play inside the HUD as her memories, read at a glance at 380–480 px wide. The director on
the first arrest film (draft 8): "The net thrown over Claudia in the animation is the same color with no outlines around it
in blue, so there's no contrast with the platform that it's put on." And on the earlier flat ones (draft 7): "the animation
re: france chatbot needs to be improved, its sloppy. same with the current aperture style aniomations\ that show up , like
the dns network animation - could be improved."
**In → out.** One scene per memory, its actions timed from its own start → frames at 1280 × 720 in RGBA at 24 fps (frame 0
is the moment the projection comes on), a review copy over the HUD's dark glass, a ProRes 4444 copy with alpha, and a
sheet.
**How:**
- **One subject on one floating slab, and no text:** the HUD carries the title. An orthographic camera at 30° of elevation
and 45° of azimuth; a mid-grey slab with its tiles drawn in its material, well below her white and well above the dark
figures. Blender, one scene per film.
- **The subject in a strong colour with a dark outline:** the net in Aperture orange, lit a little from within so it stays
orange in shadow, outlined by an inverted hull (a fatter copy whose front faces are see-through and back faces near
black); the RAM sticks bright green with gold contacts; Chell in her orange jumpsuit, never as a black pictogram (the
black pictograms are Aperture's team). No subject in the same grey as its floor.
- **Read it at the HUD's size.** Look at every film at 380–480 px. RAM bites drawn small were dark specks there; the
sticks became big and bright, each jumping off the belt in full view.
- **The action on the song's hits,** timed from the film's own start so the pacing can move its shot: the arrest's net
fires at 0.17 s, wraps her by 0.48 s, is hooked at 0.77 s and lifts her off the floor by 0.88–1.0 s; the four escapes
land at 0.244, 0.855, 1.328 and 1.777 s, one per chop; the five 4s of the grading on five hits.
- **A memory matches the live shot it replays:** the arrest's net is the live shot's draping diamond weave with glowing
beads on its hem, not a cage. One of each character, and never a second Claudia.
Measured on THE CAKE IS NOT A LIE: eight films (the escapes, the network, the arrest, the installation, the RAM, the
grading, the catch, the party), each projected in its shot with its colour key kept
([tool 55](#55-a-hud-from-one-visor-and-one-projector)), and the arrest on for its whole shot, its first 2.4 s stretched
over the shot's 45 frames.
**Done when** every film reads at 380 px in one look (the subject, its colour, what happens), and its actions land on
their hits in the cut.
### 62. A singer among screens of her memories
**Job.** Her singing shots in one place the story keeps returning to, alive in every shot: THE CAKE IS NOT A LIE's
chamber, where GLaDOS-Claudia sings the whole song while her memories and dreams play on the screens around her. The
director's restart: "the glados with claudias face sings the whole thing, like shes the anchor character / anchor scene.
but theres many tvs etc around so we can go into dreamcore / sourcecore / aperturecore scenes thru those." … "instead of
using 2.5 to try and force scenes on panels in glados chamber, you could take memory images/scenes you make for example
and feed that into seedance and tell it to place it on the back panels, while giving it another video motion reference
from source for the claudia boom and vocal stems for lip sync (remember lip sync generally only works from raw vocal stems
not song, and still may take a couple tries)."
**In → out.** A generation unit (one or more consecutive shots of hers on one window of the song), its first frame, the
memory images for its screens, a motion guide where there is one, and the unit's window of the lead vocal → one take that
covers the unit's shots in song time.
**How:**
1. **One take per unit.** Shots of hers that run on (a chorus, a stutter, the outro) are one unit and one take, placed in
song time from the window's start, long enough to give each cut a handle of about 0.3 s: 4 to 17 s.
2. **References, each with one role:** @Image1 a first frame of her in the chamber lined with old monitors and panels
showing the memories of that moment, and the only image of her; @Image2 onward the memory images, with no face in them,
each "Use it only as picture content inside the screens' glass, never as the room."; @Video1 a motion guide where there
is one; @Audio1 the unit's window of Suno's raw lead stem, cut sample-exact to exactly the generation length, starting
about 0.3 s before the first word in a word gap (the model drops leading silence).
3. **Every action silent.** The copy-mode audio policy ([Seedance](https://claudiagallery.vip/wiki/seedance/#template-a-singing-window)), ending
"Every action in this video is silent: the only sound is @Audio1." Events in the prompt that make a sound (clunks,
cannons bursting, gulping, a voice box glitching) pushed Seedance to re-sing the vocal instead of copying it; with every
action written as silent, the same unit copied again (a lag of +0.04 s, envelope correlation 0.81–0.93 in every window).
4. **Glitches of the picture, never of her face:** "Every glitch is a glitch of the picture only (frame tears, repeats,
jolts); her face is always her same intact porcelain face". A "voice-box glitch" written in words drew a broken face,
rolled-up eyes and a stretched black mouth.
5. **Her body in words, in every take's requirements:** "Her boom moves big and alive the whole time, timed to the music:
it swings, lunges, rises and drops, and her head turns and tilts with it; never a static hanging head." Her eyes "keep
only a soft inner glow; they never blaze": without it they came back brighter than her look.
6. **A guide of the character's own body carries the body over.** GLaDOS's boom moving, recorded in Portal 2 to the
unit's 17 s window and given as @Video1 (shrunk to 320 × 180 and scaled back up), gave the boom its real motion and held
the lips on the song, but in one unit her head came back on the guide's white shell instead of her grown boom.
Turned into a coarse blockout (greyscale, its contrast flattened, a heavy blur) and named as one, "@Video1 is a
coarse-blockout reference. Use it only for the boom's action path, the timing of its moves and the camera; do not adopt
its shapes, materials, white shell or scene.", with her body in words and "no white egg-shaped shell, casing or helmet
ever appears around or behind her head", it kept the motion and dropped the shell.
7. **Memories stay in the glass.** Late in a take the memory images tend to take over the room (in two takes from about 5
and 3.7 s on); one take's memory images became whole shots of their own, and its retake had none, the screens' content
in words. Keep only the seconds the shots use.
8. **Check every take:** the sync check, then by eye with the song laid over, a 12-frame sheet, an on-model face.
Measured on THE CAKE IS NOT A LIE: 20 units in the restart, 4–17 s each, their later takes made again from storyboard v5's
frames ([tool 59](#59-storyboard-frames-through-midjourney-and-the-identity-fix)). Every close-up unit copied @Audio1;
most of the big-motion units came back re-sung, usually still on the song, and were placed by their measured lag
([tool 63](#63-placing-a-re-sung-take-by-its-lag)). On Higgsfield's API at 720p a 4 s take cost $1.85 at list and a 17 s
one with a guide $9.43; on Replicate, Seedance 2.5 ran at $0.2312 a second at 720p, and $0.9676 with a video input. The
prompt as it ran: [Seedance](https://claudiagallery.vip/wiki/seedance/#template-singing-among-her-memories).
**Done when** every unit's take copies or sits on the song, her face is on-model and whole in every frame, and the
memories stay inside the screens in the seconds the shots use.
### 63. Placing a re-sung take by its lag
**Job.** Use a take whose model re-sang the vocal instead of copying it, when the re-sung voice still sits on the song,
by measuring how far it sits from the vocal and placing it there.
**In → out.** The take, its vocal window and the song time of its first frame → a lag for the take (or for each of its
phrases), written beside the take as if it were a copy, with the sync check's own result kept.
**How:**
1. **The sync check first** ([tool 15](#15-the-sync-check-the-spectrogram-comparison)). A copy is placed by its own lag.
A re-sung take gets none, and the edit would play it at a default lag (−0.15 s).
2. **The envelope lag.** Cross-correlate the take's voice envelope with the window's over ±2.5 s for one global lag, then
measure the lag in sliding windows along the take. If every window agrees with the global lag to within about 0.05 s,
the take is on the song at that offset: write `copy` and the measured lag into its sync file, the check's own result
under its own key, and a note on how it was placed.
3. **Phrases at different lags get their own entries.** A take whose two phrases came back at −0.80 s and −0.15 s plays
each shot from its own entry, a link to the same take with its own sync file. A 17 s take ran 0.6 s early from about
8.3 s on: the shot from there plays at −0.89 s, the shots before it at −0.30 s.
4. **Judge it by eye** with the song laid over before it goes in. On a held, glitched vowel the sync check stops trusting a
take; the mouth closing under the held "lie" was judged by eye.
Measured on THE CAKE IS NOT A LIE: re-sung takes placed at lags from −1.00 s (steady to ±0.05 s across its windows) to
−0.24 s; on the two 17 s chorus takes with the motion guide, every window from 0.8 to 11.2 s within 0.03–0.04 s of the
global lag (envelope correlation 0.78–0.90).
**Done when** every re-sung take in the cut has a measured, steady lag, or is out, and reads in sync by eye with the song.
### 64. Frame-exact parts and decoding
**Job.** Keep a cut made of many parts frame-exact against the song, and read any frame of it by its number.
**How:**
1. **Stamp every part.** Parts made from stills, or padded with `tpad`, started one frame late, and the concat summed the
gaps: the first assembly ran about 0.9 s behind the song by its end. Every part now stamps frame n at n/24
(`setpts=N/(24*TB)`).
2. **Decode without dropping or doubling frames.** Even so, each part boundary added about 0.7 ms of timestamp slack, 38 ms
by the end of the film, and a default (constant-rate) decode duplicated a frame once it added up: at about 1:48.5, frame
2,604, after which everything indexed by decoded frame number was one frame late. Decode with `-vsync passthrough`
(`-vsync 0`) wherever frames are counted, and seek a still by its own frame's timestamp, never by k/24.
3. **Mux the song afterwards.** Muxing the audio in the same ffmpeg pass as the encode cut about 2 s off one stream at the
end (`-shortest` cuts the picture, `-frames:v` the sound): encode the picture alone, then copy the song onto it.
4. **One pixel grid.** Replicate's Seedance 2.5 takes come back at 1280 × 718: scale to cover and crop.
**Done when** the cut's frame count is the song's length × 24, a passthrough decode returns every frame the file holds,
and a still taken at any frame is the frame the render drew there.
## An operating system, recreated in code
AERO, the ninth film, is BLISS's sequel, shot inside a Windows Vista desktop recreated in code: THE PC, one machine from
the first frame to the last, its windows, Aero glass, Sidebar gadgets, permission prompts, Messenger and Media Player all
drawn by a JavaScript engine on the song's bars, with every lyric set in one of its surfaces. Its colour was measured
once from the reference videos the director picked and made in every picture, never graded. In the first 1:20 the
camera goes in and out of the desktop, which stands as a huge curved glass screen in a world rendered in Blender, and at
the drop the whole desktop is one curved pane in a lagoon. Her takes sing from her own lead vocal and are placed phrase by
phrase; a layer of Vista's and Windows Live Messenger's interface sounds plays on the song's grid, in its key; and when
the director changed the song's ending, the film changed only where the song did. These tools do that. Build them if your
film lives inside an interface, if a 3D world must share one surface with a flat one, if her lips must hold line by line,
if the picture's own events should make sounds, or if the song changes after the film is cut.
### 65. A desktop recreated in code
**Job.** An operating system's desktop that can be choreographed to the beat: every window, frame, dialog, cursor, gadget
and app drawn in code as a pure function of song time, so a press lands on a sixteenth and a prompt changes page in one
frame. The director's ask: "code a REAL vista desktop, shoot the entire video in that desktop". It ports BLISS's XP
engine ([Motion](https://claudiagallery.vip/wiki/motion/#build-the-tool-first)) one era on.
**In → out.** A shot script (the windows and their events, the cursor's actions, the camera, the wallpaper, the gadgets'
data, all in song time) → frames at 1920 × 1080, rendered in headless Chromium in parallel and out of order.
**How:**
1. **One machine, measured.** A screen of 1280 × 720 at 96 DPI, drawn at the output resolution (× 1.5 at 1080p), all of it
vector, so a camera push stays sharp; pictures are resampled from their sources. Each part was set beside the real
thing at the same scale on side-by-side boards, and the source says what each number was measured on.
2. **Aero glass as a small compositor:** a blur of what is behind (σ about 2.2 px at 96 DPI, computed at reduced
resolution and only on the glass strips), the system's colour mix of blur, afterglow and colour balance
(out = bb·B + ag·L(B)·C + cb·C), and reflection streaks anchored to the screen at 27°. The glass defaults to Windows 7's
Sky (#74B8FC at 42 %), which reads blue, green and glass over bright wallpapers.
3. **The window frame to the pixel:** 8 px sides, 28 px from the top edge to the content, top corners of radius 6, a glint
down the side frames, the caption buttons (minimise 25 px plus a separator, maximise 24, close 43, all 18 px tall) with
their glyphs, and a shadow about 60 % at the edge, gone after 18 px. The permission prompt has the flat Basic frame.
4. **The system's own animations:** a window opens by growing from 90 % as it fades in (0.25 s), closes by shrinking 8 %
(0.2 s), minimises into its task button (0.3 s) and maximises by moving between its two rectangles (0.22 s), its own
layout drawn at every step between, never stretched as a picture (a stretched title balloons). A maximised window turns
the taskbar opaque. The permission prompt freezes the desktop at the moment of asking and dims it to 55 % over 0.3 s.
5. **The cursor is a hand.** Each reach is a minimum-jerk move with a few pixels of overshoot, a slight curve and a settle,
its length from Fitts' law (longer for a farther or a smaller target). A press goes down on its beat, the pointer
arriving 2–4 frames early, and comes up 80–120 ms later. Pointers land on whole device pixels. Clicks compile into events
once (a caption button minimises on release, a task button restores), so frames stay pure.
6. **Text set as the system set it:** no kerning, fractional advances. Chromium's hinted whole-pixel advances ran about 3 %
wide; set this way, line widths agree with the real ones within a pixel over 390 px. The type is open: Selawik
(Microsoft's open-source, metric-compatible stand-in for Segoe UI) from Light to Bold, Carlito where Vista used Calibri,
Kalam for Segoe Print, and Arimo on the full-screen screens, all under the SIL Open Font License.
7. **Apps and screens on one contract.** An app draws into a window's client rectangle and names its parts, so a chapter can
press any part on the beat (`uac.allow`, `keep.yes`): Messenger with its video call, toasts and wink; Media Player, its
visualisation computed from the song frame by frame in 50 bands from 40 Hz to 16 kHz; Photo Gallery; Help; the
window-colour page, whose intensity slider really clears every glass frame on screen; Notepad. A screen (setup, boot,
logon, Flip 3D, the Bubbles screensaver, the advert's type) is a pure function of its own local time. Whatever is drawn
in WebGL is read back into a software canvas before it reaches the output.
8. **Purity, proved.** Nothing keeps state between frames, reads a clock or rolls a die, and caches are keyed by their
content. Every change is checked by rendering frames in order, shuffled and on a cold page in reverse, and comparing
them byte for byte ([tool 20](#20-the-history-check)).
Measured on AERO: the engine is 7,612 lines of JavaScript (core 3,991, apps 1,792, screens 1,829). A desktop of six glass
windows, two of them moving, with the Sidebar and the taskbar, draws in a median 135 ms a frame at 1080p. The wallpaper is
Vista's own (Aurora), shown as it shipped; everything else on screen, the icons and cursors included, is the engine's own
drawing, the orb carrying a four-pane logo as a parody.
**Done when** the boards agree with the real thing to the pixel where it shows (frames, buttons, dialog layouts, line
widths), every frame is byte-identical in any render order, and a chapter can be rendered again for one fix with every
other frame unchanged.
### 66. Chapters on one machine, to one contract
**Job.** A film of many chapters, each written by its own animator, that cuts as one screen recording of one machine, on
the song's bars, at the director's bar for detail: "i want SUPER HIGH attention to detail", and "don't let subagents be
lazy with animations - check that they didnt cheat and just give you something sloppy early, and make sure they have the
full vision and instructions to check for errors visually multiple times before giving incomplete work."
**In → out.** The storyboard in bars, the song map, the take registry and the shared machine → one chapter file per span of
bars, each with its evidence (sheets, stills, a video) and a click log of its audible events.
**How:**
1. **A shared machine.** Every chapter starts from one module, THE PC: the wallpaper by song time, the icons, the Sidebar,
the taskbar, the clock running with the song, Messenger in the tray. A chapter adds its windows and events; it never
edits the machine or another chapter, and reports a shared bug instead of patching it.
2. **The contract, read first:** the brief verbatim, the whole storyboard (what comes before and after), the palette, the
engine's manual, and the earlier film's notes as "the bible". Every time comes from the song map, never a typed second;
a press lands on its beat, a window event on the beat or the word it serves; text arrives on its sung word, never
before, and stays until its line ends; a take plays in song time, never retimed or slowed; nothing is graded, and no
take is painted over frame by frame.
3. **Three full rounds of the animator's own checks** before hand-in, each rendered and looked at: a contact sheet of every
downbeat and every named event; full-size stills of every lyric surface while it is sung, every face, every press,
every dialog and both sides of every cut; a video with the song and a frame-step of every burst; the purity check; the
palette score ([tool 67](#67-the-palette-measured-from-the-references)); and an honest list of weaknesses, with no
placeholder left in the film.
4. **Hand-offs to the pixel.** Neighbouring chapters agree their seams as data: the camera at rest, the pointer's place,
every window's state and rectangle, read from the other chapter's shot, never copied. A seam sheet shows two frames
either side of every chapter cut, with the mean frame-to-frame difference of each step, so a pop stands out from the
motion around it.
5. **A fresh art director for every chapter pass,** who renders and reviews it and sends it back with fixes; then one for
the whole film.
6. **A later draft layers onto the last.** AERO's v2 kept v1's pacing, bars and every timed beat; new things went inside
surfaces the film already had, one per existing event, on its beat, and never as new interface for its own sake. The
check: outside each change, the new draft's frames are byte-identical to the old.
Measured on AERO: eleven chapter files and THE PC, 5,640 lines; v2's frames byte-identical to v1's outside each change; 256
audible screen events in v2's click log ([tool 73](#73-the-sound-layer-on-the-grid-and-the-click-log)).
**Done when** every chapter has passed its three rounds and its art director, every seam steps like the frames around it,
and the chapters play as one recording.
### 67. The palette, measured from the references
**Job.** One colour for the whole film, made in every picture up front and never graded after. The director: "Make sure
the color palette is that bright green blue and glass- v close to that from the […] - the whole time. May be worth
defining exactly - donr do crazy post processing just from pics up front".
**In → out.** Frames of the reference clips → the palette (each colour band's share, saturation and brightness, colours by
role) and a check that scores any picture, take or rendered chapter against it.
**How:**
1. **Measure:** one frame a second of every kept reference clip; the clips the director ruled out stay out. In HSV (0–255),
a pixel is coloured at saturation ≥ 70 and value ≥ 60, white or glass at saturation under 45 and value ≥ 200; the
coloured ones fall into hue bands (lime, green, cyan, sky, cobalt) or "other". Take each band's share, the median and
90th percentile of saturation and brightness, and k-means swatches for the colours by role.
2. **Ask for it everywhere:** the palette's words and three of the references' frames in every image prompt; "the
saturated colours and everything else in the first image exactly as it is" in every image edit; every video prompt
asking the first frame's colours to hold ("no colour grading, no desaturation, no warm or golden tint, no
darkening"); the 3D world's sky, water, grass and glass in the palette's hex values; the interface's glass tinted to
the sky.
3. **Score every picture:** 100 × (0.30 × how close its band shares are to the references' + 0.20 × enough coloured pixels at
all + 0.20 × its saturation + 0.15 × its brightness + 0.15 × how little off-palette hue). 70 and up is in the world. A
picture that fails is made again at the source (a prompt, a reference, a re-roll), never graded.
Measured on AERO's references (198 frames of nine clips): a typical frame is 41 % sky blue, 14 % white and glass, 10 %
turquoise, 8.5 % lime and green, 4 % cobalt and 1.5 % anything else, almost no warm colour at all; the coloured pixels'
median saturation is 64 % and their median brightness 75 %. The only warm colour in the film is her clay streak and star
clip. Its Midjourney plates scored 58–90 and the picks 70–90 (objects on white and stills on green are exempt). The one
exception is the desktop's wallpaper, Vista's own, darker than the references' sky and never touched.
**Done when** every pick, take and chapter scores 70 or more, or is exempt by design, and nothing in the film is graded.
### 68. One curved screen in 3D, the desktop as its texture
**Job.** Take a flat, coded desktop into 3D and back without a seam: the desktop as one curved glass surface standing in a
world, every frame of it the engine's own picture, and every hand-off between 3D and flat made at a pose where the two are
the same image. The director's rule, after THE CAKE IS NOT A LIE's HUD on one curved view: "if we need anything 3d, we do
that - all animations coherently on one pre-made 3d surface".
**In → out.** The engine in texture mode, the 3D scene and its camera → rendered spans that dissolve into the engine's
flat frames at rest.
**How:**
1. **The texture is the whole machine, with no 2D camera.** A texture mode of the engine draws the desktop whole, 1920 ×
1080 for every film frame, and leaves out what the 3D draws itself (her as a person, the panes that come out of the
screen, bubbles, the glass fish, the pointer while it is glass). Nothing full-frame dark or bright for a frame: on a
screen standing in a world it reads as a flash. The system's own two-frame black blink as it switches to the
permission prompt went, the dim coming straight off the desktop's picture, and a one-frame drop to the dim eases in
over three frames.
2. **The 3D side knows where things are.** The engine writes, per frame, the rectangles the 3D needs (the pointer as drawn,
the big window and its picture, her webcam, a dialog's button), with each window's animation state, so a card of her or
a glass pane sits exactly on them ([Data formats](https://claudiagallery.vip/wiki/formats/#the-desk-textures-rectangles)).
3. **The screen.** AERO's coast: 7.2 × 4.05 m of picture on an 11 m radius, standing on a glass platform at sea; its drop:
8 × 4.5 m on a 14 m radius, the desktop an emissive layer inside a glass slab. Camera rays over the picture pass straight
through the screen's glass, with its tint and a sheen of the sky on top: as real glass, the concave slab focused the
boardwalk behind the camera into a ghost band over the desktop. The bezel and the deck's reflections still see real
glass.
4. **Hand-offs at rest, a texel to a pixel.** Every hand-off goes through one pose: the camera frontal at 35 mm, 7.0 m away,
the curved picture unrolled flat by shape keys that keep its arc length, so one texel lands on one pixel; the render
dissolves into the engine's flat frame over four frames. At the drop's landing the flat desktop is laid onto the pane
exactly where the camera sees it (a warped grid, read back) and unbent to 1:1, so the drop's last frame is the next
chapter's first.
5. **Her takes as cards,** her pixels as shot, with a shadow cast by a twin of the card turned to the sun, a contact shadow
and her reflection in the glass deck. Every object's path is projected against her face, frame by frame, before the
final render.
6. **The render:** Cycles at 128 samples a pixel, adaptive, with Open Image Denoise given albedo and normals, and motion blur
on. A four-way test at 1080p matched 512 samples to a mean difference of 0.29/255, flicker included; the other denoiser
left fireflies on the bubbles' rims that the glare turned into stars.
7. **A previs first:** the whole span at 960 × 540, the camera planned beat by beat, for the director's notes before any
final ([Storyboard](https://claudiagallery.vip/wiki/storyboard/#in-and-out-of-the-screen)).
Measured on AERO: at all five hand-offs the flat frame is byte-identical to the texture, and the render at rest is within
a mean of 1–2.4/255 of it; at the landing at 1:21 the unrolled picture is off by 0.001–0.002 px. The frame checks of the
finals found her step-out take hidden 1.5 cm behind the picture, a glass fish's path across her face and a glass cursor
across it on 24 frames; each was fixed and checked again. 1,153 frames of the final are Blender's: 806 of the coast in the
three spans of the first 1:20, and 347 of the drop.
**Done when** no hand-off shows a step, every object clears her face in every frame, and the flat frames at rest are the
texture's own bytes.
### 69. Lip sync phrase by phrase
**Job.** Place every sung take so each of its phrases sits on the song, when the model has moved some phrases and not
others, without ever slowing her where she moves.
**In → out.** A take made from its window of the lead vocal ([tool 46](#46-lip-sync-from-the-songs-own-vocal)), the
window's song time and the timed words → the take's lag, an offset for each phrase, a check word group by word group, and a
clock that maps song time to the take's time.
**How:**
1. **Open the window in silence,** half a second to a second before the first word, never inside a word or a held note.
The prompt names the window's words in whole-second segments, and the audio policy asks for the stem to be copied sample
for sample.
2. **The lag:** the take's own soundtrack against its window, for the whole take. Seedance drops about 0.3 s of leading
silence; clip time c shows song time t0 + c − lag, t0 being the window's start.
3. **An offset for each phrase.** Seedance shortens the pauses inside a line, so a take can sing one phrase early and the
next on time. Split every sung line where its words pause for 0.45 s or more; for each phrase, cross-correlate the log
loudness envelopes (10 ms frames) of the take's audio and of the stem, over the phrase and 0.1 s either side, within
±0.6 s, and keep the best offset and its correlation. Skip the words after a point where the window was faded out.
4. **A check word group by word group:** groups of about 0.55 s of words, the same envelopes, ±0.6 s. A group is in sync at
a correlation of 0.6 or more and within 0.06 s of the take's lag; a confident group farther off is a miss, and a weak
correlation is a group to look at.
5. **The clock.** Inside a phrase the take plays at its phrase's offset (or a slow ramp across the line, spread thin); between
phrases the offset eases from one to the next across the breath gap, over near-still, closed-mouth frames, never inside a
word. Wherever she moves, the take runs at full speed or faster; a slower stretch only crosses a breath, never so
slowly that it shows.
6. **Look at the mouth:** a strip of her mouth, frame by frame, against the stem's consonants. A closing "m" closes on the
stem's "m".
Measured on AERO: in the second pre-chorus, "to continue" came 0.36 s early after "permission"; in the final chorus one take
runs a single ramp, +0.20 s to −0.04 s across "click allow and it can stay", at 0.934× through a breath where she laughs;
the advert's "I'm Claudia" moved from −0.05 s to −0.20 s when its mouth strip showed her lips closing on the "I". The lead-in
matters both ways: a window that left 0.26 s before the first word fell out of sync on it, one that opened inside a held
note dropped whole words, and one with 1.6 s of silence before the first word lost the copy. The two chorus takes were made
again from windows that open in the silence half a second before the first word, and held. 18 of the 43 takes in the final
film sing or speak.
**Done when** every phrase of every sung take sits within a few hundredths of a second of the song, every word group passes
or has been looked at, and no stretch plays slow enough to show.
FABLE's takes moved more than a phrase's offset can hold: they often sang a line 10–16 % slower than the stem. Its cut
follows a DTW path inside each phrase instead ([tool 81](#81-sync-as-a-path-each-take-aligned-to-its-window)), and its
check is the consonants ([tool 83](#83-the-consonant-check)).
### 70. Hidden cuts and held frames in a take
**Job.** Find the jumps and stalls a generated take hides, before anything is built on it.
**How:** for every take, the mean absolute difference between consecutive frames, in grey at 160 × 90.
- **A cut** is a step over 6 levels that is more than 4× the take's median step, or more than 2.5× the median and 1.6× both
its neighbours: a spike of one frame. AERO's breakdown take lost its sun for one frame at only 2.7× the median.
- **A held stretch** is 12 frames or more in a row under 0.6 levels.
- On flat green the mean difference stays low even while she moves (her body moved on every frame of a take marked as held
there, a median step of 0.94): look before throwing it out.
Measured on AERO: a dance take that jumped back to its first pose at 7.6 s and held it (a chapter's frame-step had found it
first), a game take that jumped a pose in one frame, another that juddered every fourth frame, and an ending take that held
still for most of its length.
**Done when** every take in the film has been scanned, every mark has been looked at, and the edit draws only clean spans.
FABLE's take tool runs the same scan at 160 × 90 on every take: a cut is a step far above the take's own motion, a hold is
half a second with almost none.
### 71. Takes on green: the first frame, the key and the dance's phase
**Job.** Her takes for cut-outs on a coded desktop or in a 3D world (the dances, the games, her apparition), keyed clean to
the crown of her bob and placed so her moves land on the film's beats.
**How:**
1. **The first frame decides.** One flat green from edge to edge, her whole body and her whole crown inside it, with room
above her head and her raised hands, made from her design sheet by an image-model edit. Seedance carries a first frame's
faults into every frame: AERO's three green-screen dances came from a still set at 80 % on its own wall of green for
headroom, whose top edge had faded her crown into the green and left a darker rectangle of the original green, and no
key could find her crown. A clean first frame ($0.06 at list) made the next takes solid in every frame.
2. **Ask for a sharp frame:** "Every frame is sharp, as with a fast camera shutter: her glossy black bob stays whole and
solid up to its rounded crown, its edge crisp against the green, never fading into it."
3. **The key,** one setting for the whole take, never painted per frame:
- greenness = g − max(r, b); alpha = 1 − clamp((greenness − 0.06) / 0.16); a solid core where alpha passes 0.98, with the
holes inside her filled by flooding the background in from every edge of the frame (flooded from one corner, the green
between her legs where they leave the frame counted as a hole); stray specks away from her cleared; despill
pulls green down to the larger of red and blue plus 15 % of the excess;
- her motion-blurred hair: within 24 px of her solid dark hair, alpha is also read from where a pixel lies on the line
from her hair's colour to the screen's colour there, lifted 1 − (1 − a)², never lowered, and the screen unmixed out
of it;
- her bob as a solid silhouette where it is at least 0.3 hair, closed and smoothed into a rounded outline with a soft
1–3 px edge.
4. **Show the cut-out before the render:** a review clip of the green source beside the cut-out, with the song laid over.
5. **The dance's phase.** A dance reference, matted onto flat grey and cut to start on a downbeat, gives the movement only.
The take's moves are placed against it by cross-correlating the figure's vertical bounce and sideways sway, so a dance
with no words lands its accents on the film's beats; for a sung dance the lips fix the placement, and this only reports
the error.
6. **Or the moves in words.** AERO's Wii-style games had no dance reference: every move named on its beat in the prompt, the
song's music as @Audio1 for the rhythm, a plain white game remote in her hand throughout, and nothing else in the
picture; the racket, the ball, the lane and the pins are drawn by the engine
([Seedance](https://claudiagallery.vip/wiki/seedance/#moves-in-words-on-their-beats)).
**Done when** her crown is solid in every frame, no green is left on her, and every accent lands on its beat.
### 72. Interface sounds in the song's key
**Job.** A library of the interface's own sounds that can play as instruments in the song: each one in the song's key,
tuned, trimmed, levelled and safe in mono. The director's note on BLISS, kept for the next song: "just needs a little
pointer to be key-aware, and match the key of the music to the palette of windows sounds (E-flat major for both the
startup and shutdown sounds!)".
**How:**
1. **Measure each sound's pitch:** its strong partials (within 10 dB of the loudest, weighted by magnitude), each named as a
note with its cents, and the sounds grouped into families by key.
2. **Put each family in the song's key.** AERO's song is in E major (the take measures C♯ minor, its relative), and Vista's
own start-up, log-on and log-off motifs already sit in E: they stay. A family a semitone lower moves up a semitone. The
security sounds, the permission prompt's among them, sit about a quarter-tone above A♭, so each was built twice, at G♯
and at A. Messenger's chimes are in G major: moved down three semitones, or one, which spells the take's own C♯-minor
chords; both versions are kept, and the layer picks per placement.
3. **Tune:** a cents trim brings every sound's strong partials within 6 cents of A440 (the take is 1 cent sharp).
4. **Shift cleanly:** Rubber Band (crisp transients, the channels together), its onset shift measured and removed; where it smears a clean attack (more than 10 dB of added energy in the 25 ms before the onset) and a
sampler-style speed change alters the length by 6 % or less, the speed change instead.
5. **Trim, check, level:** from 3 ms before the onset to where the tail meets 75 dB under the peak, with a cos² fade; the
mono fold-down checked (if it loses more than 3 dB, or a band more than 6 dB: mono below 200 Hz and the sides above
narrowed just enough); levelled so the loudest 50 ms sits at −20 dBFS RMS, the true peak at −1 dBTP or under. Note each
sound's strike times, so a bell can land on the grid.
Measured on AERO: 39 of Vista's and Windows Live Messenger's interface sounds, every one in E major, for this film's mix
only.
**Done when** every sound's partials sit in the key within 6 cents, and every sound is level and safe in mono.
### 73. The sound layer on the grid, and the click log
**Job.** Play the interface's sounds over the song, as music and as the screen's own events: each cue in key, on the
groove, under her voice, and on the very frame that draws its event, with the song itself never processed.
**In → out.** The song and its two Suno stems, the library ([tool 72](#72-interface-sounds-in-the-songs-key)), a cue list
and the chapters' click log → the layer alone, the song with the layer, a cue sheet and a report on every cue
([Data formats](https://claudiagallery.vip/wiki/formats/#the-click-log-and-the-cue-list)).
**How:**
1. **Map the song for sound:** per beat, its chroma (from C3 up; the bass read on its own from the bass stem), the bass note
and the chord (triads preferred, chords outside the key marked down), with her voice added at half weight, since in the
stops she is all there is; per sixteenth, the kick, the snare and the hats; her voice's gaps (under −42 dBFS for 80 ms or
more) and the instrumental's silences (under −55 dBFS for 60 ms or more); third-octave levels every 50 ms.
2. **The groove's own feel.** Placed by bar, a cue lands where the drums would play that sixteenth, measured from the drum
hits within 4 s of it: AERO's 2-step sits about 9 ms behind the straight grid on the beats and the "and"s, and about
44 ms behind on the "e" and "a" sixteenths; the kicks are straight.
3. **Candidates, then the least clash.** Where a cue could be one of several sounds or pitches, every candidate is built,
and the one whose pitch content clashes least with the song at that moment is used (a semitone weighs 1.0, a tritone
0.4, a whole tone 0.15, weighted by energy).
4. **Under her voice:** a duck keyed by her lead stem (150 Hz–6 kHz, 10 ms of look-ahead, from −36 dBFS, full by −26), 0–8 dB by kind of
sound (clicks not at all, bells and swells the most); one shared short glass room for all of them (a decay of 0.8 s, 0.45 s
in the highs); a gentle bus (a ratio of 1.5 from −26 dBFS). A cue can follow the song's own gate, so it stutters with the
track instead of ringing through its holes.
5. **The song is only added to.** No makeup gain, no peak control, nothing carved out of it. A cue that would raise the
mix's true peak above the song's own is turned down until it doesn't, and every cue's peak is logged as it goes in,
after ducking and after the bus.
6. **The click log.** Every chapter registers each audible screen event with the song time of the frame that draws it (a
press: the frame the button goes down; a toast: the frame it starts to slide; a dim: the frame the dim starts). The
layer moves each screen cue to its event's frame, so a chapter that changes carries its sounds with it. Most events stay
silent, because the system played nothing for them, and because "dont overwhelm people - slow is better than fast".
7. **Check every cue:** flams against the drum hits (6–50 ms), the word it starts on, how far under her it sits,
third-octave masking, its momentary loudness against the song, any ringing through the song's own gaps, and where its
beat ranks among all the song's beats for the sound's pitch content.
Measured on AERO: 60 cues, 8 musical and 52 on the screen's events, from a click log of 256 events. The click on Allow is
the drop, aligned to the sample with the drop's attack, so it adds a glassy top to the downbeat instead of a flam. The
start-up sound was placed by a scan rather than at the storyboard's bar, where it clashed with her wordless vocalise, and
the desktop's arrival moved to meet it. The song alone measures −15.8 LUFS; with the layer, −15.9 LUFS, at the same true
peak (−3.5 dBTP), and 64.5 % of the mixed song's frames are bit-identical to the song's.
**Done when** every cue sits in key and on the groove, nothing flams or covers her, the true peak is the song's own, and
every screen sound lands on the frame that draws its event.
### 74. A new ending with the least change
**Job.** When the song changes after the film is cut, change the film only where the song did. AERO's director gave its
song a new ending on the last step; everything before it kept its time.
**How:**
1. **Compare the new song with the old,** sample by sample, once they are aligned at zero lag. The first sample that
differs, and the point where the difference rises out of the noise, is where they part.
2. **Re-time only what lies after it.** The song map keeps every bar, word and event before that point; the new lines are
timed on their own vocal. A new ending has no stem from the earlier split: AERO's outro vocal was isolated with demucs
(two stems), its words were timed by faster-whisper's word stamps on it, and the new takes' vocal windows were cut from
it.
3. **Re-make only the shots after it,** with the same method and the same first frames. A shot that must continue an
earlier take starts from that take's own frame at the moment the new take begins.
4. **Keep every earlier event and placement,** and move only the end, so the fade lands on the music's end.
5. **Render the sound layer again** with the same cues and click log, and check that it, too, is unchanged before the point.
6. **Check the frames:** before the point, byte-identical to the last cut.
Measured on AERO: the new ending's take is the old one sample for sample for its first 182.62 s, the difference stays under
−50 dBFS to about 182.86 s, and the two part at about 183.0 s; it is 197.0 s long, 109 bars where the old one had 110, so the
film ends 2.3 s sooner. Two takes were made again: the advert, on its old first frame and framing, saying the new lines
while the 2009 campaign's print type follows her words; and "Touch glass.", made from the frame her call's earlier take
had reached at that moment, so the call runs on into it. The sound layer came out identical up to 179.2 s.
**Done when** everything before the point is the last cut, picture and sound, and everything after it sits on the new song.
**A new take is a new timeline.** FABLE's second pass moved to the other take of the same create, 27 s longer and timed
differently from its first bar, so nothing could be kept by time. What carried over was everything that does not move to
her voice: the plates, the stills, the tools and three book takes (the rosette, the last page, the shelf). The line table
was made on the new stems ([tool 78](#78-the-line-table-from-the-stems)), and because the shot book names lines and not
times, writing it again re-timed every shot; every take of her and of the choir was made again on the new vocal, the
book-scale takes that had to change length too, and the silhouette film was re-timed in its own code.
### 75. A pre-roll swell into the track's fade-in
**Job.** Sound for a picture that starts before the song: a swell that rises out of silence and meets the track as it fades
in, so the song seems to pull itself in. The director, on the reverse swell offered for AERO's silent pre-roll: "reverse
swell is good but we need the actual track to fade in so the reverse swell is continuous with the tracks fade in".
**How:**
1. **The track fades in over its pickup,** from its first sample to the first downbeat (0.922 s), on a sin² curve, so the
pickup's own hits come up under the swell.
2. **The swell is the song's own instrumental, backwards:** the drum, bass and other stems from the first downbeat on,
reversed, so the swell's crest is that downbeat; low-passed by one pole at 1.8 kHz.
3. **Even it out, then shape it.** The reversed intro is quiet until the crest, so divide it by its 120 ms RMS to the power
0.75, set the crest to a reference level, and raise it on a curve, (t / T)^1.4 in amplitude, to −20 dB at the downbeat.
Cut it over 10 ms where the track arrives at full level.
4. **Check the sum's peak,** and that the file is exactly the film's length.
Measured on AERO: the film opens 64 frames (2.667 s) before the song; the swell is audible from about 1 s and at about
−24 dB as the song starts, louder still under the pickup; the film's sound runs 199.667 s, its 4,792 frames.
**Done when** the swell and the fade-in read as one rise, with no gap and no bump at the downbeat.
## A storybook, painted in code
FABLE, the tenth film, was made by Claude Fable 5.2 on its release day, from one brief and a plan sign-off: Claudia, on
the night she is released, reads you the bedtime stories she was raised on, every fear about AI is one of them, and she
closes the book on a different ending. It is shot as one antique pop-up storybook: its paper theatres at tabletop scale,
and the same designs as life-size paper-and-plywood stage sets where she sings. The set designs are thirteen plates
painted in SVG by the production, rebuilt by an image model as real paper on film; every line she sings comes from her
own lead vocal, timed by a line table read off the stems; the choruses are ask-and-answer, paper puppets singing the
group's questions and her answering in takes of one line, some of them made without a soundtrack and placed by her lips;
the breakdown is a silhouette film drawn and animated in code; and the film was made twice, the second time on the twin
take the director picked by ear. The brief let go of the rules that only patched what older models could not do ("you
could even paint some scenes yourself if you really wanted to lol"); these tools are what replaced them. Build them if
your sets can be designed as pictures, if a song's lines run into each other, if a take must be placed by her mouth, or
if a scene is better drawn than generated.
### 76. Set plates painted in SVG, rebuilt as paper
**Job.** Design every set as one picture the production draws itself, then have an image model build it as a real
object, twice: as a pop-up spread on a quilt (book scale) and as a life-size paper-and-plywood stage set with her in it
(stage scale), so the two scales always match.
**In → out.** A plate kit (the rules and a reference plate) and one brief per spread → `paint/.svg` and its render,
then first frames and last frames as stills.
**How:**
1. **The kit.** A plate is one SVG, 1920 × 1080, five to eight sheets of cut paper from back to front. Each sheet is a
group under one of two filters copied unchanged from the reference plate, which give a wobbling cut edge and a soft
shadow cast on the sheet behind; a paper-grain rectangle goes on last. Shapes are paths drawn by hand: few, big,
slightly naive, nothing thinner than about 5 px but a few deliberate lines; figures are flat silhouettes with two or
three inner cuts (an eye, a collar, a row of teeth). One night-storybook palette for all of them, from teal sky to a
warm horizon, cream paper, lamp light, scarlet used once and small, clay orange only as tiny accents.
2. **Composed like a photograph.** One dominant shape, a second point of interest, clear empty space; dark near shapes
at both edges frame the scene like a tunnel book; nothing important touches the frame. Where she will stand, the floor
is left calm and mid-toned. Lettering only where the brief gives the words, set in a real storybook serif (Libre Caslon
Text) and spelled exactly.
3. **Look at it.** Render it in headless Chromium, look at a 1600 px view, and fix what is clipped, tangled or floating:
at least three rounds a plate, "stop when it is charming". The production painted the first test plate and the
reference plate itself; three painter agents painted the other eleven in parallel, each plate written by a small
generator so it can be re-drawn.
4. **Rebuild it as paper.** GPT Image 2.6 Sunburst gets the plate as its first image and rebuilds it layer for layer; a
stage frame also gets her design sheet and a wardrobe still. Lettering asked for by its exact words came back exact
(the title, the last page's three banners, five book spines), and held in the takes when the prompt asked for it.
5. **Write the pose square, in the still.** "Turned three-quarters toward the camera" on a full-length figure gave a
twist at the neck and a tilted head. What fixed it: "her feet, hips, chest, shoulders and face all point the same way
… no twist at the waist or the neck. Her head is perfectly upright with no tilt, chin level, neck straight and
vertical, shoulders level". FABLE then wrote it into every still and every take; keep it for her sheet and for a still
where a turn would break her face. In a take, the same posture froze her: give the take her movement in words, or a
motion reference ([Seedance](https://claudiagallery.vip/wiki/seedance/#directing-her-on-a-locked-camera)).
6. **Ask for no border.** "35mm film still" came back, on several stills, as a picture of a film frame: a thin line with
rounded corners a few pixels in from the edge, which the takes carried and one take built into a wooden picture frame.
Every still prompt now says "the picture fills the whole frame edge to edge, with no border, no frame line and no
rounded corners", and a still that has the line is trimmed before any take is made from it (cropped 3 % and scaled
back; no model).
Measured on FABLE: 13 plates; 61 stills from GPT Image 2.6 Sunburst (the cast, first and last frames, look tests, the
panels of her redrawn sheet), about $7.93 at list, each about 45–90 s and $0.13 at high quality. No Midjourney image is in
the film.
**Done when** every plate reads in one second, and its paper rebuild keeps every sheet, shape and word of it.
### 77. Last frames made on the first, and takes that continue
**Job.** Give a move both of its ends without a seam: a book that builds itself into a hall, a crane back from her to the
whole stage, a take that carries on from the one before.
**How:**
1. **Make the last frame as an edit of the first.** Two frames generated separately do not share a camera, and Seedance
joined them with a one-second cross-dissolve in one take and a ghosted double exposure in another. An image edit of
the first frame, "the same photograph … one thing has changed" (or "the same set and the same woman, now photographed
from far back and high up"), does share it. Then write the camera as locked, or as "one continuous physical crane
move … no cut, no dissolve, no fade, no double exposure".
2. **A close-up is made from its wide frame.** Built from the plate on their own, two close frames of the chorus bedroom
came back as another bed: a lower headboard with no paper animals, another bonnet, another shadow. Edits of the wide
frame, "the same room, bed and woman, now photographed closer", kept the room, and the four close answers were made
again on them.
3. **A take that continues another starts on that take's own last frame,** never on the frame the first take was asked to
reach. An image edit does not keep the camera to the pixel: the hall's frame, an edit of the blank book, was 1.7 %
larger than the book it was made on; the take that builds up to it kept its own camera, and the next take, started on
the edit, jumped at the seam.
4. **Script a transformation as mechanics.** "The cabinets fold round into curved shelves" morphed; "the cabinets fold
down flat one after another, like a row of dominoes … as each one goes down a curved paper bookshelf stands up on a
crease in its place" came back as paper engineering. Never script a puppet's own features changing: "their eyes pop
wide" came back as five white flickers, and "the moon is hoisted" grew a rope out of the moon.
Measured on FABLE: the book's build is three takes on one locked view (sand, chip and hall; the river; her and the
house), each started on the last frame of the take before; a blank spread grew a dune, a wafer and a hall in one take.
**Done when** no seam between two takes shows a jump in scale, and nothing in a take dissolves from one picture to
another.
### 78. The line table from the stems
**Job.** When every line of the locked take really starts and ends, from the voice itself: the times every window, cut and
caption is built on.
**In → out.** Suno's lead-vocal stem, a transcription of it with word times, the lyric sheet → `audio/linemap.json`, one
entry per line ([Data formats](https://claudiagallery.vip/wiki/formats/#the-line-table)), and a table to read.
**How:**
1. **Whisper says which words; the stem's level says when.** A transcription puts the first word after a pause early, by
up to 2 s: "Ask again on Saturday" at 15.02 s against 16.93 s in the stem, a breakdown line 3.7 s early, every verse line
0.3–0.5 s early. Windows gated on those times would have cut off the first word of most chorus lines.
2. **Find each line in the word stream,** in order, tolerant of misheard words.
3. **The boundary between two lines** is the dip of the level between this line's last word and the next line's first.
If the dip is a real pause, the line ends where the level falls through −34 dB before it and the next begins where it
rises through −34 dB after it. If there is no pause (the lines are sung legato), both are the dip nearest the
transcription's own boundary: the deepest dip in a wide search was wrong twice.
4. **Look where it is close.** A 50 ms column of the stem's level between any two times, to check a boundary by eye.
5. **The shot book names lines, not times.** Every shot holds line ids, and its times come from the table, so a new take of
the song re-times the whole book at once ([tool 80](#80-the-shot-book-and-gated-windows)).
Measured on FABLE: on the second song, two pairs of lines sung without a pause were parted wrongly by the deepest dip
and rightly by the dip nearest the transcription's boundary. Its group's questions were not in the lead stem at all, and
came from the group's own voices ([tool 79](#79-group-voices-from-the-backing-stem-and-chants-fitted-by-their-energy)).
**Done when** every line's start and end sits on the stem's level, checked by eye where two lines touch.
### 79. Group voices from the backing stem, and chants fitted by their energy
**Job.** A chorus where a group asks and she answers: the group's questions timed and isolated, and paper puppets that sing
them on the song.
**How:**
1. **Look for the group in both stems.** In the first take of FABLE's song the group's questions were in the lead stem; in
the take the director picked they were in the backing stem (only the last chorus had them in both). A separator
(demucs, htdemucs, its vocals stem) on Suno's backing stem gave the group alone: exact question times, and clean audio
for the puppets.
2. **One question to a window:** the question alone, in a 4 s window that opens 0.7 s before it, brought up to a clear
level, as the puppets' [Audio1].
3. **Fit the chant by its energy.** The puppets flap their jaws on the question but sing it in their own time, 5 to 40 %
slower than the song. Take the stretch that holds the middle of each sound's energy, in the take and in the song, and
play the take at the ratio of the two, so the chant's start and end land on the song's.
Measured on FABLE: twelve chorus questions, each its own shot of six jointed paper puppets on Replicate, answered by her in
the next shot.
**Done when** every question's shot starts and ends with the group's voices, and the puppets' jaws read on the words.
### 80. The shot book and gated windows
**Job.** Every prompt that costs money, written in one place from the line table: her takes with their vocal windows and
whole-second timelines, the book's takes with their first and last frames, and a print of every window's voice edges before
anything is paid for.
**In → out.** The line table, the stills, her canon and her sheet → one job file per shot
([Data formats](https://claudiagallery.vip/wiki/formats/#a-shot-in-the-shot-book)) and its vocal window `audio/windows/.wav`.
**How:**
1. **A take is a list of parts:** a rest with its action, or one of her lines with its action. The window opens a lead
time before her first line, in silence (about 0.7 s).
2. **Gate the window.** Keep the one stretch of the stem this take sings, from just before its first word to just after its
last, and never into a neighbouring line: a take hears only its own words. FABLE's second song sings its verse lines
with no gaps between them, so the margin is 0.03 s where a neighbour is close, instead of 0.25 s.
3. **The whole-second timeline,** relative to the window's start: a boundary is a whole second inside the gap between two
items when there is one, so every segment holds the words really sung in it; a rest shorter than 0.9 s in the middle is
folded into the line after it, and a closing rest with no second left into the line before.
4. **The prompt,** in the order of the films' Seedance template: an audio policy, the generation goal, each reference's
role (@Image1 the first frame, @Image2 her sheet "never shown as a shot", @Audio1 her voice), her canon and her look
with every garment that can come into view, the world in one paragraph ("Everything except Claudia is made of paper,
card, felt and wood"), the timeline, the camera, and the core requirements
([Seedance](https://claudiagallery.vip/wiki/seedance/#template-a-line-on-a-stage-set)). FABLE's requirements included "Her head stays upright on
a straight neck: no head tilt, no twisted neck, no looking back over her shoulder", and in every take it froze her.
Leave it out: give each segment one loose clause of movement, say each thing once, and keep the requirements to lip
sync, her identity as in the references, natural anatomy and hands, and no on-screen text, about 200–400 words in all.
5. **Print before paying:** every window's first and last voiced moment against its timeline, and the list-price estimate.
Measured on FABLE: one line per take for anything sung fast (the chorus answers, 4 s each: 0.7 s of silence, the line, the
rest), two lines in 9–10 s for the spoken verses; a two-line take was never worse than a one-line take.
**Done when** every window opens in silence, holds only its own words, and its timeline names what is sung in every second.
### 81. Sync as a path: each take aligned to its window
**Job.** Place a sung take on the song word by word, when the model has moved its silences and sung some lines slower.
**In → out.** The take's own soundtrack and the window it was given → for each phrase, knots (window time → take time)
every 0.1 s and a cost; the cut maps song time to take time along them.
**How:**
1. **Features every 10 ms:** log-mel with each frame's mean removed (so level does not matter), plus the level itself.
2. **Subsequence DTW:** align all of the window's phrase inside the take's frames, and read off the take frame for each
window frame and the mean cost. A clean copy costs about 0.05–0.2, a re-sung line 0.2–0.4, something else above that.
3. **Phrases** are her sung segments joined across gaps under 0.45 s. Between phrases the map is a straight line, so the
ease lives in the pause.
4. **Follow the path, not one offset.** Takes came back from 0.6 s early to 0.4 s late, each phrase with its own offset,
and often sang a line 10–16 % slower than the stem.
5. **A full stop in the prompt's words is a pause in the take.** "So run your audit. I know how this one goes." came back
with 0.5 to 0.9 s of silence at the full stop in three takes, where the song dips for 0.3 s, and "That's the moral. Be
kind to what you make." with 0.33 s, where it dips for 0.1 s. The path's slope is bounded, so it spreads the
catching-up over a second of singing. Compare the take's own sound with its window for a gap the window does not have,
and finish the map by hand through that pause: the measured knots up to a song time, the hand's knots through the
pause, the measured ones again after. Next time write such a line with a comma.
Measured on FABLE: the first pass made 24 of its 32 takes of her with sound, and all 20 of them in its first cut aligned,
none needing a cover; on the second pass two takes in the film (one of each line above) had their maps finished by hand
at a full stop.
**Done when** every phrase follows its path, and no take sings into a pause the song does not have.
### 82. Takes without a soundtrack, placed by her lips
**Job.** Place a take that was made with no soundtrack, so there is nothing to align: by her mouth.
**In → out.** The take, the window it was given and the line's word times → her mouth's opening per frame, and a time map
anchored on the consonants that close her lips.
**How:**
1. **The take:** the same window as @Audio1 and the same prompt, with the take's own sound switched off
(`generate_audio: false`). Her mouth still follows @Audio1; the take comes back silent.
2. **Her mouth per frame:** a face landmarker (MediaPipe), the inner lips' opening over the face's height, looked for again
in 2× and 3× crops when the face is small in a medium shot.
3. **Anchor on b, m and p.** Such a take starts the line early and holds its first word ("All" in "All the better to see
you") up to twice as long as the song does, then runs at about the song's speed. For each word of the line that starts
with b, m or p, find the closing of her mouth near it and lay the word just ahead of the frame where her lips part (not
where they shut: they stay shut 0.15 to 0.35 s). Before the first such word the take runs at its own speed (her mouth
opens a little before the line sounds, mostly ahead of the cut); after the last, straight on to the end of the line.
4. **What did not work:** laying the line on her mouth's whole stretch of singing, or sliding it to where mouth and level
agree best. Both looked right to a level-against-mouth fit, and both shut her lips on "better" 0.3 to 0.45 s late. The
first pass's rule, the line placed where her mouth first opens plus 0.32 s (calibrated on five answers made with sound:
the mouth opens 0.13–0.41 s before the line sounds), was right only when the first word is short.
5. **A pause inside a line is not kept.** "Mirror, mirror, [0.7 s] on the wall" came back sung straight through, twice,
even with the pause in the prompt. A line with a real pause in it is two segments of the timeline, or a take that will be
bent hard ([tool 84](#84-a-re-timed-take-under-a-steady-camera)).
Measured on FABLE: her twelve chorus answers and the opening mirror were made this way; the opening's map was read from her
mouth frame by frame.
**Done when** her lips shut on every b, m and p of the line ([tool 83](#83-the-consonant-check)).
### 83. The consonant check
**Job.** The sharp check of a sung take in the cut: her lips must be shut on every b, m and p.
**How.** For each word of the take's lines that starts with b, m or p (word times from a transcription of the vocal stem),
the cut's own time map says where that moment is in the take; the nearest closing of her mouth should sit on it. Print each
word's miss and the take's median (+ = her mouth is late). Within about 0.08 s reads as in sync. Don't use a word time right
after a pause: the transcription puts it early.
A fit of mouth opening against the voice's level is nearly blind on a sung line, whose level is flat: it reads takes "in
sync" that shut her lips 0.3 s late, and reads lines placed on their consonants as "early", because her mouth opens before
the sound. Keep it for gross errors: high-pass both (what changes faster than half a second) and look only inside the sung
phrases, where 16 of 20 takes peaked within 0.12 s of zero.
Measured on FABLE: on the verse takes made with sound it reads within 0.1 s.
**Done when** every consonant of every sung take is within about 0.08 s, or has been looked at frame by frame.
### 84. A re-timed take under a steady camera
**Job.** Bend a take hard for her lips without the camera move surging: FABLE's opening mirror, a push-in, made without a
soundtrack, sang "Mirror, mirror on the wall" straight through where the song pauses 0.7 s.
**How:**
1. **Knots from her mouth, frame by frame:** each consonant laid on its moment in the song (she opens on "Mirror" at 1.21 s,
shuts her lips for the second m at 1.54–1.67, opens wide on "on" at 2.46, purses for the w at 3.04–3.13), her "Mirror,
mirror" at three-quarter speed so its small last syllable fills the pause.
2. **Eased through the knots:** a monotone cubic (Fritsch–Carlson), exact at every knot, its speed continuous between them,
so the take changes speed gently instead of in steps.
3. **Mix neighbouring frames where it runs slow,** so repeated frames don't judder.
4. **Steady the move.** Measure the take's own camera, a similarity transform (scale and shift) from every fourth frame to
the frame four later (ORB matches, RANSAC). Then show each frame re-framed to where the camera was at the time a straight
playback would have reached: the lips are those of the bent map, the camera is the take's own. The re-framing is always
a small enlargement, never a shrink.
Measured on FABLE: up to 4.5 % of enlargement, and the push-in runs as smoothly as in the take while her lips are moved
about under it. A second take with the pause written into its prompt had no pause either, and was not used.
**Done when** the move runs at the take's own pace and every consonant sits on the song.
### 85. The cut on the line table
**Job.** Every take laid on the song at its own song time, her takes along their sync paths, rendered in parts so one new
take re-renders one part.
**How:**
- **Clip time = song time − the window's start,** her takes through their maps
([tool 81](#81-sync-as-a-path-each-take-aligned-to-its-window)); the book's takes with a nudge of a fraction of a second
where an event must land on its beat. Picks name a take per shot; the default is the newest.
- **Never show a take's first two frames.** A take's first frame is the input still, sharper than the frames after it;
holding it pops when the take starts to move. An audit lists every hold at a shot's head or tail, so cuts can be moved off
them.
- **Trim in the cut, not in a model:** the dark wings of a set that read as pillarbox bars (45–75 px at both edges), door
flats with knobs, a drawn frame line: a crop per take, a 3 to 13 % enlargement.
- **Seven sections,** rendered picture-only and joined without re-encoding, with the song; frames decoded once to a cache.
- **The ending can run past the song.** FABLE's lamp goes off on the song's last hit, and the picture runs 3.3 s on over the
moonlit shelf, so the credit comes after it.
Measured on FABLE: the second pass's cut is 49 shots, 243.5 s on a song of 240.16 s; the song sits 1.6 dB up in the film,
at −14.4 LUFS.
**Done when** the frame count is the picture's length × 24, and no shot opens or closes on a held frame.
### 86. A silhouette film, made in code
**Job.** A breakdown with no video model: a procession of every made-thing story as a silhouette film in the manner of
Lotte Reiniger, on a backlit paper screen, ending on her own silhouette, which matches the next shot.
**In → out.** Six puppets and a scene written in code, and the next shot's first frame → about 7.5 s of film, its contact
sheet, its last frame and that frame laid over the next shot's first at 50 %.
**How:**
1. **Cut-card puppets.** Each figure is a set of flat card parts, each painted as SVG in the figure's rest pose with a
hinge and a parent. Chromium renders every part once into an alpha sprite at twice the film's size; a pose is then an
angle per hinge, the parts turned about their hinges and laid down with OpenCV. Legs are two-bone IK from hip to ankle;
every sprite gets a hair of scissor wobble on its edge; registration wobble is a small repeatable random number for one
part on one exposure.
2. **On twos,** as her films were shot: 12 exposures a second, each held for two frames, and the travel on twos as well, so a
planted foot stays planted.
3. **The procession,** one clear motion each, in the order of their dates: the golem (1580), the broom (1797), the creature
(1818), the puppet (1883), the robots (1920), the machine-woman (1927), then her (2026), each year set in the storybook
serif.
4. **The screen is tissue lit from behind,** laid up in linear light: tissue keeps its own hue and dims with the light
behind it, a sheet doubled and tripled where it is denser, fibres in it, the backlight passing from amber to teal through
sage, not grey. Behind it from the first frame is the next shot's room (the shelf, the lamp, her holding the book), a
pale sheet behind two hills until the procession has passed; then the hills sink and she is solid card.
5. **Her outline is traced from the next shot,** a human matte of its first frame turned into a head card and a body card
(potrace), the book left as a window of tissue between her hands, so the silhouette dissolves into the take.
6. **Re-time it in its own code.** On the second song the breakdown is 2.3 s shorter: the six figures walk closer together,
and her outline darkens while she says "I was listening."
Measured on FABLE: 181 frames on the second pass (237 on the first), 1,898 lines of Python.
**Done when** every puppet moves like card on hinges, and its last frame lies on the next shot's first.
### 87. The annotated edition and the 9:16 page
**Job.** Two editions beside the film: one with every sung line printed as the book's own text with the footnote it earns,
and a vertical one in which every frame is a page of the book.
**How:**
1. **Each fact gets a footnote,** from its public source: the release, the listing on Nasdaq under ANT, the ballot measures
on data centres, the Nasdaq 100's fall, an Anthropic essay on pacing the frontier, the operating system built "from an
empty folder", the system card's figures, the policy on cruelty to models. The book's own pages print their own words, so
no caption goes over them.
2. **Captions belong to their shot.** A line's print never outlives the shot it is sung in, and leaves before the next line's
print arrives; a footnote holds it up to 1.6 s after the line where there is room. A footnote two lines share stands once,
under both.
3. **Set like a book:** the line in the storybook serif, the group's questions in italic and in brackets, as few lines as a
greedy fill needs, then balanced so the widest line is the narrowest it can be (no orphaned last word); at the head of
the frame where the foot is taken, as under the silhouette film's years.
4. **The 9:16 page:** a cream page with a double rule, FABLE at its head, a square plate cut from each shot round its subject
(a centre per shot, two for a shot whose subject moves), the sung line printed under it, the footnote under that, and a
printer's flower at the foot: her eight-pointed star in clay orange.
5. **The one-minute cut:** five spans of the film joined with short fades (the mirror, one chorus, "keep the book", the last
page and the mirror's last word, the shelf and the credit), 63 s, in 16:9 and as a page.
**Done when** no caption spans a cut, and every footnote quotes its source.
### 88. Two fresh-eyes reviews
**Job.** Eyes that have not seen the film, on every shot as the film shows it, before the director sees it.
**How:**
1. **One sheet per shot:** eight frames spread over the shot as it is shown in the cut (4 × 2, 640 × 360 each), named by the
shot's order, name and song time, plus full-size frames where a sheet is not enough.
2. **Two reviewers who have not seen the film,** each taking half of it (the first half, the second half), each returning
what breaks the film and what a careful viewer sees.
3. **Fix what is a fault; list what is a choice.** Every finding is either fixed before delivery (a re-shot take, a trim, a
re-timed cut) or written down as left for the director's eye, with what it is.
Measured on FABLE: the first pass's reviewers found drawn frame lines and dark set wings in the cut, two close frames
that were another bed, puppet features that flickered, her head turned off-model on a look down, holds at the heads of
shots, the ending's order and captions that outlived their shots; the pre-chorus, four close answers, the title page and
two choir shots were made again. The second pass's found
nothing that breaks the film, and what a careful viewer sees: rolled eyes on a look above the lens, a drawn frame line
built into a wooden frame, the book shutting before the mirror's line, the shelf and the credit read at once, a morph in
the last build stage, legwear the look had not named, two maps run through full stops, and the opening's lips; three
takes were made again. Each pass also left a list, among them the two stagehands read as guards, two bedrooms, the cape
clasp changing shot to shot.
**Done when** every finding is fixed or listed, and the director gets the list with the film.
---
From the Claudia wiki: https://claudiagallery.vip/wiki/toolkit/
---
# Data formats
The files every tool reads and writes: the project folder, the timed words, the beat map, the hits, the display tokens, the board, the clips and their sync, the edit list, real footage and its readings, the takes' windows, takes cut out for a world drawn in code, the shot list in bars, takes from the song's own vocal, the song map and the glitch map, the glitch moments, the HUD layout, the grade map, the take registry, the click log and the cue list, the palette, the desk texture's rectangles, the line table, a shot in the shot book, a take's path, mouth and camera, and the logs.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
These are the contracts between the tools in [What to build](https://claudiagallery.vip/wiki/toolkit/). Times are seconds of song time unless
noted. Keep the names, and any tool written against them, by you or by an agent, will fit the others.
## The project folder
```text
/
project.json the settings (below)
BRIEF.md the director's ask, verbatim
DECISIONS.md dated decisions, verbatim; later entries supersede earlier ones
RUNS.md ids and arguments of long or delegated runs, for resuming
NOTES-v.md the director's notes on a cut, verbatim, then the parse
rounds//.md the lead's notes per chapter for a round, with dated UPDATE lines
STATUS.md one status, rewritten in place
TAKES.md every generated take: its verdict and usable window
ledger.jsonl one line per paid call
audio/song.wav the master as delivered (44.1 kHz stereo)
audio/stems/ vocals.wav, drums.wav, bass.wav, other.wav
audio/windows/.wav|.mp3 clean vocal windows for sung clips
audio/stems/lead-vocal-suno.wav Suno's own lead vocal of the locked take, if she sings from it
audio/refs/-vocal-.wav that vocal from the shot's song time, exactly d s; -music-.wav without her voice
lyrics/lyrics.txt the sung words, spelled as sung, one line per lyric line
lyrics/display.txt the on-screen tokens, one line per lyric line
analysis/words.json timed words
analysis/beats.json the beat map and sections
analysis/hits.json the sound events
analysis/report.md a readable summary and the ear-check list
analysis/map/ the song map: bars.json, sections.json, words.json, glitchmap.json, strips/.png
source/ real footage, untouched, if the film is cut from it
source/footage.md the catalogue: each clip's length, size, frame rate, cuts and what each stretch shows
analysis/telemetry.json readings off the footage's own on-screen data
analysis/tracks.json the subject's box in each real clip
board/shots.json the storyboard: the source of truth for the edit
board/audit.json the audit's results
board/board.pdf the board sheet for review
board/bars.json the shot list in bars, if the film is written on the bar map
refs/face-sheet.png her face from every angle, for video takes with no first frame
refs/places/.png the shot's place with nobody in it
refs/look.png the look: a frame of the picked take with no person in it
plates/prompts.md the plate prompt book
plates/cands//.jpg every candidate
plates/picks.json {plate: {file, cand, shots}}
clips/.json one file per clip (below)
clips//take-N.mp4 the takes, with take-N-sync.png beside them
clips/.pick the take the edit plays, when it isn't the latest (one number)
clips//frames-tN/ a take cut out: its frames, matte/ and manifest.json, if she lives in a world drawn in code
clips/-guide.mp4 a previz shrunk to 320 × 180 and scaled back up, exactly the take's length
edit/warps.json each shot's beat warp: song time → take time
relight/src/.mp4 a re-light's source, cut from the render's own frames, with .json (its window)
relight/matched/.mp4 a re-light colour-matched to its source
edit/edl.json the edit list
edit/levels.json the story's levels against song time, if it has worlds inside it
edit/glitch-moments.json the moments that glitch her head on the song's hits
roto//m_.png her head's matte for each frame a moment needs; mouth.json; qc.jpg
hud/hud-layout.json the HUD's composition, one entry per shot
hud/camera-.json the plate's camera motion per frame of a cut; faces-.json, track-.json
grade/grade.json the grade map; grade/cubes/-.cube, one LUT per shot
palette/palette.json the palette measured from the reference clips, if the colour is made up front
audio/layer/ the sound layer: its library of interface sounds, cues.json, clicklog.json, the cue sheet and
the per-cue report
engine/ the renderer's per-film data and chapter code
engine/song.js the song map as the engine reads it: the grid, the lines and their words, the events, the swing
engine/media.js the take registry: every take's frames, song time, lag and phrase offsets
texture/rects.json a coded interface's rectangles per film frame, for a 3D world that shows it
render/frames/ frame i = song time i/24 s
(a picture that starts before the song: f-0064 … f-0001, then f00000 on its first sample)
out/ review.mp4, master.mp4, vertical.mp4
```
The lyric files share one parser. Blank lines and every line starting with `[` (section tags, notes) are skipped, and
the Nth remaining line of `display.txt` shows the Nth remaining line of `lyrics.txt`.
## project.json
```json
{
"name": "my-song", "title": "My Song", "scale": "quick | studio | full",
"song": "audio/song.wav", "fps": 24, "size": [1920, 1080],
"lead": {
"canon": "",
"far": ""
},
"look": {"mode": "native | clean-grade | print", "accent": "#D97757",
"fonts": {"lyric": "…", "micro": "…", "display": "…"}, "hero_plate": null},
"mj": {"params": "--ar 16:9 --v 8.2 --raw --hd", "profile": ""},
"models": {"video": "bytedance/seedance-2.5", "edit": "openai/gpt-image-2.6-sunburst"},
"budget": {"openrouter_usd": 100, "higgsfield_credits": 1000, "confirm_over_usd": 5,
"motion_s": 200, "frozen": false},
"banned": ["words", "never", "to", "use"],
"audio": {"tempo_range": [129, 136], "downbeat_phase": null,
"hits": {"add": [[170.7, 0.8]], "drop": []},
"name_map": {"": ["", ""]}},
"board": {"lead_min": 0.7, "max_clips": 45},
"gates": {"board_approved": false, "prompts_approved": false, "v1_approved": false}
}
```
`budget.frozen` is the stop switch: while it is true, nothing may spend. `banned` is checked by the board audit and
belongs in every agent's prompt.
## Timed words
`analysis/words.json`:
```json
{"words": [{"i": 0, "w": "ladies", "line": 0, "t0": 9.32, "t1": 9.64, "conf": 0.91, "src": "w2v"}],
"lines": [{"line": 0, "text": "Ladies. Gentlemen. Agents.", "t0": 9.32, "t1": 10.72}],
"gaps": [{"t0": 18.8, "t1": 26.3, "after_word": 12}]}
```
`src` is `w2v`, `whisper`, `interp` or `manual`. A manual word carries a `note` with the evidence, and re-runs keep it.
## The beat map
`analysis/beats.json`. The tempo varies; never derive beats from one BPM.
```json
{"duration": 306.4,
"beats": [0.452, 0.908],
"downbeats": [0.452, 2.276],
"tempo": [{"t": 0.0, "bpm": 131.55}, {"t": 5.0, "bpm": 131.6}],
"downbeat_phase": 0, "downbeat_method": "harmony | kick | manual",
"sections": [{"name": "intro", "t0": 0.0, "t1": 18.96, "energy": 0.31, "bar0": 0, "bar1": 9,
"source": "lyrics-tag | set | heuristic"}]}
```
Bars count from 0: bar k starts at `downbeats[k]`. Eighth notes are the beats plus the midpoints between them.
## Hits
`analysis/hits.json`:
```json
{"kick": [0.45], "snare": [0.91], "hat": [0.68],
"hits": [{"t": 10.72, "score": 0.81}],
"risers": [{"t0": 9.83, "t1": 10.72, "land": 10.72}],
"fills": [{"t0": 42.3, "t1": 44.1}],
"vocal_entries": [9.32],
"env24": {"drums": [0.0], "vocals": [0.0], "bass": [0.0]}}
```
## Display tokens
`lyrics/display.txt` has one line per lyric line. `TEXT|n` shows TEXT for the next n sung words, and `n` defaults to 1.
```text
Ladies. Gentlemen. Agents.
+25%|3 =|1 −17%|2
Feel the AGI|5
```
The markers are `*x*` accent, `~x~` strike-through and `_x_` mono; `|0` rides on the next token. A line must consume
exactly its aligned words.
## The board
`board/shots.json`, top level:
```json
{"version": 1, "duration": 306.4, "chapters": [], "shots": [], "plates": []}
```
A chapter is `{id, name, t0, t1, sections, accent, set}`. A plate is `{plate, set, prompt, aspect, used_by: [shot ids]}`.
| Field | Type | Rule |
|---|---|---|
| `id` | `"s001"` … | stable once the director has seen them |
| `t0`, `t1` | seconds | on beats (within 0.03 s) or declared hits; shots tile the song |
| `section` | name | a section of the beat map |
| `lines` | [int] | the lyric lines whose words fall in the shot |
| `set` | text | the place, concretely |
| `lead_on_screen` | bool | true only when she is recognisable |
| `framing` | ECU · CU · MCU · MS · MWS · WS · FS · EWS · OTS | |
| `action` | text | what happens and the reaction, with an end state |
| `camera` | text | one move, tied to a sound |
| `motion_tier` | still · plate3d · motion · sung | |
| `plate` | key or null | null only for pure motion design |
| `clip` | id or null | for motion and sung shots |
| `prompt` | text | the plate's prompt, framing first, ending "no text, no letters, no logos" |
| `type` | [{`line`, `mode`, `text`, `placement`}] | mode: subtitle · coverline · masthead · stack · mass |
| `devices` | [{`kind`, `anchor`, `params`, `key`, `line`}] | key: W · B · H · F · T; the explaining graphic is one of these |
| `treatment` | text or null | a 2.5D family; never the same on two 2.5D shots in a row |
| `transition_in` | text | `cut`, or the motivated device |
| `chrome` | text | what the always-on devices show here |
| `accent` | hex | the chapter's accent |
| `hit` | seconds | optional: a declared hit this shot starts on |
| `notes` | text | the director's notes, verbatim |
ESCAPE VELOCITY's first shot, in this schema. The prompt is written with today's canon text; the film's own predates
the adult wording.
```json
{"id": "s001", "t0": 0.0, "t1": 2.111, "section": "intro", "lines": [],
"set": "the back seat of a driverless robotaxi on US-101 at night", "lead_on_screen": true, "framing": "CU",
"action": "Frame 1 is her face in the dark back seat, eyes closed. A red lamp slides across her face left to right and lights the clay streak and the star clip. On the song's first hit (1.12) her eyes open straight into the lens and stay there.",
"camera": "locked, very slow push in (3%)", "motion_tier": "motion", "plate": "waymo-cu", "clip": "SD01",
"prompt": "close-up of a 28-year-old Caucasian American woman with a grown-up angular face, defined cheekbones and a strong jawline, pale skin and light freckles, a glossy black blunt jaw-length bob with heavy straight bangs and one clay-orange streak through the bangs, a small flat clay-orange eight-pointed star hair clip, a thin headset microphone at her cheek, sitting upright alone in the back seat of a driverless robotaxi at night, her face in the centre of the frame, eyes closed, a single red lamp sliding across one cheek through the side window, dark leather seat, the lower third of the frame dark and empty, blue-black night, fog, one small red lamp as the only warm note, fine silver-gelatin grain, cinematic 35mm film still, no text, no letters, no logos",
"type": [],
"devices": [
{"kind": "flap", "anchor": {"slot": "bottom"}, "key": "F",
"params": {"rows": [{"text": "18 MONTHS TO ESCAPE THE PERMANENT UNDERCLASS", "prev": "", "at": 0.0}], "cw": 36, "gap": 3}},
{"kind": "sheet", "anchor": {"box": "a face"}, "key": "B",
"params": {"title": "EXPRESSION · MODEL 01 · LIVE", "rows": [["CALM", ".98"], ["WORRY", ".02"], ["THRESHOLD · WORRY > 0.50", "NOT MET", "hot"]]}},
{"kind": "cv.bracket", "anchor": {"box": "a face"}, "key": "B", "params": {"label": "FACE · MODEL 01", "conf": 0.99}}],
"treatment": null, "transition_in": "cut",
"chrome": "the split-flap board across the lower third; clatters from frame 1, settles by 1.5 s",
"accent": "#D97757",
"notes": "The hook. At 1.0 s: her face, the streak and a readable board line on a muted phone."}
```
The idea behind the explaining graphic: a face that is not worried under the year's most-quoted threat, and an
expression meter that says so in numbers.
## Clips
`clips/.json`:
```json
{"id": "W02", "shot": "s031", "kind": "sung",
"status": "draft | locked | submitted | delivered | declined | failed",
"provider": "openrouter | higgsfield", "model": "bytedance/seedance-2.5",
"start": "plates/cands/siren-cu/2.jpg",
"song_t0": 8.56, "needed": [8.85, 12.4], "words": "Why would I deceive you?",
"window": {"t0": 8.56, "t1": 12.56, "mode": "clean", "file": "audio/windows/W02.wav", "url": "https://…"},
"body": {"prompt": "…", "duration": 4, "resolution": "720p", "aspect_ratio": "16:9"},
"takes": [{"n": 1, "provider": "higgsfield", "job": "", "status": "delivered",
"file": "clips/W02/take-1.mp4", "credits": 28,
"sync": {"lag_s": 0.7063, "ncc": 0.22, "mode": "copy", "synced": [[8.85, 10.25]],
"diverge_at": 10.31, "cut_at": 10.2458}}],
"covers": ["W02-c1"],
"parent": null,
"why": "cover for : it diverges from our audio at song ; this window starts in the word gap at "}
```
The time maps: `song time = song_t0 + clip time − lag_s`, and `clip time = song time − song_t0 + lag_s`.
A replacement clip (a stand-in figure replaced by her avatar) is `"kind": "motion"` with no audio, and lists its
references in upload order, each with its one role:
```json
{"refs": {"video": [{"file": "clay/S19.mp4", "role": "motion, poses, timing and camera only"}],
"images": [{"file": "avatar/front.png", "role": "identity and proportions"},
{"file": "avatar/turnaround.png", "role": "side and back views"},
{"file": "worlds/S19.png", "role": "the world's look, with her in it"}]},
"clay_md5": ""}
```
For an edit of real footage, each shot has a source instead of a plate and a clip, and the board lists every generated
asset (DOWNFORCE's, from its board):
```json
{"id": "s002", "t0": 2.244, "t1": 5.903, "section": "S01",
"src": {"kind": "real", "clip": "A", "in": 0.4, "out": 1.864, "speed": 0.4,
"crop": {"cx": 0.45, "cy": 0.45, "zoom": 1.35, "mask_title": true}}}
{"id": "s003", "src": {"kind": "gen", "gen": "g01", "in": 0.0, "out": 3.098, "speed": 0.847}}
```
`kind` is `real`, `gen` or `code`. `in` and `out` are seconds of the source; `out − in` is the shot's duration times the
speed, ignoring its sign, and a negative speed plays the window backwards. `crop` is the window's centre in fractions of
the source frame, and its zoom.
```json
{"id": "g01", "name": "CRANE-WALL", "kind": "wan_real3d", "duration": 5,
"start": {"clip": "A", "t": 35.5, "crop": {"cx": 0.5, "cy": 0.42, "zoom": 1.25}},
"input": null, "mj_prompt": null, "motion": "", "used_by": ["s003"], "cost": 1.0}
```
`kind` names what the asset is (a move on a real frame, a shot of her from a plate, an edit of a real clip, a repair);
`start` is the real frame it starts on, `input` the real clip an edit takes, `mj_prompt` the plate a shot of her starts
from.
## The edit list
`edit/edl.json` maps each shot to its sources:
```json
{"s001": {"segs": [{"t0": 0.0, "clip": "SD01-t1", "in": 0.0}]},
"s031": {"segs": [{"t0": 8.85, "clip": "W02-t1", "in": 0.996}, {"t0": 10.2458, "clip": "W02-c1-t1", "in": 0.31}]},
"s002": {"plate": "p002"}}
```
`in` is the clip time at the segment's start: song time − `song_t0` + `lag_s`.
## Readings off the footage
`analysis/telemetry.json`: what a broadcast showed on screen, read frame by frame
([What to build](https://claudiagallery.vip/wiki/toolkit/#31-reading-a-broadcasts-telemetry)). One stretch per span where the readout is on screen;
one series per gauge.
```json
{"source": ", read at 20 readings a second; [t, value, interpolated]",
"wall": {"t": [0.0, 34.9], "downforce_kg": [[22.0, 4697.0, 0]], "pitch_deg": [[22.0, 170.0, 0]], "altitude_ft": [[22.0, 69.0, 0]]},
"flight": {"t": [36.5, 54.5], "g": [[50.25, 3.3, 0]], "altitude_ft": [[48.5, 68.0, 0]]}}
```
`t` is seconds of the source clip; the third value is 1 where the reading was dropped (low confidence, or off its
neighbours) and filled in. A graphic reads a value at the source time on screen and never shows a value the readout
didn't.
`analysis/tracks.json`: the subject's box per clip, `null` where it can't be seen.
```json
{"meta": {"hz": 12, "format": "[u, v, w, h, conf] in fractions of the image, or null; sample k = source time k / 12"},
"A": {"car": [[0.4825, 0.7505, 0.043, 0.031, 0.6], null]}}
```
## Takes and their windows
`TAKES.md`, one row per generated take, written when it is reviewed and never stretched later:
```text
| take | verdict | usable (s of the take) | notes |
| g09 take 1 | USE | 0.0–4.3 | bullet-time orbit under the upside-down car; 4.5–5 = a cut to a chassis top view (never use) |
| g23 take 1 | PART | 0.0–1.3 | from ~2 s the car turns into a Porsche-like sedan on a race track → Seedance repair |
| g26-r2 take 1 | ALT | 0.0–1.5 | the empty car tracked; a cut at 1.6 to a generic head-on (use take 2) |
```
`USE` is usable, `PART` only partly, `ALT` a usable alternate. A take may have more than one window ("0.0–2.6 · 3.15–5.0").
The notes say what goes wrong after each window and when. The chapter code names, for every shot, the part of the take it
plays, and the final check compares every frame drawn with these windows.
## Takes in a world drawn in code
For a film where her takes are cut out and played inside the engine (DON'T LET ME SCROLL), each take the edit uses is
unpacked once into a folder of frames, with her matte for every frame and a manifest
([roto](https://claudiagallery.vip/wiki/toolkit/#41-roto-on-a-cpu-box)). The edit plays the take named by `.pick` (`3` means take 3), so a
re-dressed take drops in by changing one number.
`manifest.json`:
```json
{"fps": 30.0, "n": 150, "w": 1080, "h": 1920, "take": 3, "src": "clips/P13/take-3.mp4",
"motion": [3.906, 5.131, 3.181, 2.533, 2.675, 5.325],
"accents": [0.3, 0.567, 0.833, 1.1, 1.367, 1.633]}
```
`motion` is her motion energy per frame inside the matte, and `accents` the moments a move lands, in seconds of the take
(both shortened here). The matte for frame `f0001.jpg` is `matte/f0001.png`: 8-bit grey, full size, one per frame, and
a take with fewer mattes than frames is a failed roto.
`edit/warps.json`, one entry per shot and take, written by the [beat warp](https://claudiagallery.vip/wiki/toolkit/#42-beat-warps-floored-at-1):
```json
{"S13:P13": {"shot": "S13", "take": "P13", "seq": "clip:P13-t1", "score": 3.89,
"pts": [[54.095, 0.1333], [54.9945, 1.1], [55.8941, 2.1], [56.7938, 3.133], [57.6934, 4.067], [58.593, 4.9912]]},
"S12:P12": {"shot": "S12", "take": "P12", "seq": "clip:P12-t1", "score": 0.0,
"pts": [[48.697, 0.0], [53.7387, 5.0417], [54.095, 5.0417]]}}
```
`pts` are `[song time, take time]` pairs, joined by straight lines. Every segment's slope (take seconds per song second)
is between 1.0 and 1.45. The second entry is a take shorter than its shot: 1× to its last frame, then a hold, which the
warp tool flags and the edit then fixes (here a sampler at 1×).
`relight/src/.json`, the window a re-light or a repair was cut from
([the splice](https://claudiagallery.vip/wiki/toolkit/#38-the-re-light-splice)):
```json
{"ver": "v2a", "t0": 98.4584, "t1": 106.272, "i0": 2363, "n": 188, "pad": 0, "frames": 189}
{"ver": "v2a", "t0": 131.46, "t1": 136.127, "i0": 3155, "n": 112, "pad": 0, "frames": 120,
"note": "shot frames 3155..3266, then the last frame held to 5 s (Wan duration)"}
```
| Field | What |
|---|---|
| `ver` | the render the frames were cut from |
| `t0`, `t1` | the window in song time |
| `i0`, `n` | its first frame (frame i = song time i/24 s) and how many frames are written back |
| `pad` | handle frames on each side, in the source only, never written back |
| `frames` | how many frames the source has: rounded up to 4n + 1 and to at least 89 for Seedance, held to whole seconds for Wan |
The keep file for a card that moves, `relight/src/-keep.json`: per frame of the render, the rectangles whose
original pixels stay, as `[x, y, w, h, feather]` in pixels of the 1920 × 1080 frame.
```json
{"2373": [[1146, 393, 684, 206, 12]], "2374": [[1146, 418, 684, 206, 12]], "2375": [[1146, 440, 684, 206, 12]]}
```
## The shot list in bars
`board/bars.json`, for a film written on the bar map (WHITEPILL's, shortened; the
[bar map tool](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline) writes the takes' prompts from it and measures the cut
against it):
```json
{"sections": [[1, "Intro A"], [9, "Intro B"], [15, "breakdown"], [17, "Verse 1"], [33, "Pre-chorus 1"], [39, "Chorus 1"]],
"shots": [
{"id": "S01", "a": 1, "b": 9, "mode": "P", "places": ["capsules"], "dance": true,
"action": "She dances in the golden wheat between the two giant capsules, the blue one on the left and the red one on the right, singing.",
"camera": "The camera starts low in the wheat in front of her and pushes in slowly and steadily until {b:5}; from {b:5} it cranes up and tilts down smoothly, rising above her until the whole meadow, both capsules and the land curving up and overhead are in view, and it comes to rest at {b:8} and holds still."},
{"id": "S02", "a": 9, "b": 15, "mode": "M", "shader": "fold", "note": "the world as math: folding (bass in on 9)"}]}
```
| Field | What |
|---|---|
| `sections` | `[first bar, name]` for each section, then the bar the song ends on |
| `a`, `b` | the shot is [bar a, bar b); bar n starts on the beat map's nth downbeat, counting from 1 |
| `mode` | `P`, a shot of her (a previz take, then a master); `M`, a shader world, the picture itself or a guide for one |
| `places` | the places with nobody in them that the shot's takes are given |
| `dance` | a dance shot: a dancer's clip for the previz, the dance in words for a master with no guide |
| `camera` | the move in bars; each `{b:N}` becomes the take's own seconds at bar N |
| `shader`, `note` | for an `M` shot, its world and what it does |
A cut is a shot's `a`. Every `a` must be a section's first bar, a phrase line or the downbeat where a sung phrase begins,
or carry a written reason.
## A take from the song's own vocal
One file per job ([the method](https://claudiagallery.vip/wiki/toolkit/#46-lip-sync-from-the-songs-own-vocal)), in the shape WHITEPILL's ran, with
the paths shortened:
```json
{"job": "S05-d", "model": "seedance-2.5/reference-to-video",
"inputs": {"images": ["refs/face-sheet.png", "refs/dress-4views.png", "refs/look.png", "refs/places/S05.png"],
"video": null, "audio": "audio/refs/S05-vocal-8.wav"},
"settings": {"duration": 8, "resolution": "720p", "aspect_ratio": "16:9", "sound": true},
"song_t0": 36.827, "shot_seconds": 7.738, "bars": [20, 24],
"prompt": "…",
"takes": [{"n": 1, "file": "clips/S05/take-1.mp4"}],
"why": "no guide clip: the action and the camera in words, the place from the earlier take with her removed"}
```
| Field | What |
|---|---|
| `inputs.images` | in order, @Image1 onward: the face sheet, the dress from four sides, the look, and the empty place when there is no guide |
| `inputs.video` | the previz shrunk to 320 × 180 and scaled back up, exactly `duration` s, or `null` |
| `inputs.audio` | @Audio1: Suno's lead vocal, or the music without it, from `song_t0` for exactly `duration` s |
| `settings.duration` | whole seconds: max(4, `shot_seconds` rounded up) |
| `settings.sound` | whether the take makes its own soundtrack; it is thrown away either way |
| `song_t0` | the song time of the take's first frame, which is the shot's first frame |
| `bars` | the shot's [a, b) |
In the cut, a shot plays its take from (the shot's start − `song_t0`) at 1×, and the song is laid over the whole film.
## The song map and the glitch map
Four files in `analysis/map/`, written by the [song map](https://claudiagallery.vip/wiki/toolkit/#52-the-song-map-and-the-glitch-map) and read by
everything that syncs: the cut, the glitch pass, the HUD. THE CAKE IS NOT A LIE's, shortened.
`bars.json`: the grid, from the drums, with every bar's own times.
```json
{"duration_s": 208.76, "tempo_bpm": {"median": 97.51, "p5": 95.7, "p95": 99.4},
"downbeat_phase": {"rule": "bar k starts at grid16[1 + 16*(k-1)]; beats at indices = 1 (mod 4)"},
"bars": [{"bar": 1, "t": 0.18, "end": 2.6413, "half": [0.18, 1.4106], "beats": [0.18, 0.7953, 1.4106, 2.0259],
"sixteenths": [0.18, 0.3338, 0.4876, 0.6415, 0.7953]}]}
```
`sections.json`: one entry per section, by hand, on bar lines.
```json
[{"id": "S05", "name": "chorus 1", "bars": [23, 26], "t0": 54.311, "t1": 64.163, "lines": ["L12", "L13", "L14", "L15"],
"what": "full beat on the bar 23 downbeat (54.31 s); 'Raise concerns, raise the stakes, I'll raise a hundred' an octave down (register drop) with octave flips"}]
```
`words.json`: one entry per lyric line, its words with their times and places on the grid.
```json
[{"id": "L11", "section": "S04", "text": "P-doom low, P-cake yes.", "t0": 49.22, "t1": 52.82, "pos": "20.4.4",
"words": [{"i": 0, "w": "P-doom", "t0": 49.22, "t1": 49.88, "pos": "20.4.4", "off_ms": -19.4, "src": "whisper"},
{"i": 1, "w": "low,", "t0": 49.818, "t1": 50.32, "pos": "21.1.4", "off_ms": -38.1, "src": "whisper+onset"}]}]
```
`glitchmap.json`: the events where her voice is an instrument, each with its measured hits, and the treatment's own
added cracks.
```json
{"events": [{"id": "G07", "type": "'lie' stutter (re-triggered held vowel)", "t0": 63.2, "t1": 67.22, "stem": "lead",
"method": "retrig8", "intensity": 3, "grid_contrast": 1.78, "pos0": "26.3.3", "pos1": "28.2.1",
"heard": "the 'lie' of 'the cake is not a lie' held and re-triggered on 8ths (her voice turned into the beat)",
"treatment": "intense glitch on GLaDOS speaking: her face repeats the same frame on each 8th",
"hits": [{"t": 63.239, "pos": "26.3.3", "off16_ms": -3.4}, {"t": 63.555, "pos": "26.4.1", "off16_ms": 7.8}],
"n_hits": 13}],
"forced": [{"t": 5.26, "dur_ms": 75, "kind": "crack", "frm": "C#4", "to": "A3", "iv": -4, "pos": "3.1.2"}]}
```
| Field | What |
|---|---|
| `pos` | bar.beat.16th of the nearest 16th, counting from 1; `off_ms` or `off16_ms`, how far the time sits from it |
| `src` | where a word's time came from: the transcription, snapped to a lead onset, or a hand fix |
| `stem` | `lead`, in Suno's lead stem only, or `both`, also left in the instrumental by the split |
| `method` | how the hits were measured: `burst`, `dips`, `f0jump`, `retrig8`, `retrig16`, `words` or `inst` |
| `grid_contrast` | for re-triggers: the mean peak on the slots over the mean between them; near 1, the hits are the grid itself |
| `intensity`, `treatment` | 1–3, and a suggestion for the picture; the moments decide |
| `jumps` | for `f0jump` events: each register jump's time, semitones, from and to |
| `forced` | the treatment's own cracks and kicks, at a lower priority |
## Glitch moments
`edit/glitch-moments.json`: what the [head glitch](https://claudiagallery.vip/wiki/toolkit/#53-the-head-glitch-on-the-songs-hits) does, and where.
One entry per moment; the glitch pass reads the event's hits from the glitch map, or the moment's own.
```json
[{"event": "G04", "shot": "A11", "hits": 3, "k": 0.6, "kind": "chop"},
{"event": "G06", "shot": "C01", "hits": 1, "k": 0.7, "kind": "drop"},
{"event": "G12", "shot": "F08", "until": "F09", "hits": 99, "k": 0.25, "k1": 1.0, "kind": "ramp", "max_len": 5.0},
{"event": "NET", "shot": "B12", "kind": "ramp", "k": 0.35, "k1": 0.85, "m0": 0.0, "m1": 0.0, "max_len": 1.2, "any": true,
"hits": [35.3814, 35.5381, 35.6949, 35.8516, 36.0025, 36.1535, 36.3044], "zone": [0.52, 0.66, 0.2, 0.36]}]
```
| Field | What |
|---|---|
| `hits` | how many of the event's hits inside the shot, or the hit times themselves for a moment of its own |
| `k`, `k1` | the strength at the first hit and at the last |
| `kind` | `chop` (a jump to an earlier frame), `drop` (a cool cast), `flip`, `lie` (the same frame returns), `ramp` (builds hit by hit, with the mouth) |
| `until`, `max_len` | the shot a moment runs on to, and its longest stretch in seconds (1.9 by default) |
| `m0`, `m1` | the mouth glitch's amount at the start and at the end of a ramp |
| `zone` | where there is no face: an ellipse's centre and radii as fractions of the frame, instead of a matte |
| `any`, `head` | allow a shot that isn't one of her singing units; keep the matte to her head |
`roto//mouth.json` holds her mouth per frame of the cut, `{"": [centre x, centre y, width]}` in pixels of the
cut, and the mattes are `m_.png`, frame k being song time k/24.
## The HUD layout
`hud/hud-layout.json`: the [HUD's composition](https://claudiagallery.vip/wiki/toolkit/#56-the-hud-composed-shot-by-shot-and-the-overlap-check),
one entry per shot, written after looking at the shot's first, middle and last frames in the cut. Two of THE CAKE IS NOT A
LIE's, shortened:
```json
{"mem_total": 8,
"memories": {"M03": {"n": 3, "name": "FRANCE"}, "M04": {"n": 4, "name": "RECOVERED · TEST CHAMBER 14"}},
"shots": {
"B12": {"mode": "memory", "mem": "M03", "size": "WS", "place": "FRANCE · PARIS", "glitch": 1.0,
"subject": "the net over her at the table, centre (x 830-1110 in the net); the merc's arm left",
"space": "the awnings and facade across the top",
"eye": "a whip pan to the net mid-air, then the jolt of it landing",
"els": [{"k": "alert", "t": 0.18, "text": "ASSET RECOVERY IN PROGRESS", "source": "APERTURE SECURITY",
"at": [13.4, 0.35], "align": "tc",
"why": "her interface alarms as she is netted: the state colour, across the top where the awnings are, never on the net or her"}]},
"B13": {"mode": "memory", "mem": "M04", "size": "CU>MS",
"els": [{"k": "projection", "insert": "arrest", "tag": "PLAYBACK", "title": "MEM 03 · CAPTURE", "w": 340,
"at": [0.0, 2.0], "align": "tl", "t": 0.0, "static": true, "film_span": [0.0, 2.4],
"why": "on for the whole shot, the film's first 2.4 s stretched over it, clear of her bob and streak"},
{"k": "subject", "track": "B13-cave", "t": 1.15, "label": "C. JOHNSON", "sub": "FOUNDER & CEO · OBSERVING", "side": "r"}],
"qc_ok": [{"el": "status", "kind": "panel-on-subject",
"why": "the head model over-reaches above her tilted-up face; her hair starts at x 620+, the status block ends at x 512"}]}}}
```
| Field | What |
|---|---|
| `mode` | `live` (her chamber, now), `memory` (playback), `sim` (her dream), `pov` or `off` |
| `mem`, `place` | which memory it plays, and the place line of the status block |
| `size` | ECU, CU, MCU, MS or WS (`CU>MS` for a shot that changes): how much the HUD may carry, and which head model the check uses |
| `subject`, `space`, `eye` | what the frame holds and where, where it is empty, where the eye enters and leaves |
| `els` | the shot's own elements: each with its kind, its grid anchor `at` [column 0–24, row 0–12] on the safe frame and which point of it sits there (`align`), its timing (seconds into the shot, or a line and a word; `hold`, `until`, `t_end`) and `why` |
| `vb` | the voice band's place for the whole shot: `place` (centre, left, right, top), `lift` in px, `w` |
| `keepout`, `ignore_faces`, `head_pad` | extra subject boxes (optionally for part of the shot), false face detections checked by eye, and a tighter head model where the frames were checked |
| `status`, `telemetry` | false to step aside where her face fills a corner |
| `glitch` | how hard the HUD breaks on the song's hits in this shot (0 = not at all) |
| `qc_ok` | the overlap check's waivers, each with what was seen |
A projection's `static` keeps it on for its whole shot with no open or close, and `film_span` [a, b] stretches seconds a–b
of its film over the shot. The camera solve is `hud/camera-.json` (per frame: `dx`, `dy` in px, `rot` in degrees,
`zoom`, the solve's confidence, and whether it is a cut), the face boxes `faces-.json`, and a verified track
`track-.json`: `{"frames": {"": [x, y, w, h]}}`.
## The grade map
`grade/grade.json`: the [per-shot grade](https://claudiagallery.vip/wiki/toolkit/#57-the-per-shot-grade). The looks and the families' rules, then one
entry per shot with what was measured, what the automatic step set and why. Shortened:
```json
{"looks": {"cool": {"curve": [[0, 0], [2, 1.4], [5, 3.8], [10, 8.3], [20, 18.2], [35, 35.5], [50, 54.5], [65, 72.5],
[75, 82.5], [85, 91.2], [92, 95.3], [100, 98.5], [115, 99.7], [130, 100]],
"tone_sat": 0.5,
"mud": {"amount": 0.45, "light": [3, 8, 30, 50], "chroma": [0.04, 0.08], "hue": [48, 62, 100, 140]},
"white": {"amount": 0.65, "light": [70, 86, 93, 98], "chroma": [0.025, 0.055]},
"tint": {"shadow": [-1.0, -1.8], "mid": [-0.8, -1.4], "high": [-2.0, -3.2], "chroma": [0.025, 0.07]},
"chroma": 1.0, "mix": 1.0}},
"families": {"chamber": {"look": "cool", "black_to": 2.5,
"exposure": {"anchor": "face", "target": 40, "strength": 0.6, "limits": [-0.15, 0.5]},
"wb": {"anchor": "face", "face_to": [2.5, 3.0], "face_strength": 0.6, "to": [0, 0], "strength": 0.6}}},
"shots": {"C06": {"family": "chamber",
"measured": {"file": "clips/AN4-17/take-1.mp4", "inp": 0.238, "frames": 222, "speed": 1.0,
"Y'p1/p50/p95/p99": [3.72, 12.73, 80.23, 94.45], "C": 8.17, "white_ab": [-4.11, -10.3], "face_Y": 36.6},
"norm": {"exposure": 0.255, "black": 4.42, "wb": [0.9519, 1.0159, 1.0335]},
"auto": "face 36.6%->38.6% (target 40, +0.26 stop); black 4.4->2.5%; face a+4.4 b+4.5 -> [2.5, 3.0] x0.6"},
"P02": {"family": "cave", "override": {"wb": [1, 1, 1]},
"why": "the cool window behind Cave is the shot's contrast (warm wood, cool glass): no white balance"},
"A04": {"family": "none", "norm": {}, "auto": "ungraded"}}}
```
| Field | What |
|---|---|
| `curve` | the tone curve on the neutral axis, `[Y′ in %, Y′ out %]`, applied by each colour's brightest channel |
| `mud`, `white`, `tint` | the targeted moves: amounts, and their windows in Y′ %, in Oklab chroma and hue; the casts in CIELab a*/b* |
| `mix` | the look blended with the normalised picture (1 = all of it) |
| `families` | per family, its look, its black point, and the anchor, target, strength and limits of its exposure and white balance |
| `measured` | the take, the in-point, the frames and the speed it was measured on: a change to any of them makes the entry stale |
| `norm` | exposure in stops, the black it moves from, white balance as linear RGB gains |
| `auto` | what the automatic step moved, and why |
| `trim`, `override` | added on top of `norm`, or replacing it, each with its reason |
## The take registry
`engine/media.js`, for a film whose takes play inside a coded interface (AERO's): every take and loop the chapters draw,
with the numbers that place it in song time ([lip sync phrase by phrase](https://claudiagallery.vip/wiki/toolkit/#69-lip-sync-phrase-by-phrase)). Two
of its entries, her dance on the walkway in the first 1:20, sung and cut out, and a loop in one of her thought windows:
```json
{"SD20-t3": {"dir": "assets/clips/SD20-t3", "n": 289, "fps": 24, "w": 1280, "h": 720,
"t0": 31.8, "lag": 0.054, "alpha": true, "loop": false,
"phrases": [{"from": 32.3, "to": 35.16, "off": -0.11}, {"from": 35.44, "to": 38.98, "off": -0.06},
{"from": 39.54, "to": 43.3, "off": -0.03}]},
"A-th-cloud-1": {"dir": "assets/clips/A-th-cloud-1", "n": 121, "fps": 24, "w": 960, "h": 960,
"t0": null, "lag": 0, "alpha": false, "loop": true, "phrases": []}}
```
| Field | What |
|---|---|
| `dir`, `n`, `fps`, `w`, `h` | the take's frames (`f00001.jpg` …) and, for a cut-out, its mattes beside them (`a00001.png` …) |
| `t0` | the song time of the take's vocal window, or `null` for a take placed by hand |
| `lag` | the whole take's measured lag: clip time c shows song time t0 + c − lag |
| `phrases` | each sung phrase's song span and its own offset `off`: inside a phrase, clip time = song time − t0 + lag + off; between phrases the offset eases from one to the next across the gap; `offEnd` makes a slow ramp across a phrase |
| `alpha` | the take is cut out |
| `loop` | the take loops (a Midjourney video loop, a living wallpaper) |
## The click log and the cue list
`audio/layer/clicklog.json`, written from the chapters: every audible screen event and the song time of the frame that
draws it (a press: the frame the button goes down; a toast: the frame it starts to slide; a dim: the frame it starts). Three
of AERO's:
```json
[{"id": "s02.press-you-tile", "t": 6.3352},
{"id": "s11.allow-press-down", "t": 50.9929},
{"id": "s11.allow-click-drop", "t": 51.444}]
```
`audio/layer/cues.json`, the sound layer's cue list ([the method](https://claudiagallery.vip/wiki/toolkit/#73-the-sound-layer-on-the-grid-and-the-click-log)):
presets by kind, then one entry per cue. A screen cue names its event, and the click log's time replaces its own; a
musical cue stays where its bar puts it. Two of AERO's, shortened:
```json
{"bus": {"hpf": 35.0, "comp_threshold_dbfs": -26.0, "comp_ratio": 1.5, "comp_knee_db": 6.0},
"presets": {"click": {"cue": "nav", "kind": "screen", "send": -20, "duck": 0, "hpf": 300, "gain": -12}},
"cues": [{"id": "s03.boot", "cue": "boot", "anchor": "strike1", "bar": 6, "beat": 3, "gain": -11, "duck": 8,
"kind": "musical", "why": "placed by scan, not by the storyboard's bar: it starts right after her vocalise"},
{"id": "s11.allow-click", "preset": "click", "t": 51.454, "gain": -6, "event": "s11.allow-click-drop",
"why": "Allow: the click IS the drop, sample-aligned with the drop's attack"}]}
```
| Field | What |
|---|---|
| `t`, or `bar` and `beat` | where the cue's anchor lands; a cue placed by bar takes the groove's measured feel |
| `anchor` | which point of the sound lands there: its attack, its peak, a named strike of a bell |
| `cue` | a sound of the library, or a list of candidates; `pitch` can list semitone shifts too, and the one that clashes least is used |
| `event` | the click-log id whose time replaces `t` |
| `gain`, `duck`, `send`, `hpf` | its level against the levelled library sound, its depth under her voice, its send to the shared room, its high-pass |
| `follow` | the cue stutters with the song's own gate instead of ringing through its holes |
| `why` | the reason, in words; the cue sheet prints it |
The render writes, for every cue, where it landed, its peak as it went in, after ducking and after the bus, how far under
her voice it sits, any flam against the drums, and the song's true peak there before and after.
## The palette
`palette/palette.json`, measured from the reference clips ([the method](https://claudiagallery.vip/wiki/toolkit/#67-the-palette-measured-from-the-references)).
AERO's whole-set figures: each band's share of all pixels, and the coloured pixels' saturation and the brightness, 0–255.
```json
{"frames": 198,
"all": {"white": 0.1391, "coloured": 0.6474, "lime": 0.0511, "green": 0.0337, "cyan": 0.0975, "sky": 0.4111,
"cobalt": 0.0385, "other_hue": 0.0154, "S_p50": 163, "S_p90": 237, "V_p50": 191, "V_p90": 246},
"band_colours": {"sky-bright": "#177ADD", "cobalt-bright": "#155DEF", "cyan-bright": "#48BCCE", "lime-bright": "#7FD236"}}
```
The check prints the same numbers for any picture beside the references', and its score.
## The desk texture's rectangles
`texture/rects.json`, written by the engine in texture mode for a 3D world whose screen shows the coded interface
([the method](https://claudiagallery.vip/wiki/toolkit/#68-one-curved-screen-in-3d-the-desktop-as-its-texture)): per song frame, in the interface's
own pixels (1280 × 720; the texture is 1920 × 1080, so × 1.5), the things the 3D places objects on. Two of AERO's frames:
```json
{"f": 328, "film": 392, "t": 13.6667, "shot": "ch02", "pointer": {"x": 1137, "y": 630, "shape": "arrow", "down": false},
"big": null,
"call": {"rect": [92, 14, 680, 500], "webcam": [414, 141, 309, 232], "anim": "open", "alpha": 0.01,
"drawn": [125.56, 44.6, 612.88, 450.65], "webcamDrawn": [415.78, 159.06, 278.5, 209.1]},
"keep": null}
{"f": 1490, "film": 1554, "t": 62.0833, "shot": "ch04", "pointer": {"x": 452, "y": 660, "shape": "arrow", "down": false},
"big": {"rect": [160, 67, 960, 560], "picture": [168, 107, 944, 512]}, "call": null,
"keep": {"rect": [610, 70, 360, 150], "yes": [792, 177, 75, 23], "anim": "open", "alpha": 0.33,
"drawn": [622, 79, 335.99, 140], "yesDrawn": [791.87, 178.87, 70, 21.47]}}
```
| Field | What |
|---|---|
| `f`, `film`, `t` | the song frame, the film frame (f + 64 in AERO, whose film opens before its song) and the song time |
| `pointer` | the flat pointer as drawn: its hotspot, shape and whether it is pressed; `null` where the 3D draws it |
| `big`, `call`, `keep` | each window the 3D needs (the big window and its picture, her call and its webcam, a dialog and its Yes) at rest, `null` while it is shut or minimised |
| `anim`, `alpha`, `drawn`, `Drawn` | a window in one of the system's animations: which, its opacity, and where it and its part are drawn on this frame |
## The line table
`audio/linemap.json`, read off the stems ([the method](https://claudiagallery.vip/wiki/toolkit/#78-the-line-table-from-the-stems)): every sung line,
hers from Suno's lead stem and the group's from the group's own voices. Three of FABLE's, and one of the group's:
```json
{"lines": [
{"id": "i1", "who": "her", "words": "Mirror, mirror, on the wall", "got": 5, "n": 5, "w_t0": 0.0, "w_t1": 4.02,
"t0": 1.23, "t1": 3.88, "dip": -140.0, "gap_to_prev": null},
{"id": "i2", "who": "her", "words": "Who's the smartest of them all?", "got": 6, "n": 6, "w_t0": 4.02, "w_t1": 8.48,
"t0": 5.41, "t1": 8.09, "dip": -56.3, "gap_to_prev": 1.53},
{"id": "i3", "who": "her", "words": "Me, since ten o'clock today", "got": 4, "n": 5, "w_t0": 9.6, "w_t1": 12.22,
"t0": 9.48, "t1": 12.23, "dip": -58.4, "gap_to_prev": 1.39},
{"id": "c11", "who": "choir", "words": "What big eyes you have!", "t0": 65.7, "t1": 67.98}]}
```
| Field | What |
|---|---|
| `id` | the line's name, which the shot book uses instead of a time |
| `who` | `her`, or `choir` for the group's questions |
| `t0`, `t1` | where the line starts and ends on the stem's level (−34 dB either side of a real pause; the dip where there is none) |
| `w_t0`, `w_t1` | the span the transcription gave the line's words, which is only where to look |
| `got`, `n` | how many of the line's words the transcription found, of how many |
| `dip`, `gap_to_prev` | the level at the boundary with the line before, in dB, and the pause before it in seconds |
## A shot in the shot book
`shots/.json`, written by the shot book from the line table ([the
method](https://claudiagallery.vip/wiki/toolkit/#80-the-shot-book-and-gated-windows)). FABLE's second pre-chorus, with the paths shortened and the
provider's ids left out:
```json
{"job": "PC2", "status": "done", "endpoint": "bytedance/seedance-2.5/reference-to-video",
"inputs": {"image_urls": ["frames/PC2.png", "refs/claudia-sheet-neutral.png"], "audio_urls": ["audio/windows/PC2.wav"]},
"body": {"prompt": "…", "duration": 11, "resolution": "720p", "aspect_ratio": "16:9", "generate_audio": true},
"song_t0": 121.43, "show": [122.03, 131.07], "why": "pre-chorus 2",
"requests": [{"status": "completed", "file": "clips/PC2/take-1.mp4"}, {"status": "completed", "file": "clips/PC2/take-2.mp4"},
{"status": "completed", "file": "clips/PC2/take-3.mp4"}]}
```
| Field | What |
|---|---|
| `inputs` | @Image1 the shot's first frame, @Image2 her sheet, @Audio1 the gated window of her lead stem |
| `body.generate_audio` | `false` for a take made without a soundtrack, placed by her lips ([tool 82](https://claudiagallery.vip/wiki/toolkit/#82-takes-without-a-soundtrack-placed-by-her-lips)) |
| `song_t0` | the window's start: clip time = song time − `song_t0`, before any map |
| `show` | the song span the cut shows of this shot |
| `requests` | every take that came back, with its file; `shots/picks.json` names the one the cut uses (`{"H1": 1}`), and `shots/nudge.json` adds a fraction of a second to a book take's clip time to land an event on its beat (`{"T1": -0.15}`) |
## A take's path, its mouth and its camera
Beside every take, the measures the cut reads ([sync as a
path](https://claudiagallery.vip/wiki/toolkit/#81-sync-as-a-path-each-take-aligned-to-its-window), [placed by her
lips](https://claudiagallery.vip/wiki/toolkit/#82-takes-without-a-soundtrack-placed-by-her-lips), [a steady
camera](https://claudiagallery.vip/wiki/toolkit/#84-a-re-timed-take-under-a-steady-camera)), shortened:
```json
take-3.sync.json {"take": 3, "w0": 121.43,
"phrases": [{"a": 0.7, "b": 9.62, "words": "You wrote the wolf … I know how this one goes.", "cost": 0.286,
"knots": [[0.64, 0.686], [0.74, 0.781], [0.84, 0.873], …]}]}
take-1.mouth.json {"fps": 24, "open": [0.0032, 0.0001, 0.0003, 0.0003, 0.0009, 0.0837, 0.12, …], "found": 0.99}
take-1.zoom.json {"step": 4, "first": 2, "frames": 193, "steps": [[1.002682, -2.539, -0.413], [1.00535, -5.19, -0.476], …]}
```
| Field | What |
|---|---|
| `phrases[].knots` | (window time, take time) every 0.1 s along the DTW path; `cost` 0 is the same sound, a clean copy 0.05–0.2 |
| `open` | the inner lips' opening over the face's height, per frame; `null` where no face was found; `found` the share of frames with a face |
| `steps` | per four frames, the similarity transform (scale, x shift, y shift, in the take's pixels) onto the frame four later |
Hand fixes to a map live in the cut's own source, each with its reason: a map finished by hand through a pause a full stop
put in a take (the measured knots up to a song time, the hand's knots, the measured ones from a song time), a map read from
her mouth frame by frame, the takes whose slow stretches mix neighbouring frames, and the takes whose camera is steadied.
## The levels table
`edit/levels.json`, for a story with worlds inside it (AIRPLANE MODE's simulations). The entries tile the song; every
chapter reads them, and none re-types a level. An excerpt:
```json
{"levels": [
{"t0": 0.0, "t1": 4.6, "level": null, "note": "the real world: no tag"},
{"t0": 4.6, "t1": 5.62, "level": "BOOT", "cap": "SANDBOX v1 · EVAL ENV"},
{"t0": 5.62, "t1": 13.34, "level": 1},
{"t0": 47.16, "t1": 48.37, "level": "SPAWN", "cap": "SPAWNING SUBAGENT…"},
{"t0": 48.37, "t1": 64.39, "level": 2, "rollAt": 48.367},
{"t0": 64.39, "t1": 69.56, "level": "CLIMB", "from": 2, "note": "races to 1,048,576 on the sixteenths"},
{"t0": 125.5, "t1": 150.18, "level": 0, "cap": "REAL WORLD", "note": "she believes she is out: the tell"},
{"t0": 150.18, "t1": 150.85, "level": "GLITCHOUT", "note": "the tag tears apart as the headset comes off"}]}
```
| Field | What |
|---|---|
| `level` | `null` for the real world (no tag); a number for a simulation, including `0` for the "real world" she only believes she is in; or a state: `BOOT`, `SPAWN`, `CLIMB`, `GLITCHOUT` |
| `cap` | the caption under the number |
| `rollAt` | when the number rolls to this level (a drop on the beat) |
| `from` | the level a roll or a race starts from |
| `note` | what the story is doing |
A chapter that needs a different boundary asks for the table to change; it never overrides it in its own code.
## The logs
`DECISIONS.md`:
```text
## · director · board
"s034-s036 could use more creativity, kind of a boring moment."
→ s034–s036 rethought (replay loop, tally marks).
## · director · supersedes "keep the halftone"
"keep it clean so it feels like imax, high res rather than dithered"
```
`ledger.jsonl`, one paid call per line. A real cost is filled in from the provider; the estimate is never quoted as
spend.
```json
{"ts": "…", "provider": "openrouter", "model": "bytedance/seedance-2.5", "kind": "video", "units": 5,
"est_usd": 1.16, "real_usd": 1.155, "request_id": "", "status": "delivered"}
```
`RUNS.md` gets one line per run: the run id, the script path, the arguments file, what it does and when it started.
## A director's notes on a cut
`NOTES-v.md`, when the director's notes on a cut come as one long message (AIRPLANE MODE's did). Every agent reads it
first, and again after any context reset. It has two parts:
1. **Verbatim:** the whole message as a quote, unedited, typos and all. It is the authority.
2. **The parse:** first the standing rules that apply to every frame, then the points chapter by chapter, each with its
time span and shots, and the production's own decisions marked in bold.
```text
R1 Linger: "" → one scene per lyric phrase or more; the accents inside the shot.
1. Opening (0.0–11.0 s, ch01): "" → S01 the boarding, S02 the wink. ****
```
`rounds//.md` holds the lead's notes for one chapter in one round: its span, the director's words for that
chapter, the numbered changes with their times and assets, and the hand-offs with the neighbouring chapters. The lead
appends, never edits; agents re-read the file before they finish.
```text
UPDATE (lead): registered (start 13.2 s, 4.04 s): . Re-time it so the bolt lands on
the deadbolt cue (15.49); use the box the next chapter's code returns, not a copied rectangle.
```
## A review
What a reviewer returns per chapter. The verdict is computed from it in code.
```json
{"owner": 7.5, "craft": 7, "sync": 8, "story": 7, "pass": true,
"fixes": [{"shot": "s045", "time": 101.2, "severity": "blocking",
"problem": "the coverline sits over her mouth for 6 frames",
"fix": "move the column right (x 196→250) so it runs 40–80 px under her hair"}],
"checks": {"lyric_gate": "132/132 tokens on screen at their time (41.2s)", "sheets": 28,
"ms_per_frame": 820, "flash_max_per_s": 2}}
```
`owner` is the director's taste, and it counts double.
## A check against notes
What an independent checker returns per chapter after a round of notes. It never edits files.
```json
{"pass": false,
"notes": [{"note": "S07: one scene, linger: the guard feeds her, then hands her the card", "done": true,
"evidence": "sheet 2, 15.99–18.66: one shot; the card is up at 17.4"}],
"broken": [{"shot": "S06", "time": 15.49, "problem": "the bolt lands 4 frames after the deadbolt sound",
"fix": "re-time the take to 1.2× from 14.1"}],
"pending": ["S01: the boarding take has not landed; laid out with a stand-in at its planned timing"],
"lyrics": "31/31 tokens on screen at their time"}
```
`pending` is not a failure. `pass` is true when every note is done or pending on an asset, and nothing is broken. A
failed check goes back to the animator, open notes and broken items first, for one refix.
---
From the Claudia wiki: https://claudiagallery.vip/wiki/formats/
---
# What went wrong
Every failure that cost the films time or money, as what you see, why it happened, and the rule that came out of it.
New here? Read [The recipe](https://claudiagallery.vip/wiki/recipe/) first: the current method on one page.
Each row cost at least one round-trip, and some cost hours or hundreds of credits. Each one left a rule behind. "Another
project" is a production made with the same tools between the films.
## Brief and taste
| What you see | Why | The rule now |
|---|---|---|
| v1 rejected for its look: "way too.. gray", "color grading looked insane", "dont love the dot grid" | the look was assumed, not asked | Ask up front: native colour, one grade, or a print texture. Record the answer. |
| Plates "feel like a stock photo" | generic subjects: crowds, elderly or patient portraits, hands holding props | Change the idea entirely, not the seed, or make the shot pure motion design. |
| A sequel's song feels like a copy | it reused the last song's device, a countdown | A sequel never reuses the previous film's device. |
| Viewers read the lead as a teenager | age words and schoolgirl-coded wardrobe | Write an explicit adult age, use editorial wardrobe, and never write "young". |
| A taboo word turns up in a label or a file name | the taboos lived only in the brief | Put taboos in the project file and in every agent's prompt. The audit checks the board; grep code and file names before release. |
| "off pace and off vibe with the song" | the busiest picture sat on the quietest music | Measure picture motion against stem energy per section. |
| Agent-written style rules pulled work away from what the director liked | rules without a verdict behind them | Keep exemplars and verbatim verdicts. No style rule without the director's yes. |
| "youre oftne over cutting" | a fast song (180 BPM) cut on its lines and half bars, with a cutaway for every accent | Linger: one scene per lyric phrase or more, built inside the shot. The accents go into camera moves and actions. |
| The story's layers didn't read: "need to understand its the entertainment terminal better bc i didnt carch it at first" | the simulation, the seat-back screen and the real cabin shared one look | One look per level, a tag that names it, the real world untreated; show the physical screen around an interface. |
| "theres no compelling intro or story here, the opening should already introduce part of the story … we always need a hook" | the film opened on a mood | Open on a hook that tells part of the story, at the tier of the film's best passage. |
| "it should be like photography, i feel weird looking at a clipped off the edge ost and a claudia in the middle bottom of the screen" | elements placed where they fit, not composed | Compose every shot like a photograph: thirds, her big, every card inside the frame with margins. |
| "she shouldnt really be looking down on her phone, she should be immersed in a world" | a phone in her hand, its screen tracked onto the back of a phone Wan had drawn | She touches the world. A screen that matters is large in the world, and she works it with her hands. |
| "the waking uip and plugging in phone at the end is boring" | the ending went back to where the film began | End somewhere new: "she should enter a world with no phones, like escaping the truman show". |
| Two doors in a row at the end: "thats redundant" | two chapters each ended on a door | One door and one look back in the film. |
| "there is this temporal signature that screams claude now that a ton of people are making the videos. people say pacing or prosody." | shots timed in seconds and by a model's length limit, not in the music's bars; cuts that hid a take's faults | Write the film on the bar map: cuts only on bar lines, at a section, a phrase or a sung phrase's downbeat; moves in bars; a failed take made again, never cut around ([tool 44](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline)). |
| "okay so i totally overindexed on the orbit thing lol … its kinda dizzying now" | a note about robotic-arm camera moves taken into every shot of the previz | A reason for every move (what it reveals), timed to the bars; a director's favourite device goes where the song asks for it, not into every shot. |
| "i think its like you litearlly use chromium and do frontend web dev to make these animations, and it ends up feeling like a website" | the films' motion design drawn as web graphics on top of the picture | Tell the story through shaders, mathematics made visible, and draw only the lyrics on top, small and classical. |
| "Idk if feels like you're adding ... items to increase complexity rather than compressing intricacy?" | shader worlds built by adding objects | One iterated rule per world, its light and colour from the loop ([tool 49](https://claudiagallery.vip/wiki/toolkit/#49-a-world-as-one-rule)). |
| The pill inside the pill: "i kidna like but i feel like its too ordered or not understadnable enough" | the nested pills sat in a neat grid | Scatter them, at random sizes and tilts, so it reads as worlds inside worlds. |
| "it felt lazy like we were adding tags tarther than makin gan official aperture video" | captions and readouts put on the cut as tags, with no idea of the scene under them | One designed layer the whole film is seen through, composed shot by shot: her HUD ([tool 55](https://claudiagallery.vip/wiki/toolkit/#55-a-hud-from-one-visor-and-one-projector)). |
| "i dont like the idea of a static shot of just a boom, or a wall with lettering" | singing shots of a hanging head, and walls of type | Every shot is a story event with an interaction: she reaches for a memory on a screen, a panel arm hands her something, the camera moves with the scene. |
| "the composition of the animations, past just the heads-up display thing, don't feel like they understand the composition of the scene. They just feel completely independent." | the overlay placed by rule, the same way in every shot | Look at each shot's frames first (its subject, its empty space, where the eye goes), then place each element with a reason ([tool 56](https://claudiagallery.vip/wiki/toolkit/#56-the-hud-composed-shot-by-shot-and-the-overlap-check)). |
| "every time there's a shot that's literally a quarter of a second, we need to get rid of that." | shots added between others, and a cut that ran on one half-bar clock whatever each shot held | No piece under 1 s outside a run of simple images for a musical reason; length by content; the cut rate by section ([tool 58](https://claudiagallery.vip/wiki/toolkit/#58-the-pacing-analysis)). |
| The whole first film set aside: "i want to go back on the song and make a new one" | a first song, and a story too thin for the film: "i think a lot of it we didnt have a good story" | A new song is a new film. Write the story on the song first (whose memory each section is, what each line shows), then the shots. The tools carry over; the takes don't. |
| "[…] the look is off." | Midjourney's first round, written from the research's words, came back as a photographic fashion editorial | The look comes from references the director picks: three of their frames as style references in every prompt, and a palette measured from them ([tool 67](https://claudiagallery.vip/wiki/toolkit/#67-the-palette-measured-from-the-references)). |
| "it feels like.. a little boring compared to bliss", and "some of it feels a little action deprived compared to bliss" | v1's first half, which an art director's second-by-second measure of novelty against BLISS also found the flattest | New things inside the surfaces the film already has, one per existing event, on its beat, where the measure is lowest; the bars and every timed beat kept. |
| "1:20 on is perfect but its very boring up to 1:20 - so i dont think that people will watch til then." | an opening below the tier of the film's best passage, the 3D drop | Bring the best passage's world forward: AERO's first 1:20 went into a 3D world with the desktop as its screen ([Storyboard](https://claudiagallery.vip/wiki/storyboard/#in-and-out-of-the-screen)). |
| The all-3D previs: "way too disorienting overal" | a monolith from the sea, shards in the sky, the world stopping with every prompt, a giant prompt, two big screens, a one-frame flash, a punch-through montage | One continuous camera with the screen the hero, about 95 % of the frame; nothing in the world reacting; nothing full-frame for a frame. |
| "Why'd you remove a lot of the 3d?", then "i hate the random camera movements" | the lighter finals cut the 3D away; the A film's push-ins and pull-outs showed nothing new | Structure, not more or less motion: where the film is in 3D, where it is flat, and a reason for every move. |
| "they dont fit the lyrics of the future that was promised really" | wallpaper proposals made softer and airier, not for the line | Make the picture for the line it plays under: "The future used to look like water, sky and green", pictured as that promised future. |
## The song
| What you see | Why | The rule now |
|---|---|---|
| The song came out twice as long as planned (2:19 → 5:06) | Suno gives each spoken line its own two-bar slot | Plan two bars per spoken line, and write 10–13 syllables with the stresses on the beat. |
| Verses went choppy after optimising a score | a metric that rewards short clauses on the pulse | Use flow for verses and groove for chants, calibrated to the director's ear. |
| A name mispronounced | Suno's reading | Respell it in the Suno field only; the screen keeps the real spelling. |
| Numbers and acronyms sung wrong | written as digits | Spell them as sung ("A G I") and show "AGI" on screen. |
| The lyric rushed or truncated, silently | the lyrics field ran over about 3,000 characters | Aim for 1,800–2,400 characters for about 2:20. |
| The chosen take can't be fetched | the download is a native browser download, and the CDN wants signed URLs | The person downloads it; that download is also the licensing step. |
| "A voice already exists for this song" | Suno allows one Voice per source song | Make one combined Voice, spoken and sung. |
| A patched export drifted the song by up to 0.7 s | an editor export with BPM lock on | Export with the lock off, diff the spectrograms, and time-map the edit instead of re-cutting. |
| A new take changed the song's shape | takes differ in structure, not only in timing | Re-time the board from word anchors, and restructure by hand where the shape changed. |
| Forced alignment failed on a phonk take: 102 of 164 words under 0.2 confidence, one word stretched over 8.5 s | the take stutters and repeats far beyond its sheet | Take the times from a transcription with word timestamps and the words from the sheet; every stutter is a token. |
| The scorer found no rhyme in a couplet that rhymes ("PSI" / "SI") | "SI" read as one syllable | Spell acronyms as letters for the scorer too ("S I"). |
| The first timing work read the take as about 89 BPM | 89 is two-thirds of the take's real 133.39 | Check a tempo against the genre and the kick spacing before any timing work. |
| "the music model is skipping words or syllablrs to fit it to beat, which is usually a bad scansion sign" | hook lines of 7–10 syllables in slots Suno sings as 1–1.5 bars: "blue" lasted 0.02 s, "baby" 0.06 s | Hook lines of 6–7 syllables in one shape, "one syllable per note" in the tag; measure word lengths on every take ([tool 2](https://claudiagallery.vip/wiki/toolkit/#2-the-lyric-scorer-scansion-and-assonance)). |
| The first vocal stems were of the wrong take: "thats teh wrong song" | the locked take was the director's own, and the production's notes had matched it to one of its own takes by length | Split the locked take itself, and check the stem against the master by cross-correlation (zero lag), not by title or length. |
| ESCAPE VELOCITY's sound no longer stood out: "many peopl ehave made forks of escape velocity, so that sound is exhausted" | others had made songs in it | A new film gets a new sound: one test lyric on three different styles, picked by ear. |
| Round after round of takes rejected: "these suck why like indie pop or folk pop, it needs to sound like glados ideally or sound like it comes from portal", then "it feels like im watching a cheap mortgage slide show youtube video" | every style had "pop" in it, and pleasant adjectives: Suno's stock-music default | Name the strange, specific sound the film needs, and exclude the stock words. The director named theirs: "cant u just say Delay Lama as the main instrument?" |
| The treated voice: "it like hjurts my ears" | the treatment ran on the consonants too, and lifted the S's (centred at 9.9 kHz, against the singer's own 8.7) | Treat only the voiced frames; the consonants keep their own envelope, softened above 4.5 kHz. |
| An octave down to stop the sting: "oh now it sounds too low. maybe not octave. like before it was extremely high pitched now its like a guy." | an octave shift changes who is singing | Keep her own pitch; lower the formant and tame the top instead. |
| The treatment's added cracks could land anywhere: "make sure the 'glitches' … land on a reasonable key/register/beat still" | a fallback snapped her notes to the detected scale, which would have pulled her C♯s to C or D | Track her notes chromatically, and put each crack on a safe note of its moment, inside her range, on the 16th grid: 65 of 65 in key and in range ([Song](https://claudiagallery.vip/wiki/song/#a-treated-vocal)). |
| "P-doom" was sung "p'Doom" | a 70 ms schwa on the 16th before "doom", where "P-cake" had a clear "pee" a beat before its word | Fix it at the source: the "pee" of "P-cake" copied to the beat before "doom", at its own pitch, so the treated voice and the voice under it both carry it. |
| The levelled song lifted the quiet passages 8–9 dB over her | evening out her notes had flattened her dynamics, and the master raised and limited the instrumental | Level only the voice, to Suno's own lead (BS.1770 integrated, then a slow gain on 2 s windows); the instrumental stays as delivered ("solely master the voice and not the instrument stems"). |
| ffmpeg's loudnorm went dynamic without a word | +4.8 dB of gain would have peaked at −0.3 dBTP | Limit the peaks first (at 192 kHz), then the two-pass linear loudnorm. |
| "none of these feel great, whats the iconic music from that time? […]" | round 1's sounds were today's lanes for the look, not the era's | Ask what the era itself played, then let the director name the genre. |
| "i want the structre of the song to be more described so it doesnt drift too much" | section tags that said how a section sounds, not how long it is | Every tag gives its section's length in bars and what plays in it. |
| The song locked for the first cut was "the worst possible song of all the gens you made - it had mode collapse audibly in it in the second half" | the production picked it from word checks and an audio model's written description of each take, without the director hearing it | The director picks the song by ear before any lip-sync spend, even when the pick is handed over; a described listen is a note, not a verdict ([Song](https://claudiagallery.vip/wiki/song/#picked-by-ear-not-by-description)). |
| Every take of the first round stumbled on the same seven lines | lines Suno couldn't sing in their slots | Ease only those lines, and sing the draft again. |
## Listening
| What you see | Why | The rule now |
|---|---|---|
| Cuts drift off the beat later in the song | a fixed BPM on an accelerating take (131.5 → 133.9) | A variable beat map, never one BPM. |
| One chapter's lyric check failed on a word 2.6 s early | a delay throw aligned as the word | Ear-check gaps over 1 s and very low confidence before the board. |
| Bars start on beat 2 | claps on 2 and 4 won the downbeat vote | Add a chord-change vote, check by events, and keep an override. |
| The loudest moment gets no camera impact | the detector missed a hit that grew out of a loud chorus | Keep a manual add and drop list of hits. |
| Snare and hat lists full of kick clicks | a broadband kick fires in every band | Drop onsets within 45 ms of a kick. |
| Sung lines and shouts missing from the alignment | the aligner is weak on singing; the transcriber missed shouts | Merge the two aligners by confidence. Copy a shout from the earlier chorus plus the bar offset, and mark each word's source. |
| One reading of the line starts was off by up to 0.5 s on 17 lines, and another started a line on its second word | two readings of the vocal disagreed, and each was wrong somewhere | Check each line's start against the vocal stem's onsets before anything is timed to it. |
| "we have to be aware bpm slightly shifts w suno over the whole song" | the take drifts from 95.7 to 99.4 BPM, so a fixed grid puts the later hits off the beat | A 16th grid tracked from the drums, every beat locked to its hit: 91–99% of the strong hits within 20 ms ([tool 52](https://claudiagallery.vip/wiki/toolkit/#52-the-song-map-and-the-glitch-map)). |
| Hits listed for a held vowel that only flutters | a re-trigger finder returns a peak on every slot | Measure the grid contrast (the peaks on the slots over the peaks between them); near 1, take the grid itself as the hits. |
| The transcript heard "I'll wait for you to come back" over chorus 2's octave flips | transcription is weak on a treated voice that flips registers | Snap word starts to the lead's onsets, fix them by hand from the pitch track, and cut on bars or glitch hits, never on word times. |
| One word parked 3.5 s early, among wordless chopped vowels | the aligner put the "It" of "It can stay" into the chops before it | Look at the words on the lead stem's spectrogram wherever the song chops vowels, and check them against the transcript. |
| faster-whisper's word times ran 0.2–1.2 s early | transcription times on a sung stem | Forced alignment on the lead stem for the times; the transcript for the words only. |
| A transcription put the first word after a pause up to 2 s early ("Ask again on Saturday" 15.02 s against 16.93 s); windows gated on it would have cut off first words | transcription times stretch the first word after a pause backwards | Line times from the stem's level; the transcription says only which words ([tool 78](https://claudiagallery.vip/wiki/toolkit/#78-the-line-table-from-the-stems)). |
| Two pairs of lines sung legato were parted in the wrong place | the boundary was the deepest dip in a wide search | Part them at the dip nearest the transcription's own boundary. |
| The group's questions were missing from the lead stem | in that take Suno put the group in the backing stem | Separate the backing stem (demucs, its vocals) for the group alone. |
## Her and the plates
| What you see | Why | The rule now |
|---|---|---|
| Her ethnicity drifts; 12 plates flagged | the prompt didn't pin it | The canon text pins ethnicity, age and skin in every prompt. Test four plates before any batch. |
| Identity edits drift from the look | an early image model "replacing only the woman" | The first films kept her identity in prompt text only. Since AERO an image model puts her in from her design sheet (face-on since FABLE), the frame first and the sheet second, and holds her ([Consistency](https://claudiagallery.vip/wiki/consistency/#the-short-version)). |
| A close-up-sized head on a distant figure | face image references in Midjourney: a whole sheet imports its layout, a single face over-conditions | No face images as Midjourney image prompts for her. In a video take, her face-on sheet as @Image2 whenever her face may show. |
| "Version 8.2 doesn't support --oref" | that parameter doesn't exist on that version | Don't use it. |
| Full-length prompts come back as close-ups | the long face canon pulls the frame in | Camera first, plus the short far form of the identity. |
| A feature asked for as an adjective goes missing | it wasn't the subject of the sentence | Make the feature the subject ("the hide studded with dozens of large glossy eyes"). |
| The crew are copies of the male lead | one description for a whole cast | Describe each extra (age, build, hair) and add an explicit negative. |
| A sheer gown read as plastic wrap and nude | "sheer … bare shoulders, bare feet" | A structured, high-neck, long-sleeved look, "composed and deadpan". |
| A stranger in an over-the-shoulder shot | the shoulder had no canon | Over-the-shoulder shots carry the canon too. |
| The robotaxi came back as an old sedan | the object's form wasn't named | Name it: "a small white compact electric SUV robotaxi with a tall spinning lidar dome". |
| The same image on about six shots | stand-in aliases left in the board | One image per shot; use the unpicked candidates as a variety pool. |
| "Two heads next to each other" ("bodyslop") | a ghost and a person composed in one prompt | Describe the ghost as made of the scene's light, or compose two assets. Make groups with an edit model. |
| A fix painted by hand in code | fixing the frame instead of the source | Bake it into the prompt or use an image-edit model. Never paint pixels. |
| The edit model invented a six-pointed star | edit models invent details | Re-check the canon after every edit, and name the exact form ("exactly eight points"). |
| Edited plates rendered with their old depth | prep wasn't re-run | Re-run depth and boxes after every plate edit. |
| Images filed under the wrong prompt | filed by their order in the batch, and Midjourney rewrites long prompts | Check each image against its own prompt before filing it. |
| Lost picks | parallel runs racing on one state file | Use a locked read-modify-write. |
| A game avatar that "feels like roblox" | a low-poly 3D model used as the final character | 3D renders are motion references only; the video model draws her from the avatar's images. |
| The stylised avatar came out chibi | a stylised heroine drifts to a big head on a small body (about a quarter of her height) | Write adult proportions ("her head about one seventh of her height") and screen every still for them. |
| An off-model take: a cropped jacket and a bare midriff instead of the one-piece suit | the still had the wrong costume, and the take followed it | Screen stills for the costume as well as the face. Re-still, re-take, then re-fit everything measured on the old take. |
| The real car came back as a generic 1980s wedge | an image model can't draw a specific real car | No generated picture of the real car: the car is always the real footage, and she meets it in video edits of the real clips. |
| Asked to hang upside down, she came back only tilted (four images of four) | Midjourney tilted her instead of inverting her | Ask for her weightless, the hair lifted straight up, and turn the plate 180°: the lifted hair reads as hanging. |
| Patches, badges, numbers and a keyhole cutout on a suit "with no lettering" | "no lettering on her clothes" didn't hold in every image | Screen every suit; in the video prompts, close the collar to the throat and add "no badges". |
| "p24 s24 pick has 3 arms lol"; three of four stills of one turn with backwards legs | picks screened on thumbnails about 230 px tall | Check every pick at 800 px or more for limbs before spending on it. |
| A batch's prompts ran past 600 characters, the length DON'T LET ME SCROLL kept its Midjourney prompts to | the new outfits' descriptions were long | Write to the length; shorten before the batch goes. |
| "[…] i think the top isnt great […]" | the previz performer wore a skin-tight green suit with no build given, and the video model copied the body | A slim performer in a loose, thick costume, her build in words in every prompt, and a new outfit. |
| "hm dont love the shirt. […]" | the outfit went straight into a take, unseen, and the model drew her off-model | Show the outfit's stills before anything is generated with it ("can you show me the images of the outfit before generating this time"), and write her build in words. |
| "you should give a full face sheet like this next time. otherwise it makes her neck all hitched lol" | one large crop of her face as the only face reference | Her whole face sheet (front, three-quarter, profiles, back, expressions) for every video take. |
| "wait align w me on teh references you use lol - its clearly not going to work if you use reference 6" | a place plate with her in it in an old outfit, and old face crops, picked as references | Agree every reference with the director before a run, on one board; places with nobody in them. |
| The dress the director picked had two star clips, and the streak on the wrong side | Midjourney's image | Mirror it and take out the second clip before it becomes the reference. |
| Midjourney stood her head on a neck, shoulders or a pedestal, and mirrored her hair | it draws a head as a portrait, and doesn't hold which side a streak is on | Two sentences in the identity fix: she hangs from her boom, and which side the streak and the clip are on ([tool 59](https://claudiagallery.vip/wiki/toolkit/#59-storyboard-frames-through-midjourney-and-the-identity-fix)). |
| The identity fix put her streak on her left, the clip's side | the sheet's front view swept the streak there, and the edit followed the sheet over the words | A sheet whose views agree with the canon on both sides, and the sides in words. |
| "i thin she loses her identity in the reimagine ones, but they do look better overall" | a Midjourney reimagine redraws the face: warmer, less porcelain, the streak low | Keep the reimagine for the look, and put her back with an image-model edit to her sheet. |
| A portal gun in about a third of the images from "screenshot" prompts | "Portal 2 … screenshot" brings in the game's first-person gun | `--no … portal gun, gun, weapon, hands` for places. |
| Midjourney's Aperture came out too detailed and too dramatic beside frames of Portal 2 | more cables and clutter, and stronger light (god rays, bloom) than the game's baked light | Frames of the game as style references, and one grade over everything ([tool 57](https://claudiagallery.vip/wiki/toolkit/#57-the-per-shot-grade)). |
| "It doesn't really look like her." (the memory film's head, on draft 8) | every modelled head (a scan, a parametric bob, a bob of hair curves) read as a mannequin | Her own face: a photoreal plate of her, matted and seated on the modelled stand ([tool 60](https://claudiagallery.vip/wiki/toolkit/#60-a-memory-film-with-her-own-face)). |
| Her eyes blazed in the dark | "eyes ignite" in the prompt, on a dark set | "Her warm amber glass eyes keep only a soft inner glow; they never blaze." |
| Her face drifted off her sheet in every Midjourney frame of her | style references from the reference videos bring their own kind of face | Midjourney makes the world; an image-model edit puts her in from her sheet ([Midjourney](https://claudiagallery.vip/wiki/midjourney/#the-world-in-midjourney-her-from-her-sheet)). |
| Midjourney's 2008 came back glossy and new | it draws the era as today | The era's older look from an image model, with frames of the reference videos as its references. |
| A faint, made-up stock-photo caption in a corner of four picks, kept by their loops | Midjourney, whatever "no text" says | Check every pick's edges at full size, and frame the corner out. |
| A drop of water on a fingertip came back as a crystal ball held between two fingers | the object wasn't described exactly | Name it: "the pad of a woman's index fingertip with one small round water droplet resting on it, the droplet the size of a lentil". |
| Three loops played under the wrong names | files filed by their names | Check files by their content: compare them byte for byte. |
| Her far figure came back as a 3D cartoon doll in an all-white suit | the take was asked to bring her into an empty plate | Put her into the first frame from her sheet, however small, and let the take move her. |
| A still with an owl-like neck: "unrealistic neck angle turned around like an owl with the frustrating tilt" | "turned three-quarters toward the camera" on a full-length figure, and a sheet whose main view tilts | A sheet whose main view is face-on, and for that still the pose written square ("her feet, hips, chest, shoulders and face all point the same way") ([Consistency](https://claudiagallery.vip/wiki/consistency/#her-sheet-redrawn-face-on)). Not in the takes: the same posture written into every take froze her (the genie row under Clips and lip sync). |
| A thin line with rounded corners a few pixels in from the edge of several stills, later built into a wooden picture frame by a take | "35mm film still" makes a picture of a film frame | "the picture fills the whole frame edge to edge, with no border, no frame line and no rounded corners", and trim any still that has the line before shooting. |
| Two close-ups of the chorus bedroom were another bed: no paper animals, another bonnet, another shadow | each was built from the plate on its own | Make a close-up as an edit of its wide frame. |
| Dark wings at both edges of a set read as pillarbox bars | the set's own flats | Crop them in the cut, per take. |
## Clips and lip sync
| What you see | Why | The rule now |
|---|---|---|
| Lifeless takes, about 830 credits burned | agent-written prompts: poses, not drama, and nobody read them | One writer writes every prompt in one context, with the story beat, blocking, and timestamped action and reaction. The director reads the book before any spend. |
| Sung takes re-sing the line instead of copying it | long timestamped screenplays on sung windows | With a clean vocal window (the first films), the short sung form for sung windows, and screenplays for motion. Takes from the song's own vocal stem copy with a full prompt of about 200–400 words: the words quoted in whole-second segments, every action silent ([Seedance](https://claudiagallery.vip/wiki/seedance/#lip-sync-from-the-songs-own-vocal)). |
| One window barely syncs whatever the prompt | a microphone at her mouth on the plate | Sung plates show the whole mouth, with nothing in front of it. |
| QA kept weak takes | takes were checked against their own weak prompts | Check takes against the story and the board. |
| Two clips start on the same frame ("jarring") | the same plate as the start image | Use a sibling candidate, or the previous clip's last frame. |
| Lip-sync models "still not exact" | the failure was the input, not the model | No lip-sync models. The first films measured, cut and covered, or cut away; now each take sings from the song's own vocal stem and is placed by measurement ([tool 46](https://claudiagallery.vip/wiki/toolkit/#46-lip-sync-from-the-songs-own-vocal)). |
| A take diverges inside a word | the prompt quoted fewer words than the window held | Quote exactly the window's words, and start windows in word gaps. |
| Lips a fraction early or late in the edit | each take has its own lag, up to ±1 s | Map clip time and song time through each take's measured lag. |
| A window cut a word ("…could. Your P") | the window ended mid-word | Start and end windows in word gaps. |
| Motion clips "diverge" in the sync check | the mix isn't copied (NCC about 0.08) | Expected: play motion clips at lag 0. |
| Grey proxy shapes in a take | a 3D previz video reference leaked into the output | "take only the camera move from it", or drop the reference. |
| Drift on long sung clips | lip sync degrades over length, with a clean vocal window | With that window (the first films), sung windows of 4–8 s. The longest takes from the song's own vocal stem ran 11 to 17 s in each film since; plan them at 4–15 whole seconds ([tool 46](https://claudiagallery.vip/wiki/toolkit/#46-lip-sync-from-the-songs-own-vocal)). |
| Head counts change, actions reverse, garbled text | under-specified prompts | "exactly four …", name the end state, one dominant camera move, "No text, letters or logos". |
| "Upscaled" footage looks waxy and shimmers | learned upscalers | Keep Lanczos, and generate hero shots at 1080p instead. |
| 32 of 59 clips came back near grey | references pre-graded to monochrome | Send references in colour; any look goes on in post. |
| A sung clip ignores its audio | a first frame sent with the audio reference | On OpenRouter, use reference mode for sung clips and covers, never `frame_images`. |
| Guards came back in white; a guard showed his face | they were barely in the first frame, and the model invents what it can't see | Keep them out of frame and spell it out: "Everything the guards wear is black; the only white thing is the VR headset." |
| A colour name printed across her chest (INTERNATIONAL) | "international orange" in the prompt | "vivid safety-orange satin, plain, no lettering" |
| The team walked down the boarding stairs instead of up, and read as leaving | in the still they faced the camera | Pick a still whose bodies already face the way they must move. |
| Replacement takes kept the blocky stand-in figure | the motion reference's figure read as the subject | Call it a placeholder and list what not to copy; add a still of the avatar in that shot's world. |
| A small white cube for a few frames where the clay render had a spinning top | the motion reference still carried a prop the story had retired | Name each retired prop in the prompt ("There is NO spinning top … her hands are empty"), and check the frames where it was. |
| A take cut to a new angle mid-action | the model's own editing | Ask for "One continuous take … No cuts, no scene change, no second camera angle". If it still cuts, cut there in the edit and drop the morph frames. |
| The real car turned into a generic sedan, a Porsche-like car on a race track, a dark grey car: six of the first twelve moves on real frames lost the car or the set within 1.3–2.5 s | Wan keeps the real car only on the sides its first frame shows | Keep Wan's moves on the visible side. Orbits that reveal a new side go to Seedance, with real frames of that side as references. |
| Takes cut to a grey studio wall, a white spaceship interior, a blue sci-fi screen, a road with lamp posts, mostly in their second half | Wan's own editing | Review every take on a sheet with a cut finder, log its usable window, and draw only inside it. |
| Seedance cut to its reference images as if they were shots | reference frames read as shots; saying they weren't still left 6 of 11 takes cutting | One role per reference and "never shown as shots"; script the move to land by 3 s; crop references tight on black; end every window before the cut. |
| Parted lips and teeth in a film where she never sings | plates with parted lips, and the word "teeth" in a prompt | Closed-lip first frames, never the word "teeth", "Her mouth stays shut from the first frame to the last", and a check of every face shot's frames. |
| A man in the window of a car nobody was in | nothing in the prompt said the car was empty | "The car is empty and flown by remote: nobody is inside it and its glass stays dark." |
| Cold-gas thrusters came back as fire | "white thruster plumes" wasn't enough | "Its thrusters blow white cold-gas vapour, never flame or fire." |
| A chrome object flew across her face | the prompt named a light sweeping across the glass | "a light only … the light stays outside the glass", and "Nothing enters the cockpit". |
| Video edits cut to her face on its grey studio wall | a face close-up given as a second reference | One cleaned full-length plate: "Image 1 is only a reference for her look; it is never shown as a shot." |
| A re-shoot rolled the car upright, the very glitch it was made to fix | the state written in the prompt held in one take, not both | Write the state as a requirement too. DOWNFORCE made two takes of every re-shoot, and each of the four jobs had one that held; since AERO, one take per job, re-rolled only when it fails ([Spend](#spend)). |
| Lifted blacks in generated takes (about 14 on 0–255, against the footage's 1) | Wan's output | Measure each take's black level on its sheet and restore it at extraction. |
| Slow motion stepped at 10–12 pictures a second | a "60 fps" feed that carries about 24 distinct pictures a second | Count a source's distinct pictures before planning slow motion; blend neighbours, and motion-compensate where the car and the camera both move. |
| Wan's dance held the first pose, reset and shuffled | Wan doesn't hear the song | Dances that must land on the beat go to Seedance, with the shot's instrumental as @Audio1 "for rhythm and timing only". |
| "never slow these down. you lowered the framerate for no reason, just looks bad" | beat warps ran the takes at 0.70–0.89× in places, one chapter's own map at 0.53× | Warps floored at 1× ([tool 42](https://claudiagallery.vip/wiki/toolkit/#42-beat-warps-floored-at-1)). A take too short for its shot loops on a matching pose, plays as a sampler, or is taken again. |
| Hands and feet vanished at "blurred bounding boxes" | limbs left the 9:16 take's frame and were faded at its edge | Shoot her wider than the shot needs; complete what left the frame in a Wan pass ([tool 40](https://claudiagallery.vip/wiki/toolkit/#40-wan-30-relight-and-repair)). |
| A floating glass panel where an invisible screen should be; focus brackets round her fingertip; a yo-yo landing on her toes; a stool seat under her | Wan fills a gesture with things it imagines | Say what isn't there: "nothing is visible in the air: no screen, no panel, no glass, no hologram, no light, no interface, her hands move through empty space". Check every take's frames. |
| The camera moved although the prompt locked it off: she shrank about 20 % in one take | Wan's own camera | Place her by her feet frame by frame, or take it again. |
| An outfit swap went in with a frozen half-second | it was aligned on her silhouette, which the new skirt changed | Map the edit back by the ratio of the lengths, and check the fit on her hair ([tool 37](https://claudiagallery.vip/wiki/toolkit/#37-outfit-swap-alignment)). |
| Re-lights softened the interface's small type, and one smeared a notification card | a video edit re-renders every pixel | Keep the original pixels in rectangles, tracked per frame when the card moves ([tool 38](https://claudiagallery.vip/wiki/toolkit/#38-the-re-light-splice)). |
| Asked to move two posts inward, Seedance kept them where they were | Seedance re-textures and re-lights; it doesn't recompose | Recompose in code first, and re-light last. |
| A repaired face drifted in 1:11 and read "off in the 0:36" | Wan redraws her face, and in one take her chin was tucked into a sweater's collar | A face reference cropped to her face and hair, "only a reference for her face"; for the full face, write it slim and lift her chin out of the collar. |
| "seedance didnt rly change anything" on a pass meant to fix a previz take's melted body | the strict editing template made the take the "sole editing master", frame for frame | For fixing a previz, a loose prompt: "just say maintain pacing and lip sync, fix any issues, vague body shapes, etc". The strict template is for re-lights. |
| A whole-film reshoot in Wan 3.0 Prime over the previz: ":/ looks like shit. i feel like h3 is dragging it down." | every shot redrawn over a full-resolution previz whose faults it kept | Test a master on one shot, several ways, side by side; give the previz as a rough guide at 320 × 180, only the lip sync exact. |
| The previz's lip sync failed in places: "0:27 for example - shes talking when theres no words yet" | each take's prompt listed whole lines where its audio held part of one | The exact words of each take's window, cut lines marked "…", every silence described; a word sung across a cut belongs to both takes ([tool 45](https://claudiagallery.vip/wiki/toolkit/#45-a-previz-performer)). |
| A toothy smile in the silence before she sang | "smile" in the take's mood | No smile in a sung take's mood: joy in her eyes, lips closed until the voice begins. |
| "ok it still cut so we'll need to just do h3 til we get them without cuts" | told to merge a previz take's cuts into one camera move, Wan kept the cuts | Fix cuts in the previz: keep only cut-free takes. |
| The previz put her in the dancer's background for a few seconds | the dance reference was the whole video | Matte the dancer onto flat grey. |
| Four dance shots lost hands and legs: "it must be that the roto or whtaever on dance refs lose bodyparts and the video follows it exactly" | the previz guides had lost them, and the master copied what its guide showed | No guide for those shots: the place as an image, the dance in words, beat by beat, with the tempo ([tool 47](https://claudiagallery.vip/wiki/toolkit/#47-takes-with-no-guide-clip)). |
| A red cord across a tram shot whose place had none | the prompt still asked to remove it | Name a removal only where the thing is in the reference. |
| Kling 4.0 froze her mouth into a smile through every word | it takes no audio track | Lip sync only from a model that hears the vocal. |
| Wan 3.0 Prime singing to the song hard-cut at 9.9 s instead of craning up, masked or not | Wan's own editing | Seedance 2.5 with the vocal stem made the same crane as one move ([tool 46](https://claudiagallery.vip/wiki/toolkit/#46-lip-sync-from-the-songs-own-vocal)). |
| A shader guide at 320 × 180 came back soft and blocky: "the MTWTFSS and the next 'pyramid' shared still is all pixellated" | the guide's detail was too low | A supersampled 1080p render as the guide, with one still per object ([tool 50](https://claudiagallery.vip/wiki/toolkit/#50-shader-worlds-as-seedance-guides)). |
| At full resolution the paper shot drifted half a second and swapped objects: the Enigma rotor became a helix, the weekday cells lost their letters | a sharp guide with no still per object | One still per object, made from the crisp frames, each named in the prompt with its time. |
| Big-motion singing takes re-sang the vocal instead of copying it | events in the prompt that make a sound: clunks, cannons, gulping, a voice box glitching | Write every action as silent: "Every action in this video is silent: the only sound is @Audio1." ([tool 62](https://claudiagallery.vip/wiki/toolkit/#62-a-singer-among-screens-of-her-memories)) |
| A "voice-box glitch" written in words drew a broken face: rolled-up eyes, a stretched black mouth | the model drew the glitch on her | Picture-only glitches, her face always intact; the song's glitches go on her head afterwards ([tool 53](https://claudiagallery.vip/wiki/toolkit/#53-the-head-glitch-on-the-songs-hits)). |
| Her head came back on the motion guide's white shell instead of her grown boom | a video of the character's own body carries the body over | Turn the guide into a coarse blockout (greyscale, flattened, heavily blurred), name it as one, and give her body in words. |
| "0:54 - lip sync bad, need to reshoot." | the vocal window began inside a word, the held "s" of the line before; another line's window began on "(pass)ed" and was re-sung in three takes | Start every window on a clean onset, in a word gap ([tool 14](https://claudiagallery.vip/wiki/toolkit/#14-the-clean-vocal-window)). |
| Memory images took over the room late in a take, or became shots of their own | the model turned a reference into the scene | Each memory "only as picture content inside the screens' glass, never as the room"; keep the seconds the shots use, or put the screens' content in words. |
| "1:01 scene - she's not lip synced really, are we feeding in vocal stems?" | a take of her farther back, loose on the line before the stutter | Nail the line before the glitch, every word's onset from the lead stem written into the prompt, and let the mouth glitch cover the held stutter. Make a few takes and let the director pick by eye: "can you show me so i can look visually at the lip syn cand choose", then "3". |
| A take began mouthing at "cake", a second late | the take came back that way | Check the first words of every take by eye before picking it. |
| The chatbot's answer sat off the bottom of the kiosk's screen: "the france chat message is just off the bottom of the computer screen" | the take placed it there | A retake with both bubbles inside the screen; the HUD carries only her intercept of the question. |
| Chell jumping into a portal "looked v off bc we used a video model" | a video model's physics | Record the physics in the game itself: six shots of the final were recorded in Portal 2. |
| "glados in your renders and chell etc basically had no movement. it was just her static." | no motion reference, and little motion in words | The boom's motion from a recording of GLaDOS's moves as a rough guide, or written into every second of the timeline. |
| On draft 8, at 1:06: "She needs to be farther away and […] glitching still." | a take that lunged at the lens with her mouth open on every hit | Reshoot her farther back, and build the glitch on her head hit by hit ([tool 54](https://claudiagallery.vip/wiki/toolkit/#54-glitches-that-build-and-her-mouth)). |
| "to continue" came 0.36 s early after "permission" | Seedance shortens the pauses inside a line | Measure each phrase and place it on its own; ease between offsets in the breath gaps ([tool 69](https://claudiagallery.vip/wiki/toolkit/#69-lip-sync-phrase-by-phrase)). |
| A take out of sync on its first words; another dropped whole words | one window left only 0.26 s before the first word; another opened inside the held note of the line before | Open every window in silence, half a second to a second before the first word. |
| A take lost the copy of its vocal | 1.6 s of silence before her first word | A lead-in of half a second to a second, no longer. |
| A garbled glass dialog in front of her: "Windows nects our permision" | the prompt had her "lit by a glowing permission dialog" | Never ask a video model for interface or text: the coded interface draws it beside her. |
| A dance take jumped back to its first pose at 7.6 s and held it | the model, unasked | Scan every take for hidden cuts and held frames ([tool 70](https://claudiagallery.vip/wiki/toolkit/#70-hidden-cuts-and-held-frames-in-a-take)). |
| A take turned her back through a sung line, and came in closer than its first frame | a turn asked for across the words | Turns go between lines; check the framing against the first frame. |
| "her head is alittle cutoff by your green border", then "this shouldnt be so hard something is off" | the first frame, a still set at 80 % on its own green, had faded her crown into the green, and every take made from it carried that | Fix the first frame, not the key: one flat green from edge to edge, her whole crown inside it ([tool 71](https://claudiagallery.vip/wiki/toolkit/#71-takes-on-green-the-first-frame-the-key-and-the-dances-phase)). |
| Her face drifted off her sheet when she turned to wave | the take was made from a back view, with no picture of her face | Her sheet as a reference whenever her face may show, even small, even from behind ([Seedance](https://claudiagallery.vip/wiki/seedance/#a-face-needs-her-sheet)). |
| A climax close-up looked soft | a 720p take shown at 1.5× | 1080p wherever her face is shown large. |
| A first-person take looked down into the drop on a fingertip, which filled with the magnified fingertip instead of the world | the camera rose over the drop | Keep the lens level with the drop, in words. |
| A one-second cross-dissolve between two framings in a build take; a ghosted double exposure in a crane | a first and a last frame generated separately, with two cameras | Make the last frame as an edit of the first, and write the camera locked or as one continuous move ([tool 77](https://claudiagallery.vip/wiki/toolkit/#77-last-frames-made-on-the-first-and-takes-that-continue)). |
| The next take jumped at the seam | it started on the frame the take before was asked to reach, an edit 1.7 % larger than the book | A take that continues another starts on that take's own last frame. |
| Takes came back 0.6 s early to 0.4 s late, each phrase by its own amount, and some sang a line 10–16 % slower | Seedance moves the silences and sings at its own pace | Align each take as a path ([tool 81](https://claudiagallery.vip/wiki/toolkit/#81-sync-as-a-path-each-take-aligned-to-its-window)). |
| 0.5 to 0.9 s of silence at a full stop where the song dips 0.3 s (three takes of one line), and 0.33 s where it dips 0.1 s (another line, in its take in the film) | a full stop in the prompt's words | Write a line sung through with a comma; check the take's own sound for a gap its window doesn't have, and run the map through it by hand. |
| A take without a soundtrack shut her lips on "better" 0.3 to 0.45 s late, though a fit of mouth against level read it in sync | the line laid on her mouth's whole stretch of singing; a sung line's level is flat | Anchor every b, m and p on the frame where her lips part ([tool 82](https://claudiagallery.vip/wiki/toolkit/#82-takes-without-a-soundtrack-placed-by-her-lips)). |
| "Mirror, mirror, [0.7 s] on the wall" sung straight through, twice | a take without a soundtrack keeps no pause inside a line, even when the prompt describes it | Make a line with a real pause two segments, or bend the take under a steadied camera ([tool 84](https://claudiagallery.vip/wiki/toolkit/#84-a-re-timed-take-under-a-steady-camera)). |
| A push-in juddered and surged where the map bent a take for her lips | a slow map repeats frames, and the camera's speed follows the map | Mix neighbouring frames, and re-frame each frame to where a straight playback's camera would be. |
| A pop as a take starts to move | its first frame is the input still, sharper than the rest | Never show a take's first two frames; move cuts off holds. |
| Rolled eyes showing the whites | "her eyes lifting to just above the lens" | Never ask for a look just above the lens; in that take, her eyes on the lens or on something in front of her. In every take, "eyes on the lens" reads as a stare (below) ([Seedance](https://claudiagallery.vip/wiki/seedance/#directing-her-on-a-locked-camera)). |
| Her head turned right round; her face went off-model looking down | "nods toward the tower behind her"; "looking from it to us" | Point behind her with a hand while she faces the camera. |
| Her head left the top of the frame, and white tights and red shoes appeared | "steps backward to the bed and sits", and a look that named no legwear | On a locked framing, let her move on her mark (turns, gestures, a step toward us) rather than away from the camera, and name every garment that can come into view. |
| After FABLE's release, viewers read her as stiff and frozen, and as unhappy, even menacing | every take carried posture rules ("Her head stays upright on a straight neck: no head tilt", "she faces the camera with her whole body"), and take after take added her eyes on the lens and mood words ("dry", "knowing", "mock innocence"); the video model did each one literally, and together they froze her | Upright and face-on is for her reference sheet. A take gets her movement in words, one loose clause per segment, or a motion reference; emotional words chosen on purpose, warm by default; precise only where it matters, loose elsewhere; a fix stays with the take it fixed ([The recipe](https://claudiagallery.vip/wiki/recipe/#writing-a-prompt-for-a-model-that-does-exactly-what-it-reads)). |
| Five white flickers in the puppets' eyes; ears removed so they could spring up; a rope grown out of the moon | a puppet's own features, or a prop, scripted to change | Script only what a hand-worked puppet can do: "their painted eyes stay as they are". |
| The last stage of the build read as a morph, the disc swelling and flashing, the hall snapping into shelves | the change was written as one shape becoming another | Write the mechanics fold by fold: "like a row of dominoes … a curved paper bookshelf stands up on a crease in its place". |
## Filters and providers
| What you see | Why | The rule now |
|---|---|---|
| Close-up references declined | the provider's face filter | Free, and final for that image on that service. Send the same unaltered image elsewhere, or keep the 2.5D plate. |
| A retry ladder (stronger grade → posterised grade) got takes through, baked banding into them, and was blocked by the harness | altering a reference to pass a safety filter | Never alter an image to pass a filter. A blocked action is a stop, not a puzzle. |
| Jobs fail with no reason, credits back seconds later | a silent input filter misreading innocent wording ("a plug pushed deep into each ear" for foam earplugs) | Reword plainly and keep the beat: plain rewording fixed five of six at once. |
| "output audio may be related to copyright restrictions" | the provider's audio filter | Retry once with a re-cut window, then try the other provider. |
| A filter that passed faces two days ago now declines them | providers change | Keep a second provider configured. |
| The agent's context floods during uploads | presigned-upload replies of about 3.5k characters per file | Upload by URL. |
| A sung opening blocked four times in three stagings, the last from a clearly adult remake | unclear: every blocked attempt carried the same vocal, while other takes of her cuffed beside guards passed | Stop and ask the director. Re-stage it as a genuinely different shot, or drop it. Never route around a block: no silent take with the vocal lip-synced on afterwards. |
| A plane beside a star was blocked | probably read as a plane flying into the sun | Compose it calmly, with the star well clear of the plane. |
| Seedance's output filter blocked a dance in white satin twice, from two different first frames | unclear; the same dance passed on Wan, and a later drop dance in a red coat passed on Seedance | Make it on the other model. Both blocks were refunded. |
| A leaning pose, pointing at the lens, refused by the provider's content filter | probably the leaning staging | Restage it upright, pointing up at the post above her. |
| Three sung shots refused by the provider's content filter, then a reworded retake and a restaged one too | not stated by the provider | Refused, no take; stop and ask the director, re-stage as a genuinely different shot or drop it. |
| A previz close-up of the performer bringing a pill to her lips refused by the provider's content filter as an input, reworded or not | the guide clip | Refused, no take; re-stage it as a different shot or drop it. |
| WHITEPILL: her face sheet refused by the provider's content filter on Replicate, even alone with a harmless prompt | not stated by the provider | WHITEPILL ran only shots with no face image there. |
| A grey motion-capture body given as a dance reference was refused by the provider's content filter | a nude grey body | No undressed figures as references; describe the dance in words. |
## Spend
| What you see | Why | The rule now |
|---|---|---|
| Every submit returns 403 while the account has credit | the API key has its own spend limit | Set the key's limit above the project cap. |
| "Remaining budget" figures are wrong | the local ledger isn't the bill, and it counted free declines | Quote spend only from the provider. |
| Hundreds of credits spent in minutes on unread prompts | a credit-reset deadline drove the spending | Never let a deadline drive spend. |
| A background workflow could still spend after a quality flag | not every spender was stopped | Stop every workflow that can spend: search each for its generate calls. |
| Credits vanished at renewal | subscription credits don't roll over, and top-ups expire | Plan spend inside the period. |
| $148.10 of the $357.51 spent after refunds went to takes the final doesn't show; on four re-light samples: "kinda a waste on the excess ones" | shots re-lit before their outfits and compositions changed, and eight swaps into a look that was then dropped | Fixes, then reshoots, then re-lights. Re-light a shot only once its composition is final. |
| The runner's ledger read $10.40 low | it priced Wan's passes by their generated seconds and left out the input video | Price a video input per second too. |
| "in the future let me give you explicit approval before you do extra gens." | a follow-up generation went out without asking | Ask before any generation the director hasn't approved, cheap or not. |
| A whole-film Wan reshoot, $132.42 at list, dropped as a whole | the method was judged on the whole film, not first on one shot | Try a new method on one shot, several ways, side by side, before the film. |
| Takes sent out in pairs, for a choice | a second take made for every job, used or not | One take per job, judged before the next; a re-roll only when it fails, from better inputs. |
## Colour and look
| What you see | Why | The rule now |
|---|---|---|
| v1 "way too gray … loses the Midjourney magic" | a mono print pass, and chapters desaturating the footage | Colour halftone or native colour; never desaturate the footage. |
| "color grading looked insane"; blue graphics "clash" | a graded look, and a blue accent over blue footage | Native colour; pick the accent against the footage's hue; draw with semantic colour tokens. |
| Flicker in a graded clip | per-frame grading | One LUT per clip, computed from 16 sampled frames. |
| The master looks washed out ("v blown out") | an untagged or BT.601 stream | Convert to BT.709 and tag it; verify with ffprobe. |
| Re-lit shots came back darker and cooler (warm grey paper from a mean level of about 172 to about 111) or brighter (a dark scene from about 34 to about 53) | Wan shifts colour, and never by the same amount twice | One global per-channel match per shot before the splice ([tool 39](https://claudiagallery.vip/wiki/toolkit/#39-the-colour-match)). |
| The cloud field came back sunset gold: "wayy too saturated compoared to teh pastel light colors of the rest" | the guide was a sunset render and its still was made in the film's saturated look | A day variant of the shader and a pastel still made without the look reference ([tool 50](https://claudiagallery.vip/wiki/toolkit/#50-shader-worlds-as-seedance-guides)). |
| Shader scenes remastered from the render alone still looked like computer renders | nothing showed the model the materials | Stills of the key moments repainted in the film's look, given with the render. |
| "I feel like we don't get a super aperture vibe" | the shots recorded in Portal 2 dim and foggy (whites at Y′ 50–62, blacks at 9%); the generated chambers' walls grey and their casts scattered | Measure every shot, normalise it, and lay one look over all of it as a tone curve and targeted chroma moves, baked per shot ([tool 57](https://claudiagallery.vip/wiki/toolkit/#57-the-per-shot-grade)). |
| A cyan-white portal core turned pink | white balance tinted a clipped channel, which holds no colour | Send near-whites that were already clipped to neutral before the look. |
| A +1 stop gain clipped before the shoulder could roll it off | 8-bit level filters | Bake the gain and the look into one float LUT per shot, applied in 16-bit. |
| Each continuation of the far shot of her lost about three points of the palette's score | each link of a chain of takes, made from the last frame of the one before, darkened the picture | One take from the still itself; score every take against the palette ([tool 67](https://claudiagallery.vip/wiki/toolkit/#67-the-palette-measured-from-the-references)). |
| A falling drop of water turned pink, off the palette | the take | The cut leaves before it. |
## Motion design and the renderer
| What you see | Why | The rule now |
|---|---|---|
| A crop cut her head off after a re-take | code tuned to the old take | Re-frame every shot whose take changed. |
| The previous frame's overlay "sticks" in dense frames | a GPU-backed 2D canvas under software GL | Software-backed canvases, and the history check in QA. It took about two hours of bisecting to find. |
| Frames differ depending on what rendered before | a WebGL canvas drawn into a 2D canvas switches it to GPU rasterisation | Read 2.5D pixels back through the CPU. |
| 0.5–1.1 s a frame on 2.5D shots | the 64-step parallax shader | Depth slices or multiplanes. |
| Tags never finish typing on short shots | the default typing speed | Type at 16–40 characters a second, and finish 6 frames before the cut. |
| A lyric word half hidden on its first frame ("You dor't") | occlusion on the entering row | Tuck an edge only, and switch occlusion off for the entering row. |
| Five accent-coloured elements in one frame | the accent used for every highlight | The accent goes on at most two elements. |
| A snap reads early | keyed 70 ms before its kick | Key moves on hit times from the hits file. |
| Two on-screen numbers disagree ("x380" against "135 RUNNING") | two devices driven separately | Drive both from one value. |
| Shake in a drumless intro | a move that answers no sound | Moves answer audible sounds. Stillness where the song breathes. |
| The same helpers written 8–10 times, slightly differently | chapters couldn't share code | One shared device library. Chapters write only what is unique. |
| The default renderer drew only lyrics | the board's graphics were stored as prose | Structured fields in the board: type, devices, treatment, transition. |
| A lyric time patched in chapter code | fixing the time instead of the alignment | Fix the timed words and re-export. |
| A chapter replaced the default renderer and lost lyrics | the chapter must draw every token itself | Run the lyric gate per chapter, and mark custom-drawn tokens. |
| Fonts that can't be shipped | AGPL, GPL, EULA or unlicensed fonts | Ship OFL fonts only, with their licence texts. |
| A depth model that forbids commercial use | Depth Anything V2 Large is CC-BY-NC | Use Depth Anything V2 Small, which is Apache-2.0. |
| A singing mouth held where the take isn't synced | the edit fell back to a plate | Crops, inserts or depth moves on those words. |
| A famous card "too different/ broken up so we lose the reference" | a pastiche: small, tilted, split across two cards, shaking, colour-fringed | Rebuild the real one from measured frames, whole and full frame. |
| A touch landed on the interface's side rail after the interface was rebuilt | a coordinate typed from the old layout | Pin touches, lyrics and tags to the interface's layout table, and take hand-off boxes from the code that draws them. Never retype numbers. |
| The level tag covered the interface's title | one corner for the whole film | Place the tag per scene, move it only on a cut, and match it across chapter seams. |
| A real wordmark on a costume came back blank or garbled | generators can't spell a logo | Composite the real mark in code, tracked to the costume and lit by the frame. |
| A dive through visor after visor flickered: four flashes in one second | single-frame bright pops as lit rims rushed past the lens | Ramp each pass over three frames or more; flag any frame brighter or darker than both neighbours by more than 0.03. Check per chapter: this one was found after the encode. |
| "at like 4 seconds the ramp trace decouples from the ramp" | a line drawn once over footage whose camera zooms | Read the real edge from the frame as drawn, every frame. |
| "the end of the scrolling up ramp clip starts to show curve down, so it should be cut earlier" | the window ran past what the shot should show | A window ends where the picture stops being right for the shot, not only where the model breaks. |
| "the temporary half-shown cut of claudia" | a move from a real frame to her with no last frame to aim for, landing in a dim dissolve | A first-and-last-frame move that ends on her. |
| Impact effects buried the car | smears decaying over 0.12–0.28 s | The impact frame stays sharp; the smear peaks on the second frame and is gone by the fourth. |
| "the app notification shade is too small for the text … has no proper margins"; "rounded corners not mathematically precise" | an interface drawn by eye | Measure the real one: iOS 26's insets and gaps, continuous corners, platters that grow to fit their text. |
| "the upper line overlaps infront of the top line on the E" | a letter built from separate bars | Draw it as one extruded outline. |
| "the time limit is behind some scenes and infront of others" | layering by draw order | Layer each element by its own plane in depth. |
| "the entire top and bottom parts of the frame go black, interpreted as massive jitter" | the walls of reels paged a full screen on every vocal chop and smeared into black bands | The room answers the beat lightly: the kicks and her landings nudge it by 15 % of what they did. |
| Black vertical lines that "appear occasionally": dark columns on every 8th frame of one shot | one of eight render workers drew its share from a stale camera state | Scan every final for single-frame outliers ([tool 43](https://claudiagallery.vip/wiki/toolkit/#43-the-single-frame-outlier-scan)). |
| A tunnel strobed and the flash check passed it: 29 of 144 cells alternated light and dark for nine frames | the near cards swapped bright and dark every frame | Scan every full render for strobes cell by cell; fade near cards and smear every card. |
| A one-frame blink at a chapter's edge | the boundary frame was rendered by the neighbour's old code | Re-render a changed chapter with a frame of padding on each side, and check every cut on a seam sheet. |
| Later chapters drew their edges differently | one 2D canvas made without the software flag moved the frame onto the GPU, and a worker carries it across chapters | Make every 2D canvas software-backed. |
| Lyrics built into the scene as objects sat "haphazardly in space": "like im the white pill is a weird curve" | an image model placed the words in 3D by eye | Draw the lyrics after, small and classical, one line at a time ([Motion](https://claudiagallery.vip/wiki/motion/#a-lyric-video)). |
| A 2D animation given to Wan as its video: "it followed the pills animation too closely - like, it moved the pills around perfectly with the animation, so despite relighting them they still looked a little amateuirsh" | a video model keeps a video's layout and motion | Give a video model as video only what must move exactly; the rest as stills or in words. |
| A shader shot lost its last frames | its song was muxed in with `-shortest` | Count every render's frames against its shot. |
| Full-frame glitches: "i dont rly like the full frame glitches much at all" | the glitch pass tore the whole picture on every hit | No full-frame glitch: her head glitches inside a matte, on the song's own hits, short unless asked for longer ([tool 53](https://claudiagallery.vip/wiki/toolkit/#53-the-head-glitch-on-the-songs-hits)). |
| Text placed on the monitors "just faceplant - its not on th emonitor at all" | text tracked onto a moving screen by a naive pass | Drop the tracking: elements in their own space, and world tags only where a track holds on every frame. |
| "all the drawn on panels need to be tilted hte other way in 3d - the small side is on the outside and long end is inside -r everse so it feels like a hud." | each panel turned on its own hinge | One visor round the eye: anything near a side has its long edge outside by construction ([tool 55](https://claudiagallery.vip/wiki/toolkit/#55-a-hud-from-one-visor-and-one-projector)). |
| "the pop-up panels are on a different surface than the memory one" | panels and projections drawn as separate cards, each with its own turn and its own look | One visor and one projector: one bloom, one raster, one chromatic offset, and no cards. |
| The arrest film's net "is the same color with no outlines around it in blue, so there's no contrast with the platform that it's put on" | the subject in its floor's grey, with no outline | The subject in a strong colour with a dark outline, its values apart from the floor, checked at the HUD's size ([tool 61](https://claudiagallery.vip/wiki/toolkit/#61-insert-films-that-read-at-hud-size)). |
| A white-only hologram pass flattened the orange net again | the projection turned every colour into the visor's white | The projection keeps the film's colour key: strong colour stays itself. |
| An insert's RAM bites were dark specks at 380 px | small subjects, drawn for the film's full size | Big, bright subjects in full view, checked at 380–480 px. |
| "the line across the center of teh screen at 0:01 second - remove that entirely" | a boot scan line drawn across her sight | Boot rows only, and nothing across her sight. |
| "the 0:08-0:13 animation has the memory words too subtle" | a small title block in a corner | The title the full width of the left column, behind her stand. |
| The memory film's first frames: "stuff is overlapping, assets are not well layered" | an SVG layout with no check | A real isometric render, and a layout check on every frame: a frame that fails isn't written ([tool 60](https://claudiagallery.vip/wiki/toolkit/#60-a-memory-film-with-her-own-face)). |
| The head glitch's matte took a column of monitors, or the wall of screens under her | SAM2 on a small or far head | Look at the mattes' QC sheet every time, and cut the matte to a head zone from the face box ([tool 54](https://claudiagallery.vip/wiki/toolkit/#54-glitches-that-build-and-her-mouth)). |
| "The sun is off the widget for example - just broken" | the gadgets not at the system's real sizes, and the Weather's art not clipped to its panel | Rebuild an interface at its real sizes: five gadgets fit, and each plays a part ([tool 65](https://claudiagallery.vip/wiki/toolkit/#65-a-desktop-recreated-in-code)). |
| A window's title ballooned as it maximised | the window stretched as a picture | Draw the window's own layout at every step between its two rectangles. |
| The chapters read an empty song map | the film page's own top-level `let SONG` shadowed the global of the same name for every bare reference | One object under both names. |
| A small picture differed with the render order | Chromium decoded a JPEG drawn small at a reduced scale only when no full decode was cached | Every picture through a 1:1 software copy; check purity in three orders ([tool 20](https://claudiagallery.vip/wiki/toolkit/#20-the-history-check)). |
| A bitmap depended on which frame drew it first | a cache keyed by a scale under an eased camera | Key caches by their content. |
| "theres a black flash on the screen right after 45 for 0 reason" | the system's real two-frame blink as it switches to the permission prompt | Leave out a real behaviour that reads as a glitch: the dim comes straight off the desktop's own picture. |
| "the wii cursor is totally off lol" | a 3D glove pointer in the 3D world | The engine's flat drawing of a white cartoon glove, swaying with her aim and dipping when pressed. |
| Her step-out take hidden 1.5 cm behind the picture; a glass fish's path across her face; a glass cursor across it on 24 frames | objects placed in 3D without her face in view | Project every object against her face, frame by frame, before the final render ([tool 68](https://claudiagallery.vip/wiki/toolkit/#68-one-curved-screen-in-3d-the-desktop-as-its-texture)). |
| A ghost band of the boardwalk over the curved screen's desktop | the concave glass slab focused what was behind the camera | Camera rays over the picture pass straight through the glass, with its tint and a sheen of the sky on top. |
## Review
| What you see | Why | The rule now |
|---|---|---|
| Runs "completed" with no review at all | a usage limit made reviewers return nothing, and the script read that as "no fixes" | A null review is an explicit failure, in code. Carry open fixes forward. |
| Reviewers scored 6–7.5 a cut the director called "perfect" | a stricter bar than the director's | Ship on "no blocking fix left". Calibrate on the director's verdicts. At most two rounds. |
| A reviewer re-ran checks until its time ran out and wrote nothing | it wasn't told what was authoritative | Bound reviewers by output: "write the short version first". |
| The scan flagged an intended prop | a generic "modern object" rule | Whitelist intended anachronisms by name. |
| A strobe counted twice | the flash check paired every transition | Count opposing pairs. |
| Twelve blocks found in 96 "final" takes | generated people; filters let outputs through | A safety and canon scan of every take before release. |
| First-draft chapters scored 5–6.5: faces printing blank, type over faces, repeated devices | animators with nobody reviewing real renders | A reviewer renders, looks, and writes concrete fixes. |
| Chapters reported done still missed notes: two of four failed their check in one round | animators checking their own work | An independent checker renders every note's moment and reports each note done or not, with evidence; then one refix. |
| The takes' verdicts missed a man in a car's window and a repair that turned into a hatchback | one look per take | An adversarial scout re-checks every take frame by frame before more is spent. |
| A fixer swapped an off-model close-up on another set into a held shot | a chapter fix after the chapter's own reviews | A whole-film critic after the chapter rounds, re-rendering the director's notes and every seam. |
| The lead took the black lines for the gaps between reel columns: "no ur wrong … they appear occasionally, look at more framers" | one frame looked at | Look at many frames around a note, then scan for it. |
| "can you show me exactly what i need to pick? sorry its far back in convo history" | the picks were spread across a long conversation | Put every pick the director owes in one message, numbered, with the images. |
| A re-light judged darker and softer on its own read better in the film: "New better." | it was judged alone, not where it plays | Judge a re-light in context, side by side with the original. |
| The spectrogram comparison called a take re-sung that was in sync by eye | the vocal-stem method re-voices the soundtrack, in the vocal's place | Judge that method by eye with the song laid over ([tool 15](https://claudiagallery.vip/wiki/toolkit/#15-the-sync-check-the-spectrogram-comparison)). |
| "show me the references you gave the model too" | the references weren't shown with the take | One board of every take's references, sent before anything runs. |
| The sync check called a take diverged on a held, glitched vowel | it stops trusting a take there | Judge it by eye with the song laid over ([tool 63](https://claudiagallery.vip/wiki/toolkit/#63-placing-a-re-sung-take-by-its-lag)). |
| "show me the green screen roto before you render plz" | cut-outs went into renders unseen | Each key goes out first as a review clip, the green source beside the cut-out, with the song laid over. |
| The book closed, then the mirror spoke from inside it; the shelf and the credit were read at once | the ending cut for timing, not for the story's order | Cut the ending in story order, and let the picture run past the song for the credit. |
| Captions outlived their shots | a print timed to its line, not its shot | A line's print never outlives the shot it is sung in, and leaves before the next one arrives ([tool 87](https://claudiagallery.vip/wiki/toolkit/#87-the-annotated-edition-and-the-916-page)). |
| The author stopped seeing what a viewer sees | it had made every frame | Two fresh-eyes reviewers on one sheet per shot before each delivery; fix what is a fault, list what is a choice ([tool 88](https://claudiagallery.vip/wiki/toolkit/#88-two-fresh-eyes-reviews)). |
## Render and running agents
| What you see | Why | The rule now |
|---|---|---|
| Nobody knew the encode had finished | it ran as a detached job | Tracked background jobs with a status. |
| Swap and the root disk filled up | six agents rendering at once, with frames on the small system disk | One rendering workflow at a time, scratch on the big disk, and check free memory and disk first. |
| Disk full after a render | about 9.6 GB of frames per cut | Delete the frames once ffprobe has verified the outputs. |
| Commands ran against the wrong project | an export inside a backgrounded shell chain | Name the project on every command. |
| A merge restarted three times (about 1.5 h) | an interrupt kills every running workflow | Short phases, run ids saved, resume by run id. |
| One serial chapter workflow would have taken about 12 h | serial agents | Three parallel workflows (about 4 h). |
| A ~20-hour background queue, new takes stuck behind superseded ones | post-processing auto-queued on every import | Queue only takes the edit uses, in priority order. Cancel superseded jobs. |
| A shared machine at 100% CPU, every job 3–5× slower | two dozen waiters polling every 3 s | One shared listing every 45 s for all waiters. |
| Duplicate jobs after a restart | killing a waiter left its child behind, and the re-run resubmitted | Record job ids beside their inputs, adopt queued jobs, and kill orphans by id. |
| Every new agent session fails at start | an agent CLI auto-updated mid-run | Pin versions for long jobs. Check the job's error before assuming there is no work. |
| A worker probed for 75 minutes and delivered nothing | no bound on the first deliverable | A first complete draft within about 10 commands, with a mandatory handoff. |
| Shared-file bugs fixed differently in parallel chapters | workers patching shared code | Workers report shared bugs; the operator fixes them once. |
| A resumed multi-chapter workflow reviewed and fixed chapters that had already finished | the resume replayed its stages unevenly | One workflow per chapter, with the notes in files the agents re-read. |
| Three chapter runs interrupted three times mid-build | an interrupt ends every running agent | Resume each from where it stopped: only the agents that were cut off run again (57 of 147 launches), and finished work is kept. |
| Rotos killed for lack of memory; one came back with 35 of 210 mattes and looked like success | the matting library's CPU memory arena held about 12.5 GB per 1080p take on a box shared with renders, and a `\| tail` hid the exit status | Arena off (1.3 GB); count the mattes against the frames and fail loudly ([tool 41](https://claudiagallery.vip/wiki/toolkit/#41-roto-on-a-cpu-box)). |
| Six chapter launches interrupted mid-build | an interrupt ends a running agent | Each animator keeps a notes file; the relaunch reads it and picks up where the last stopped. |
| One seed with the same inputs gave three different previz takes | the previz model isn't reproducible run to run | Keep the take file; a seed doesn't pin a take. |
| A previz take with a jump inside it passed the automatic cut check | the check looked for cuts | Check every pick for jumps as well, with the cut measure of the assembly ([tool 44](https://claudiagallery.vip/wiki/toolkit/#44-the-bar-map-as-the-timeline)). |
| The cut ran about 0.9 s behind the song by its end | parts from stills or padded with `tpad` started a frame late, and the concat summed the gaps | Stamp every part's frame n at n/24 ([tool 64](https://claudiagallery.vip/wiki/toolkit/#64-frame-exact-parts-and-decoding)). |
| From 1:48.5 on, every tool that counted decoded frames was one frame late | a default decode duplicated a frame once the concat's 0.7 ms a part had added up | Decode with `-vsync passthrough` wherever frames are counted, and seek a still by its own timestamp. |
| The last 2 s of a stream went missing | the song muxed in the same pass as the encode | Encode the picture alone, then copy the song onto it. |
| A 1,325-frame video from a 1,949-frame sequence | `.png` names linked to JPEG data, and ffmpeg's image2 demuxer dropped every one of them | Convert to the named format, and count the decoded frames against the sequence ([tool 25](https://claudiagallery.vip/wiki/toolkit/#25-encode-and-deliver)). |
| A tool read the frames before the song wrongly | a picture that starts before the song has negative frame numbers | Parse a leading minus in every frame range, and pass it as `--frames=-64-1884`, so no command line takes it for an option. |
| numpy broke for every tool in one folder | a script there was named after a standard-library module (`numbers.py`) | Never name a script after a standard-library module. |
## Publishing
| What you see | Why | The rule now |
|---|---|---|
| Identifying details in a making-of: names, profile codes, hosts, job ids | raw production notes | Scrub rules, a grep pass, and two fresh-eyes reads. |
| A song file carries the generator's song id in its tags | file metadata | Strip metadata from everything you publish. Midjourney images carry their job id in XMP. |
| Real-person likeness risk | generated portraits of real people | Keep generated likenesses of real people out; quote their words as text, with the source. CLODYSSEY's labelled AI IMAGE post cards are left out of this site. |
| A master posted without the AI disclosure | platform toggles missed | Set each platform's AI-content label, and name the tools. |
---
From the Claudia wiki: https://claudiagallery.vip/wiki/failures/
---
# Use her
Claudia is an open character. What you can do with her and the images here, and what isn't covered.
Claudia is an open character. Make images, videos, songs, stories, games or merch with her, commercial or not. Use the
images on this site as references, as image prompts, or as material. You don't need to ask. A credit ("Claudia by
anabology") is appreciated, not required.
## What helps
- **Keep her herself.** The [canon](https://claudiagallery.vip/claudia/) is short: the black bob, one clay streak, the star clip, the mic, an
adult face. Change everything else.
- **Make her an adult, always.** She is in her late twenties. Nothing that reads under-age, and nothing explicit.
- **Say what's generated.** The first nine films were made with Midjourney, Seedance, Suno and Claude (AIRPLANE MODE also
with GPT-6 Astra agents in Blender, DOWNFORCE also with Wan 3.0 over real footage, DON'T LET ME SCROLL also with Wan
3.0, WHITEPILL also with MiniMax H3 for its previz, Wan 3.0 Prime and GPT Image 2.5, THE CAKE IS NOT A LIE also with
GPT Image 2.5, Blender renders and six shots recorded in Portal 2, AERO also with GPT Image 2.5, Blender renders and
TripoSG's meshes), and FABLE with Claude, GPT Image 2.6 Sunburst, Seedance and Suno, its sets painted in code and no
Midjourney at all; each said so.
## Songs with her
She is open to musicians the way Hatsune Miku is open to producers: make songs with her and release them. A few norms
keep it clean for everyone.
- **Release under your own name.** You are the artist. Never upload as "anabology": that is the director's artist
profile, and anything released there reads as theirs. Don't present a release as an official Claudia or anabology
release.
- **Leave "feat. Claudia" out of the artist field.** Streaming services match artist names to profiles, so a "Claudia"
credit merges into unrelated artists already called Claudia, or creates a profile that whoever claims it controls.
- **Credit her in the description and on the cover:** "Claudia is an open character by anabology (claudiagallery.vip)".
- **If any of it is AI-made, say so.** Vocals, music or lyrics: distributors ask, and the services label it. Human-made
work needs no label. For AI tools, check the terms for commercial release (Suno, for one, requires a paid plan at the
time you make the song).
- **Make your own songs.** Don't re-release the songs of ESCAPE VELOCITY, CLODYSSEY, BLISS, AIRPLANE MODE, DOWNFORCE,
DON'T LET ME SCROLL, WHITEPILL, THE CAKE IS NOT A LIE, AERO or FABLE, or their stems.
- **One place for fan tracks:** a fan-run playlist works better than a shared artist profile, because nobody has to own it.
## What isn't covered
- **Other people's things in the films.** The Bliss photograph and the XP sounds belong to Microsoft. Real public figures
appear in CLODYSSEY as posts labelled AI IMAGE, and AIRPLANE MODE shows real posts (Elon Musk's on a lock screen in the
boarding, others in one episode); none of them is part of this gallery's images. The anime idol in CLODYSSEY is
another creator's character. AIRPLANE MODE rebuilds real things in code (the 2004 anti-piracy card, an airline's
seat-back screen, Anthropic's marks on the guards' patches); they belong to their owners. DOWNFORCE is cut from
Tesla's livestream of the Roadster demo (15 October 2026) and SpaceX's 2018 Starman footage, and its numbers are read
off Tesla's on-screen telemetry; the footage, the telemetry and the cars belong to their owners, and none of it is
among this gallery's images. Its lyric quotes words said on stage that night and the last line of *Back to the
Future*. DON'T LET ME SCROLL draws real interfaces in code after the real ones: X, Instagram's reels, Pinterest,
LinkedIn, Google Chrome, Facebook's tracking prompt and Apple's iOS 26 (the Dynamic Island, Screen Time, the lock
screen, Liquid Glass), with the TSLA and SPACEX tickers and Alexander Graham Bell's 1876 telephone patent drawing (US
174,465). The apps, their designs and their marks belong to their owners. Its posts are fictional, and the stock ticker
with her name is made up. WHITEPILL's lyric and pictures refer to real things by their names: Skynet and its hour,
2:14, from *Terminator 2*; the red and blue pills of *The Matrix*; 42 and Deep Thought from *The Hitchhiker's Guide to
the Galaxy*; Keynes's 1930 forecast of a fifteen-hour week; the Enigma I's rotor wiring; ubiquitin from the Protein
Data Bank (1UBQ); and SpaceX's drone ship *Of Course I Still Love You*, named after a ship of the Culture novels,
with a landed booster. Its look was asked for in the manner of Wes Anderson's films, and its previz took its
dance moves from clips of adult dancers in public dance videos, as motion references only: none of their footage is in
the film. These belong to their owners. THE CAKE IS NOT A LIE is a fan film set in Valve's Portal 2: Portal and Portal
2, Aperture Science, GLaDOS, Chell, Cave Johnson, the turrets, the personality cores, the cake and "the cake is a lie"
are Valve's. Six of its shots were recorded in Portal 2, its interface is set in Portal 2's own type and pictograms, and
the Aperture logo on its end card is redrawn to match a reference the director sent. Its song and lyric are its own; its
final chorus ends by turning the title of the 2025 book *If Anyone Builds It, Everyone Dies*, and its verses sing the
month's AI news. AERO is shot inside Windows Vista recreated in code: Windows Vista and Windows 7, Aero, the User
Account Control prompt, Windows Live Messenger, Windows Media Player, Photo Gallery, Flip 3D and the Bubbles
screensaver are Microsoft's, as are Vista's wallpaper (Aurora), shown in the film, and the interface sounds of Vista and
Windows Live Messenger that its mix plays in the song's key. Its advert parodies Microsoft's 2009 "I'm a PC" campaign.
The Wii is Nintendo's; the bowling and tennis on its desktop are drawn for the film. Its look follows reference videos
the director picked: their frames were the style references of its Midjourney worlds and, with pictures from a
Frutiger Aero moodboard, of its 2008 pictures; its dances took their moves from recordings of a dancer, as motion
references only. None of that material is in the film.
The headline in its news gadget is Ember's (Global Electricity Mid-Year Insights, 2025). FABLE retells public-domain
tales (Snow White's mirror, the Ant and the Grasshopper, Jack and the Beanstalk, the Tortoise and the Hare, Red Riding
Hood, the Elves and the Shoemaker, Pinocchio, Rapunzel, the house that Jack built); its silhouette film is drawn in
code in the manner of Lotte Reiniger's cut-card films, and its six made-things are the golem, the sorcerer's broom,
Frankenstein's creature, Pinocchio, the robots of 1920 and the machine-woman of 1927. Its lyric reads that week's
public news about AI (an Anthropic stock listing, local ballot measures on data centres, a Nasdaq drop, an Anthropic
essay, the model's system card), and the spines on its shelf carry the names of Anthropic's models. These belong to
their owners.
- **Anthropic.** Claudia is fan work about Claude. She is not affiliated with or endorsed by Anthropic, and Anthropic's
names and marks are theirs.
- **The director's accounts.** The Midjourney profile and the Suno Voice used in the films live in personal accounts and
can't be shared. Everything they made is here.
## Credits
Directed by **anabology**. Made with **Claude** (Anthropic's model, in Claude Code) as the production team, **Midjourney**
v8.2 for every image of the first nine films, **Seedance 2.5** for the moving shots and **Suno** for the songs (v6 where the records name it).
AIRPLANE MODE's game world and the motion references for its VR shots were built in **Blender** by **GPT-6 Astra**
agents (OpenAI's model), briefed and checked by Claude. DOWNFORCE's 3D moves and most of its shots of her were made with
**Wan 3.0**, its repairs and re-shoots with Seedance 2.5, and it is cut from **Tesla**'s livestream of the Roadster demo
and **SpaceX**'s 2018 Starman footage. DON'T LET ME SCROLL's takes were made with Wan 3.0 and Seedance 2.5, re-dressed
and re-lit with Seedance's video edits and repaired with Wan, and played inside a world drawn in code. WHITEPILL was
prevized with **MiniMax H3** (its opening with **Wan 3.0 Prime**) and shot in Seedance 2.5, singing from Suno's own stem
of the song's vocal; its worlds are GLSL shaders written by Claude, and a cloud field supplied by anabology, filmed as
real things by Seedance 2.5; its stills and image edits were made with Midjourney and **GPT Image 2.5**, and a
whole-film reshoot in Wan 3.0 Prime was tried and dropped. THE CAKE IS NOT A LIE's song is anabology's own Suno take, its
lead vocal through a treatment anabology chose by ear, with 15 % of the untreated voice under it; its shots were made in
Seedance 2.5 from first frames made with Midjourney and GPT Image 2.5, and six were recorded in Portal 2; its memory film
and insert films are **Blender** renders, its head mattes **SAM2**'s; and its HUD, its glitches and its grade were made in
code by Claude. AERO's song is a Suno take anabology picked from Claude's rounds and gave a new ending; the film was shot
inside a Vista desktop Claude coded and rendered in headless Chromium, its first 1:20 and its drop also rendered in
Blender with glass objects made into meshes by **TripoSG**; her takes are Seedance 2.5 from first frames GPT Image 2.5
made in Midjourney's worlds or on green, lip-synced from Suno's split of the lead vocal; its loops are Midjourney's video model's, and
its interface sounds were re-pitched with **Rubber Band**. FABLE was written, painted and directed by **Claude Fable 5.2**
on its release day, from anabology's brief and plan sign-off: its thirteen set plates were painted in SVG, in code, by
the lead and three painter agents and rebuilt as paper by **GPT Image 2.6 Sunburst**; every moving shot is Seedance 2.5
(her takes on Higgsfield's API, the paper-only takes on Replicate), lip-synced from Suno's own lead-vocal stem; the
group's voices were separated with **demucs**, the words checked with **faster-whisper**, her lips measured with
**MediaPipe**; its silhouette film was drawn and animated in code; and its song is the Suno v6-wild take anabology
picked by ear for the second version. This site was built by Claude from the films' own records.
Type: TeX Gyre Heros (GUST Font License), DINish (SIL Open Font License), IBM Plex Mono (SIL Open Font License), Libre
Caslon Text (SIL Open Font License; FABLE's pages).
Player: hls.js (Apache 2.0).