Free guide ·
Why AI characters change between shots and how to stop it: a character sheet, a fixed identity block, reference images, short shots, and when to use code.
By John Shemron, graphic designer and video editor at PromptVault
Keeping the same character in AI animation means the face, body, clothes and colours stay identical from one shot to the next. With AI video tools you get there by fixing the character's identity before you animate: design it once, describe it the same way every time, give the tool reference images, and keep each shot short. If the character has to stay exactly the same across a whole series, build it from code instead, where it cannot drift at all.
A video model doesn't remember your character. Every generation starts again from your words and images, and it fills in everything you didn't pin down. Three things make that worse:
So consistency isn't one setting you switch on. It's a set of habits that remove guesswork at each step.
Before any video, make a character sheet: one still image that shows the character from the front, from three-quarters and from the side, on a plain background, with a neutral expression.
Design for recognition, not detail:
A simple design that holds up in every shot beats a rich one that changes in every shot.
Write one paragraph that describes only the character, and paste it word for word into every prompt. Keep the scene in a separate part of the prompt.
| Part of the prompt | What goes in it | Changes per shot? |
|---|---|---|
| Identity block | Face, hair, body shape, outfit, colours, signature feature | Never |
| Scene | Location, lighting, what the character does, camera | Every shot |
| Locks | "Same face, hair and outfit as the reference. No other characters." | Never |
An example identity block:
Maya: a woman in her late twenties, round face, short black bob with a straight fringe, thick black glasses, mustard-yellow raincoat over a white T-shirt, dark blue jeans, white trainers. Slim build, average height.
Then a scene part for one shot:
Maya stands at a bus stop in light rain at dusk, checks her phone and looks up the street. Medium shot at eye level, slow push-in. Same face, hair and outfit as the reference image. No other people.
Save the identity block in a note. Don't retype it from memory: one changed word ("dark" instead of "black") is enough to get a different person.
Most current video tools let you attach images of the character, so the model works from a picture as well as your words. The feature names differ:
| Tool | Feature | What it takes (checked November 2026) |
|---|---|---|
| Google Flow (Veo 3.1) | Ingredients to Video | Up to three reference images per prompt; Google recommends a plain background for character references |
| Runway | References | Works from a single reference image, or up to three in one generation |
| Kling | Elements | A character element built from two to four images, a front view first |
These features change often, so check each tool's own help page before you plan a project around them.
Tips that apply to all three:
The surest way to keep a character in a clip is to start the clip from an image you've already checked.
Small actions hold identity better than big ones. If a shot needs a full turn or a run, split it into two clips with a cut in between.
Watch each clip three times and compare it with the character sheet:
Only move on once the clip passes. A drifted clip early in a story forces you to regenerate everything after it.
| Symptom | Likely cause | Fix |
|---|---|---|
| A different face in every clip | No reference image, or a new one each time | Use the same reference files for every shot |
| The face changes halfway through a clip | The clip is too long or the action too big | Shorter clips, smaller actions, cut on the turn |
| The outfit changes colour | The identity block was retyped or shortened | Paste the saved block word for word |
| A second person appears | Nothing told the model there is only one | Add "No other people" to the locks |
| The character looks right but older or younger | Age left out of the identity block | State the age range in the block |
Everything above makes a video model drift less. It doesn't make it stop. For a one-off ad or a trailer, that is usually fine. For a recurring series, where the same two characters appear in every episode, even small drift adds up.
The alternative is to build the character from code. An AI coding agent such as Claude Code writes a 3D scene in which each character is defined as shapes, sizes and colours, and your computer renders every frame. Nothing is generated per shot, so there is nothing to drift: the character in episode ten is the same object as in episode one. The trade-off is the look. Code draws stylised characters, such as low-poly or flat cartoon styles, not photoreal people.
The clip above was made this way. The Claude & Pixel Guide teaches this exact setup for two-character dialogue shorts: you paste one prompt into Claude Code, approve the story and voices, and get a finished vertical video, with the same cast in every episode. It needs a Mac, and it is an independent course, not made by Anthropic.
For the bigger picture of the three ways to animate from a prompt, see the prompt animation guide. For the format these characters are made for, see how to make animated dialogue videos. To see what Claude can and can't animate, read can Claude make animations?.
Because each clip is generated from scratch. Without a reference image and an identical description, the model fills in the gaps differently every time. Use the same reference images and paste the same identity block into every prompt.
Start with one clear front view on a plain background. Add a three-quarter and a side view if your shots show the character from different angles. Check your tool's limit: in November 2026, Google Flow and Runway each took up to three, and Kling's character elements took two to four.
Partly. Free tiers of most video tools are limited and often watermarked, and consistency needs extra attempts, which cost credits. A character built from code costs nothing per render, but the coding agent usually needs a paid plan.
Only with their permission. Using someone's likeness without consent can break the tool's terms and the law where you live, and advertising platforms may require you to disclose that a person is AI-generated.
The Claude & Pixel Guide: Animated dialogue videos for Reels, Shorts and TikTok, made in code from one prompt. 15 text lessons in three tracks.