Character sheets: how a character stays the same across many shots

Nobody needs convincing anymore that a generated character needs a character sheet before it moves. The real work is in how you build it. The video model only knows the character from this one image. Whatever is missing or can be read two ways, it fills in itself, and slightly differently in every shot: in the second, the nose sits a little differently; in the third, the coat has new buttons.

So it pays to make every decision on the sheet deliberately: which views, how large the face, how many faces at all, whether there is text in the image. A mistake in the sheet comes back in every shot the character appears in. The sheet goes into the video model as an image reference, and the prompt defines how to read it: one person from several angles, not several people. Which image model creates it matters less. Nano Banana, GPT Image and Flux all handle the same setup.

Variant one: three views

For Pocobello, our short comedy about one last box of popcorn, every character got a lean sheet, the dog included.

Front: the full figure with proportions, costume and posture. Back: the same figure from behind. Without this view, the model fills in the back however it sees fit as soon as someone turns around. Three-quarter profile: head and shoulders, turned 45 degrees and filling the frame.

The three-quarter profile is actually the most important of the three tiles. From the front, a face looks flat. Only from this angle does the model grasp its depth: how far the nose projects, where the cheekbones sit, how the jawline runs. That is exactly the information that carries the identity when the character turns in the video.

Whatever must be recognisable from behind belongs explicitly in the prompt. For Cordelia, it is the braid around her tower of hair and the small brass key pinned into it; for Ignatz, the razor-sharp parting and the sideburns that jut past his silhouette even from behind.

A classic, and an annoying one: image models like to add a second, small head to the front view, a miniature portrait next to the figure. On a reference, that is fatal, because the video model then reads two faces where it should learn one. Strictly speaking, the same goes for the head in the full-body view. In the sheets above it is covered in grey, so a single face remains: the large one in three-quarter profile.

The prompt for Ignatz, as we used it. The first paragraph defines the grid and stays the same for every character. From «The character» on, it is your own.

Character reference sheet, three views of the same man side by side in one image, evenly spaced, identical lighting, thin neutral gap between tiles. Left tile: full body, front view, neutral standing pose with arms relaxed at his sides. Center tile: full body, back view, same pose, seen from directly behind. Right tile: head and shoulders close-up in three-quarter view turned 45 degrees to his left, framed to fill the full height of the tile so the face is rendered large and detailed.
The character: a man in his late forties, neutral relaxed expression, no smile. His hair is a rigid side-parted helmet of shoe-polish-black hair with an implausibly sharp part line — the part line and the combed lacquered layering must read clearly in the back view — with enormous flared mutton chop sideburns reaching almost to his chin, visible jutting past the silhouette even from behind, and a thin waxed moustache curled into two tight spirals. He wears a deep petrol-blue safari suit with four oversized patch pockets and a stiff epaulette on each shoulder, a pale butter-yellow ascot, matching petrol-blue trousers with a sharp crease, a wide brown leather belt, and scuffed tan desert boots. A small pair of brass binoculars hangs on a leather cord around his neck.
Background: flat neutral mid-grey seamless studio backdrop, no props, no shadows on the wall.
Lighting: soft even studio lighting from front and both sides, low contrast, neutral white balance, identical across all three tiles.
Ultra-photorealistic, visible pores and razor stubble texture at the jaw, shine on the forehead, individual coarse sideburn hairs, gabardine twill weave with soft creases at the elbows and knees, scuffed brass, worn leather.
Sharp focus throughout, 85 mm lens look, no depth-of-field blur, all tiles same height with a small margin above and below, 3:1 aspect ratio.

Variant two: the full board

Right now we are testing a second, far more extensive variant: a board with more views, detail tiles and labels. The idea behind it: the more information the sheet holds, the less the video model has to invent. Whether the extra effort is needed, we don't know yet. The lean sheet defines who a character is. The board is meant to also define what they are made of.

The top row shows five views: front, profile, back, a large close-up and a top-down view. Below come detail tiles for exactly the features a model would otherwise even out: hair texture, beard, eye colour, skin texture and the one element that defines the character, for the ninja the sword with the throwing star. A colour palette with hex codes completes the board. Even what is hidden gets defined: the ninja's curls sit under the hood and are specified anyway, for the moment the hood comes off.

Here, the labels are part of the layout. They appear in the prompt in quotation marks, and a separate paragraph rules out any other text. GPT Image puts out labels and hex codes the cleanest. For Flux and other local models, we shorten the prompt, because long text loses precision there.

The board's grid, with placeholders in square brackets:

A photorealistic character reference sheet of ONE single person, landscape 16:9, laid out like a professional casting and costume reference board on a seamless light warm-grey studio background, soft even studio lighting with gentle shadows, identical light in every view. Thin, even gutters between all panels, everything on the same background.

IDENTITY – absolute: [With reference image: The person is exactly the person in the reference image(s): same face, bone structure, eye shape, nose, lips, skin tone, hairline, hair and body proportions. Do not beautify, slim, de-age or idealize.] [Without: describe sex, age, face, eyes, skin, hair, facial hair, build, height.] Distinguishing features that must be visible wherever that body part is in frame: [tattoos with motif and position, piercings, scars, moles …].

OUTFIT – identical in every view: [every piece with colour, material, fit; prints and logos with exact text and position front/back; shoes or barefoot].

TOP ROW (about 70 % of the sheet height), five views side by side, each with a small clean uppercase sans-serif label centered above it:
– "FRONT": full body, standing straight facing the camera, arms relaxed at the sides, neutral expression, head to feet fully visible.
– "SIDE": full body in exact left profile, facing left, same relaxed pose.
– "BACK": full body seen from directly behind[, showing … (back print, hair at the back)].
– "FRONT CLOSE-UP": the largest panel, head-and-shoulders portrait looking straight into the lens, [expression], about twice the scale of the full-body views, every pore and hair sharp.
– "TOP-DOWN VIEW": camera directly overhead looking straight down at the person standing, showing [the top of the head / the cap from above], the shoulders and the feet.

BOTTOM ROW (about 25 % of the sheet height), square detail close-up photos, each with a bold uppercase label above it:
– "[LABEL 1]": [specific macro crop].
– "[LABEL 2]": [...].
– "[LABEL 3]": [...].
– "[LABEL 4]": [...].
– "[LABEL 5]": [...].
– "COLOR PALETTE": [5–6] flat colour swatches in [two rows of three], each with its hex code in small type below: [#XXXXXX (hair), #XXXXXX (skin), …].

STYLE: photorealistic studio photography, medium-format camera, 80 mm lens, natural skin texture with visible pores and small imperfections, fine film grain, true-to-life muted colours, crisp clean graphic layout.

TEXT: Only the labels listed above appear, spelled exactly as written[, plus the hex codes]. No other text, no watermark, no logos except those on the clothing.

NEGATIVE: Exactly one person, the same person in every panel – never a different face, never two people in one panel. No changes of outfit, hair, facial hair or accessories between views. No CGI look, no plastic skin, no HDR, no dramatic or coloured lighting, no props, no scenery.

From drawing to film

I first tried the detailed sheets at home, together with my daughter. We didn't start with photos but with her drawings: a guy with a cardboard sign around his neck, his perfectly round dog and a rather cheeky bird. The goal was to turn them into characters that could carry a whole film.

The first step translates the drawing into a look without touching its design. The prompt opens with: «Use the attached hand-drawn sketch as the exact design reference. Keep every design element of the sketch, only upgrade the rendering.» Only then do the sheets follow.

Animated characters live on exaggeration, and the video model has to know its limits. Their sheets therefore show the turnaround at the top and the expressions below: bored, grinning wide, shocked, embarrassed. The dog gets poses as well, from a suspicious side-eye to a jump to curling up into a ball. That way the model knows how far eyes and mouth may go before the character stops being itself. The facial features stayed remarkably stable across every shot.

For the bird, the model received two images with separate jobs: the drawing for the shape, a finished image for the style only. The prompt states that nothing else is taken from the second image. The sheet adds flight poses, because a bird rarely stands still in a film.

Starting from a drawing has another advantage that goes beyond craft: it does not depict a real person. Questions about the right to one's own image or personality rights simply don't come up. With real people, we only work with their consent. The ninja above has agreed. It is my face, after all.

And this is what came out of it.

From practice

• Neutral expression in the views. Expressions get their own sheet when needed. The views define identity only.
• Describe the material, not just the colour. Gabardine, brass, leather: the more precise the surface, the less the model interprets in motion.
• The pose follows the body. Oswald cannot let his arms hang straight. His prompt therefore asks for arms slightly away from the body, both hands open and visible.
• Text only on purpose. None on the lean sheet, or words like «Panel 1: Front» end up in the image. On the board, exactly the labels from the prompt.
• One sheet per outfit. If a character changes clothes, it needs a second one.
• Set the format in the model. Nano Banana 2 ignored the «3:1» in our prompt and returned 16:9.

Where the sheet stops

A sheet defines how a character looks, not how it performs. Places get plates of their own. And even the best sheet does not replace checking every single shot.

• Miraki: Pocobello