A character sheet before the first video credit

Lock face, wardrobe, palette in stills. Video credits are the expensive way to discover the character drifted.

Still first, motion second. That is the same doctrine as storyboarding brand videos with AI images first — this page is the credit argument for the person (or mascot, or recurring presenter), not the sequence board. The routing rule that sits on top of the sheet is lock the still before you spend on motion. Do not start a video job to find out what she looks like.

Why the first video credit is the wrong place to cast

A video generate has to invent a face and move it and, on audio models, decide how it talks. Those are three meters. Identity is the cheap one to settle, and the expensive one to fix, because every retry re-pays motion (and often audio) to sample another jawline.

Text is a lossy casting call. "Woman, 30s, brown hair, green coat" is a demographic. It is not a person. Two generates from that paragraph are two actors. A weekly series from that paragraph is a recast every Monday. You will notice on episode four, when the client notices on episode one.

Image credits are the right meter for that argument. Generate faces until someone can point and say "her." Generate the coat until it is that coat. Generate a palette strip — three stills in the same light — until the world is one world. Then, and only then, spend a video credit on making that picture move.

Uploading the approved stills into a workflow does not spend generation credits. Set up reusable characters and products is that step: open the asset cards, bind face / product / room by key, reuse across scenes. The sheet is the files. The agent is how the files stop living in a chat.

What "a sheet" actually contains

Not a collage. A small set of stills that each do one job.

• Face. Clean, well-lit, front. Neutral expression plus one in the performance you will actually use (smile, rest, talk). This is the identity contract.
• Wardrobe. The coat, the uniform, the colourway that must not drift. A second angle if the back of the jacket matters.
• Body / framing. If any shot is full-length, you need a full-length still. A face sheet will not survive a wide.
• Palette / room. One establishing still of the space and the light. Video models love to relight. Give them a room they can copy.
• Product, if any. Pack-accurate, label-legible. Viewers forgive a cheekbone. They do not forgive a word count on a bottle.

That is five stills, not fifty. Seedance 2.5's reference ceiling is a kit for people who already know what each slot is for. A sheet is the minimum kit that would let you recast this character tomorrow without a meeting.

Approve the sheet the way you approve a board: someone signs "that one." Unapproved stills are prompts with extra pixels.

The credit arithmetic, without fake prices

Image generates are cheap and fast. Video generates are not. The exact ratio moves with the model; the order of magnitude does not. Every identity decision you still have open at video time is a motion-priced coin flip.

The failure that looks like a production day: prompt a talking take from a paragraph; face is almost right; regen; coat changes; regen; room changes; face is a cousin. The afternoon is video credits spent hunting a person who never existed as a still.

The day that looks like less work: image tries until you can point at a face, a coat, and a room; save them as assets; one I2V or R2V from the sheet, with a motion prompt that describes motion only. If that generate fails, you have a motion problem. You do not have a casting problem. That is the whole saving.

Sheet versus board versus first frame

Sheet — who they are, across jobs. Reusable. Lives in the workflow as keys.

Board — the sequence of story frames for one film. The storyboarding post is that object. The sheet feeds the board; the board does not replace the sheet. Episode two should not require a new face.

First frame — the locked still for this shot. I2V's contract. The sheet's hero face is often this frame; it is not always. A new angle is a new still, generated from the sheet as reference, approved, then animated.

Do not skip the sheet because you have a board. A board without a sheet is fourteen almost-hers. Do not skip the first-frame lock because you have a sheet. A sheet without a first frame is a casting call into a new room.

When you may spend a video credit first

Motion is the idea: a whip-pan, a physics gag, a dance. Then a cheap, fast video model is a sketch, not a character. Pull a frame from the sketch, still it, put it on the sheet, and then spend the real motion credits. The sketch was research. It was not the series bible.

Single-shot avatars with no recurrence: if she will never appear again, a talking-avatar take can be the whole job. The moment there is a tomorrow, there is a sheet.

FAQ

How many stills before I am allowed to generate video?

Enough that the things which would force a regen are already decided: face, wardrobe, and any product or room that must match. For a talking head in one interior, that is often two or three stills. For a series, it is the full list above, signed. Count stills by decisions, not by a quota.

Does a good R2V model mean I can skip the sheet and just keep using one photo?

One photo is a sheet of one. That is enough for a simple identity. It is not enough for wardrobe changes, a second location, or a pack. When the job outgrows one photo, add a still — at image prices — before the next video job. Do not add it as the next video job.

Should the sheet be photoreal if the films are illustrated?

The sheet should be in the look of the deliverable. A photoreal face feeding an illustrated series is a recast every time. Lock the look the way you lock the face. Then reuse it.

Where do the files live so the next scene actually sees them?

In the workflow, as reusable assets, not in a Slack thread. Set up reusable characters and products binds the stills by key so every scene can call the same face. A folder of JPEGs the prompt "remembers" is not a sheet. It is a hope.