Character consistency in generative video: techniques that work

Consistency is the property that the same character is recognisably the same across scenes: face, proportions, costume, palette. Generative models have no memory between calls; every request starts from nothing. Consistency therefore has to be engineered into the inputs, and the techniques that work are all variations of one idea: decide what the character looks like once, in a form the model can copy, and give it that form every time.

Why it is hard

A text prompt describes a distribution, not an individual. "A fox in a blue jacket" produces a different fox every time, because thousands of foxes match the description. Even a very long description leaves room: ear shape, snout length, eye colour, the cut of the jacket. Video models make it worse, because they also have to keep the character stable across frames, and they trade fidelity to the reference for motion quality. The result, without engineering, is a film where the viewer needs subtitles to know who is who.

Technique 1: reference sheets

Draw each character once: full body, neutral pose, plain background, in the film's style. This is the visual anchor. A sheet works because an image pins down everything a description leaves open. Store it with the character so it can be reused across films, and keep the prompt that produced it so it can be regenerated in another style.

Practical rules: one character per sheet, no props that will not appear in every scene, no dramatic lighting. The sheet is a specification, not a poster. A dramatic sheet drags its lighting into every scene that references it.

Technique 2: a written look

Alongside the image, keep a 30–60 word description: species or age, face, hair, costume, colours, one distinguishing mark. Include it in every scene prompt. Words steer the model even when the visual reference is weak or absent, and they fix the things an image cannot, such as the character's name or role. The written look is also what a language model reads when it writes scene prompts, so it should be written for a model: concrete nouns and colours, no adjectives like "beautiful".

Technique 3: keyframe first

Do not ask the video model to place the character. Ask the image model to compose the scene as a still, with the reference sheets inline and an instruction to keep every design faithful. Then animate the still. Image models are markedly better than video models at honouring references, because they solve one frame rather than a sequence, and they are ten to fifty times cheaper. A failed keyframe costs cents and is regenerated in seconds; a failed clip costs dollars and minutes.

The keyframe route has a second benefit: it separates composition from motion. If a scene looks wrong, the problem is visible in the still before any motion is added, and it can be fixed by editing the prompt for that one frame.

Technique 4: one entity per entry

Language models like to economise. Given a story with a fox and a bear, a director model will sometimes produce one cast entry called "the animals". Forbid this in the contract: each cast entry describes exactly one being. A merged sheet cannot anchor two characters, and every later stage inherits the merge.

Technique 5: continuity frames

When a scene continues the previous one in the same location, start the new clip from the last frame of the previous clip rather than from a fresh keyframe. The character, lighting and set carry over automatically. Use this only for true continuations; a new location needs a new keyframe or the old set leaks into it.

What did not work

Multi-image reference endpoints on video models: acceptable for human faces, weak for animals and objects, and liable to be removed by the provider without notice. Several such endpoints have disappeared within months of launch. Text-only prompts with a name: the model has never seen that name and treats it as noise. Very long prompts: past a few hundred words, models weight the beginning and ignore the rest, so the look should come first, not last. Seeds: a fixed seed reproduces a single image, not a character across different compositions.

Measuring consistency

Pick three scenes with the same character and compare face, costume and palette side by side. Score each dimension 0–2: 2 if identical at a glance, 1 if recognisable with effort, 0 if it could be a different character. A character scores out of 6; a film averages its characters. Below 4, regenerate the sheets before regenerating any scene: the fault is almost always in the anchor, not in the scenes that reference it.

A cheaper proxy is the subtitle test: watch the film with the sound off and ask whether you can tell who is speaking. If you need the name on screen, consistency has failed.

Frequently asked

Does a photo of a real person work as a reference sheet? Only for photorealistic styles, and only with the person's consent. For drawn styles, generate the sheet in that style; a photo drags realism into every scene.

How many reference images per character? One good full-body sheet is enough for most films. A second sheet with a close-up of the face helps for scenes that are mostly faces.

Can two characters share one scene reliably? Yes, if both sheets are passed to the keyframe stage and the prompt names them by position ("the fox on the left, the bear on the right"). Three or more in one frame is where current models start to blend features.