Guides

How to keep characters consistent in AI video

The same face, the same clothes, the same person, in every shot. Why AI characters drift between clips, and a practical system that keeps them on model.

Updated 28 September 2026

Character reference sheets and looks in Poiema's lorebook

Every AI filmmaker hits the same wall. The first shot is perfect. The second shot has a different person in it. By the fifth, your hero has changed age, hairstyle and jacket twice. This guide explains why it happens and walks through a system that keeps characters consistent across a whole film.

Why characters drift

AI image and video models do not remember. Every generation starts again from what you give it. If all you give it is text, the model has to invent the character from the description every time, and a description is never specific enough:

"A young man with curly hair, a denim jacket and a friendly face."

That sentence describes millions of people. The model picks one per generation, so you get a different one per shot.

Small details drift first: eye colour, jewellery, the pattern on a shirt. Then the big ones go: face shape, age, build. The more shots you make, the further it wanders.

The fix is to stop describing characters and start showing them.

Step 1: Make a clean reference sheet

A reference sheet is an image the model can look at every time the character appears. A good one has:

In Poiema, each character in your project's lorebook has a reference sheet you can generate, refine or upload. Read how it works on consistent AI characters.

Step 2: Turn outfits into named looks

The most common consistency break is clothing. You write "she changes into a black dress" in one shot, and suddenly her face changes too, because the model is reinventing her.

Treat each outfit as a look with its own reference image:

Look When it's used
Office suit Scenes 1 to 3
Rain-soaked suit Scene 4, after the storm
Black dress Scenes 5 and 6

Then pick the look for each shot. The character's identity stays the same; only what the story changes, changes.

Step 3: Give locations and props references too

Consistency isn't only about faces. If the café changes layout between shots, or the red car turns maroon, the film breaks just as badly. Create references for:

Step 4: Keep one visual style

Mixing styles makes consistency impossible. A character rendered photoreal in one shot and painterly in the next reads as two people. Choose one style for the whole film and keep it. In Poiema the style is set per project and applied to every shot.

Step 5: Put every character's reference into every shot they're in

This sounds obvious, but it's where manual workflows fail: you forget to attach the reference on shot 23. The reliable approach is a system that attaches references automatically based on who is in the shot.

Poiema does this from your shot list. When a shot includes a character, their current look's reference goes into the generation, along with the location and props in the scene.

Step 6: Plan shots that help the model

Some shots are harder to keep consistent than others:

Step 7: Choose models that use references well

Models that accept reference images hold identity much better than text-only generation. When you choose a model for a shot, prefer one that supports reference images or a first frame. Because your references travel with the character, you can switch models between shots without losing them.

Step 8: Review like a continuity supervisor

Before you move on, check each shot against the reference:

Regenerate only the shots that break. It is faster than fixing them in the edit.

The system in one line

Show, don't tell: one clean reference per character, one look per outfit, references for places and props, one style, and references attached to every shot automatically.

Ready to try it? The AI film studio keeps your whole cast in one project, and the guide to making an AI short film shows where consistency fits in the full process.

Questions

Why does my AI character look different in every clip?

Because each generation starts from its own prompt. Words like "a woman with short dark hair" fit thousands of faces, so the model picks a new one each time. Visual references fix this; descriptions alone cannot.

What makes a good character reference image?

A clear, evenly lit image of one character, showing the full face and outfit, on a plain background, in the same style as your film.

How do I handle outfit changes?

Make each outfit a separate named look with its own reference image, then choose the look for each shot rather than describing the new clothes in the prompt.

Does the video model matter for consistency?

Yes. Models that accept reference images hold identity far better than text-only generation. Using the same references across models keeps characters stable even when you switch.

The next scene is yours

Design your cast in Poiema.

Design your cast