Consistent AI Character

What is a consistent AI character?

Image models do not remember people. Every prompt starts from nothing, so “a pale woman with a scar” becomes a different woman each time. A consistent character is the workaround: you generate the person once, carefully, and from then on you show the model who she is instead of telling it.

In the source experiments, the author organized this in a character profile called a cameo. It holds four things: the character sheet (the person from every side), shots from reference collections (poses, emotions, locations, haircuts, makeup, camera angles, light, props, colour), personality and voice, and story states — how the character changes across a film. The source workflow compiled prompts from that profile to keep the person consistent.

The examples below follow one character, Mara Vale: 29, lean, grey-green eyes under heavy lids, freckles across the nose, a thin scar through the outer end of her left eyebrow, a worn olive bomber jacket. Everything else — the snarl, the crouch, the red light, the overgrown corridor — came from reference images.

Start with the face — from an emotion reference

The very first image is the only one generated without an identity picture, so it has to invent the person. Two things go in: a concrete face inventory in words — face shape, brows, eyes, nose, mouth, jaw, skin, asymmetry — and one image from an emotion collection that gives the face its attitude.

The emotion image is dangerous: it shows a real face, often a recognizable one. The prompt has to say twice what it is for. The identity block says “an invented face, not any real actor, model or celebrity, and not any person in the reference pictures”; the role block says “take the facial expression only … do not take the face, age, hair, makeup, skin, clothes”. Generate four candidates in one request and keep the cleanest front close-up. In our run two of the four had also copied the crouching pose of the reference — a reminder that a full-body emotion reference leaks body language.

The character sheet: two anchors, then every side

The chosen face becomes the front close-up anchor. The second anchor is a front full body generated from it: head to boots, arms relaxed, neutral grey seamless. From now on these two images go into every prompt as pictures 1 and 2, and the face is never described in words again — the inventory keeps only what the anchors cannot guarantee from every angle: skin, the scar, hair, build and wardrobe.

The sheet has three levels:

• Base — front close-up and front full body. Enough to start working.
• Full — adds the left profile, the right profile and the back for both close-up and full body.
• Extended — adds three-quarter views, three waist-up shots and details that must repeat from shot to shot: eyes, hands, distinctive marks, shoes.

Every view gets the same studio block: plain mid-grey backdrop and floor, soft even front light, neutral expression, empty hands, no props. That neutrality is the point — the sheet is a reference for the model, not a portrait.

Here each view is its own image. To put a whole turnaround, a set of expressions or a range of poses into one image in a single request, see AI character sheet.

Three details needed a second try, and each taught a rule. Asking for a “crop” of the eyes made the model paste a cut-out over grey bars; describing the shot size (“macro, 20 cm from the face”) in a 16:9 frame worked. The scar vanished until the prompt said which way to turn the head. And a full-body three-quarter view came back on a different floor until the floor was named in the studio line.

Give every reference one job

Once the character exists, references do the rest. The rule that makes this repeatable: never describe what is in the reference. The model can see it. The prompt only says what to take from picture N and where it goes. Which job a reference gets follows from the collection it came from:

Each line comes with its opposite in a role block: “take the pose only … do not take the person, face, body type, clothes, hair, background and light”. Where precision matters, the formula states a condition instead of a description — “if the camera looks down, look down just as steeply”, “if the frame is flooded with one colour, flood this one too”. A condition works on any reference; a description works on one.

Camera and light: when words are not enough

Camera failed as a sentence and only worked once we changed the order of the images. Light worked in one pass once the prompt stopped fighting it.

Camera: the reference becomes the base frame

“Shoot it with exactly the camera of picture 3” gave the window bar in the foreground but lost the steep angle and made Mara too large. What held, three runs out of three: the camera reference goes first, and the prompt says “recreate picture 1 frame-for-frame and put Mara in place of the person in it”, keeping the camera, the crop, the foreground and what the camera sees around the person — the same ground plane and depth — and only restyling the surfaces into her world. The anchors move to pictures 2–3, and the output takes the aspect ratio of the reference.

Light: stop the scene from fighting the reference

In the first try the lighting reference lost to the scene: street lamps, a “restrained colour grade”, the night city. Three changes fixed it in a single pass. The light formula states a condition instead of a description — “if picture 3 is flooded with a single colour, flood this whole frame with that same colour, even the shadows”. The world line switches the place's own lamps off. The post block drops the neutral grade and names picture 3 as the only source of light and colour. With thinking set to high, the frame comes out flooded.

How many references can you mix?

We stacked ten roles on Mara, two at a time, in one frame prompt: pose and prop, then location and camera, then haircut and makeup, then emotion and light, then palette and an image idea. Every reference had already worked on its own, so every loss below comes from the mixing.

Every step also carries the two identity anchors, so a step with eight references sends ten images. Steps 8 and 10 include the light reference, so the light was added in a second pass — open them to see both. The model accepted twelve input images. Accepting them is not the same as using them: past six roles the references crowd each other out, and the two identity anchors get lost among ten other pictures.

Build it in layers

The same ten references work when the frame is built in three passes. Pass 1 makes the frame: camera, location, pose, prop, palette, idea. Pass 2 takes that frame as picture 1 and changes only her head and face: haircut, makeup, expression. Pass 3 is the only pass that sees the light reference and adds its light to the finished frame. No pass carries more than six roles.

Layered, the frame kept the overgrown corridor, the symmetry and ceiling lights of the camera reference, the skeleton bicycle, the palette until the light pass, the choppy haircut, the smudged eyes, the red light — and Mara's face. It still lost three things: the crouch (it fights “riding the bicycle”), the idea (a glowing trail on the floor), and a strong expression, because in a wide frame the face is too small to act. Mix emotions into shots where the face is large.

Update · later canon The passes above are the original measurement on Nano Banana 2 Lite. The method now builds the frame in one pass on GPT Image 2.5 and, if a reference gets lost, cuts the number of roles instead of adding passes.

Keep the character through the story

A character in a film changes. The cameo profile stores each change as a state that inherits everything from a parent and keeps only the difference: in Act II Mara cuts her own hair short and burns her hand; the finale inherits from Act II, so the short hair and the burn stay; a flashback five years earlier inherits from the canon, so there is no scar yet. Pick a state and every new prompt compiles her look for that point in the story — the anchors, the haircut, the bandage, the emotional temperature.

Four recipes from the run. Open any of them to see the exact prompt, unedited, next to its inputs. The prompts share one structure — identity, reference roles, inventory, scene, performance, camera, in-camera integration, final rules — written in English, with no notes to the model about what it is doing. In FlashBoards the anchors, references and every take sit on one board, so you can compare runs side by side and keep the one that works.

Where it goes wrong

• The face drifts. The prompt still describes the face in words next to the anchors. Delete the face line; keep skin, marks, hair, build, wardrobe.
• A reference brings its person along. The role block is missing its “do not take …” half. Name what stays behind: the person, face, clothes, background.
• The emotion does not transfer. Something else owns the frame — busy hands, a wide shot. Make the expression the subject: hands still, face large.
• The camera flattens to eye level. Put the camera reference first as a base frame and keep what the camera sees around the person.
• The light only tints the frame. Remove other light sources and neutral grading from the scene text and state the flood as a condition.
• Big mixes fall apart. Past six roles, build in layers: frame, look, light.
• Results vary run to run. For camera and light, generate two or three takes and pick one — the same prompt does not land every time.