Character Sheet 101

A character sheet is a set of reference images of one character, built before any scene is generated, that every later shot is checked against. It carries the face, the body, the wardrobe and the emotional range the story needs. When a character drifts between shots, the cause is usually upstream: the sheet was incomplete, noisy, or built from the wrong starting image.

We covered what keeps a character stable during generation in Character Consistency in AI Video. This piece is the step before that: building the reference those techniques rely on.

What goes on the sheet

A complete sheet has five parts. Skip one only when the film genuinely never needs it.

• Identity panel. Head and shoulders, neutral expression, mouth closed, looking straight into the lens. A portrait lens around 85mm keeps proportions natural. This is the face lock.
• Full body panel. Standing, arms at the sides, full silhouette from collar to shoes. When it shares an image with the identity panel, show the figure from the neck down with no head. A sheet should carry exactly one face. A second, slightly different face on the body panel gives the model two faces to blend.
• Expression or state panels. One for every emotional peak the script depends on: the scream, the shock, the smirk. A model given only a neutral face invents a weaker version of the peak on its own.
• Wardrobe panels. One clean pose per outfit the character wears across the production.
• Turnaround. Front, three quarter, profile and back, when the camera will see the character from those angles.

Write a short text description alongside the images: face marks, hair, each garment. Keep it concrete. A long, adjective heavy description gives the model more room to interpret each view differently, so it holds worse than a short, precise one.

Build it on grey

Every panel goes on a flat mid grey seamless background with soft, even frontal light, colour accurate, no grade, no lens flare, generated in the aspect ratio of the film. No environment, no text, no labels, no watermarks.

The grey isolates identity from everything else. Any background detail on a reference image becomes one more thing the model may try to carry into the scene. Burned in labels are worse: they are a visual trait like any other, and they leak.

The grey is a reference convention, not the look of the film. Lighting for the actual scenes is set in the video prompt, not inherited from the sheet.

Pick your starting path first

How you build depends on what you start with. Choose before generating anything.

• One reference photo. Go straight to the turnaround and panels from that photo.
• A written description only. Generate one seed identity image, approve it against the description, then build every other panel from the approved seed. This approval step is the one most often skipped under deadline, and skipping it means finding the error only after the whole sheet is built.
• Several partial references, such as a face photo plus a separate outfit photo. Reconcile them into one written description first, then proceed as if starting from a description. Merging two source images directly produces a character that matches neither.

Build order

Approve the face before anything else. Every later panel inherits it, so an unapproved face multiplies its error across the whole sheet.

Generate expression and wardrobe panels from the approved identity image as an image reference, not from text alone. That is what keeps the face the same when the expression changes.

With more than one character, finish and approve one sheet before starting the next. Building two characters in the same session, interleaved, lets traits leak between them: a skin tone, a jawline, a jacket. Before finalizing the second sheet, spot check it against the first.

Build order

Every later panel inherits the approved face

Then rewrite the text to match the images

Generations rarely match the brief exactly. On our short film Last One, the hero was specified in a grey tee and sneakers and came out in a black crop top and chunky boots. The images were right for the film, so the written description was rewritten to the crop top.

The rule: the prompt describes what the sheet shows, not what you originally asked for. If the text and the image disagree, the model has to choose, and you will not like every choice it makes.

When you regenerate a character, stop attaching the old sheets. Two versions of one person in the reference set get blended into a third.

The checks that catch drift early

• Expression drift. For every expression beyond neutral, confirm the face shape, proportions and skin tone held and only the expression changed. This is the most common place a face drifts. Regenerate the one view that shifted rather than accepting it.
• Invented angles. A front only source photo forces the model to invent the profile and the back. Those invented views then pose as confirmed reference. Flag them and approve or correct them explicitly.
• Extending mid production. A new outfit or expression gets generated from the approved seed or an approved sheet view, never from the latest scene output. Building from the last generation is how drift compounds across a project.

Name every view

Give each image an ID in the form CHARACTER, VIEW, VARIANT: LOLLY-FRONT-NEUTRAL, LOLLY-WARDROBE-JACKET, LOLLY-EXPRESSION-SMIRK. If your platform stores references as named elements, give the element that exact name so the ID in your brief is the same string you call in the prompt. Log the platform's own element ID too: some platforms quietly add a number to a duplicate name, and the ID is the only handle that never changes.

In the prompt itself, refer to references by role (@HERO, @VILLAIN) rather than by position (@Image1). Positional references break the moment the attachment order changes.

Where to save credits

A head and shoulders sheet alone is enough when the body is barely on screen. Spend the credits on the video generations instead. The expression panels are the ones not to skip: they cost one image each and they decide whether the emotional peak of the film lands.

Ready check

Before a sheet goes into any brief, answer yes or no to each line. One no blocks it.

• Every required view is present.
• Every view is on flat grey, with no environment and no text.
• Expressions checked for identity drift, none left uncorrected.
• Every outfit the script needs is covered.
• Every image has an ID logged.
• For more than one character, the cross check is done.

This is the same sheet discipline behind Last One, a short film whose sequences were each generated in a single pass with every character, prop and location referenced on grey. The sheet is the cheapest part of the production and the one that decides whether the rest works.