Video models have no memory, so every clip invents a new face. You need the same character from every angle plus expressions — and the only way to get those to match is to generate them in one go.
You generate a shot. The character looks right. You generate the next shot with the same prompt, and a different person walks into frame. Same hair colour, same jacket, different face.
This is the most common complaint about AI video, and the cause is simple: video models have no memory. Every generation starts from nothing. A prompt describes a person; it does not bind the model to one. "A woman in her thirties with short dark hair" matches millions of faces, and the model picks a new one every time.
Longer prompts do not fix this. More adjectives only narrow the pool. The fix is to stop describing your character and start showing it — the same character, from every angle your shots will need.
Think about what your shots actually require.
A wide shot shows the character from the front. In the next shot they turn away, so the model needs the back. They walk past camera, so it needs the left and the right side. Then you cut to a close-up where they are angry — and now the model needs to know what that face looks like angry, not a guess.
So one photo is never enough. You need a set: front, three-quarter, both profiles, back, and a close-up per emotion you plan to use.
Every angle you do not supply is an angle the model invents. And it invents a different one each time. That is exactly why faces morph the moment a head turns.
The obvious approach is to make them separately. Generate the front. Then prompt for the back. Then the left side. Then an angry close-up.
It falls apart for the same reason the video did: each of those generations is independent too. Your back view comes back with a slightly different jawline. Your angry close-up has different hair. You end up with six images of six slightly different people — and you are about to feed all six into a video model as "the same character".
You can brute-force it. Twenty candidates per angle, throw away the ones that drift, nudge the prompt each time. It costs an afternoon per character and the result is still uneven.
One generation means one decision about what this character looks like, applied to every panel at once. The panels cannot disagree with each other, because they were made together.
That is what the Character Sheet app does. Upload up to three photos of your character — a photo, a drawing, a mascot, a 3D render — and it returns one sheet with that same character repeated across every panel.
Three settings matter for video work:
What to show. Turnaround gives you one neutral standing pose rotated through front, three-quarter front, side, three-quarter back and back. That is your angle set. Expressions gives you head-and-shoulders close-ups, each a different emotion. Mix puts turnaround, poses and expressions on one page.
Run it twice — once on Turnaround, once on Expressions. Two focused sheets beat one crowded page, and you get bigger panels.
How many panels. Nine, not six. More angles covered per generation.
Background. Neutral grey. You are making references, not artwork. A plain background keeps the model on the character instead of the scene.
Then crop the panels. The sheet is not the deliverable — the panels are. Cut each one out and name it: front.png, profile-left.png, back.png, angry.png. A sheet of nine small figures handed to a video model gives it nine small faces to average. A single cropped panel gives it one face to copy.
If you need angles beyond what a sheet gives, Angles turns one image into nine camera angles of the same subject, with style and background held steady.
Everything you generate lands in your Library, and that is where a character kit belongs. Put the sheet and the crops in their own folder, named after the character. Now the kit is not something living in your downloads folder that you re-upload for every shot — it is in the app, next to the tool that needs it.
From the library, every asset has a transfer button that sends it straight into another tool and fills that tool's upload slot for you. So the loop becomes: open Create Video, send in the crop that matches the shot, write the prompt, generate. No re-uploading, no hunting for the right file, and no risk of grabbing last week's slightly different version.
That last point is the real reason to keep it in one place. Character drift usually does not start in the video model. It starts with someone feeding in a different reference than they did yesterday.
Two rules once you are there:
Generate from an image, never from text. Text-to-video rolls the dice on the face. Any shot your character appears in starts from a reference.
Hold everything else still. Same model, same aspect ratio, same crops, same lighting direction, same described outfit, across the whole project. When a shot comes out wrong, change one thing and regenerate. Changing three at once tells you nothing about which one mattered.
And let the prompt describe only what changes — setting, action, camera, light. Never the face. You already supplied the face; describing it again just gives the model something to argue with.
If your model takes several references, tell it what each one is for:
@Image1 is the character's face and hair. @Image2 is the outfit. @Image3 is the location. Keep identity, hairstyle and outfit exactly as the references.
More references is not automatically better. Two or three with clear roles beat six that compete. If two references disagree about the hair, the model averages them — and an average face is a new face.
Reference limits differ a lot, and they decide which model fits which shot. These are the limits in Opaxes, checked 3 October 2026.
| Model | Reference images | Also accepts | Max length | |---|---|---|---| | Seedance 2.5 | 30 | 10 videos, 10 audio, first/last frame | 30s | | Seedance 2.0 | 9 | 3 videos, 3 audio, first/last frame | 15s | | MiniMax H3 | 9 | 3 videos, 3 audio, first/last frame | 15s | | Veo 3.1 Fast / Lite | 3 | first/last frame | 8s | | Kling 3 | 3 elements | first/last frame | 15s | | Seedance 1.5 Pro | 2 | — | 12s | | Wan 2.6 | 1 | — | 15s |
If a shot needs the character plus a specific outfit plus a location, use a model that takes several references — Seedance or MiniMax. Veo takes three, enough for a character and one other thing. Wan takes one, so it gets your front crop and nothing else.
Length matters as much as the limit. A 30-second generation keeps its own continuity from start to finish. Four separate 8-second generations each roll the dice again. Where the story allows it, generate long and cut it down afterwards.
One more setting: most models let you pin the first frame of a shot, the last, or both. Pinning the first is almost always a win — the shot starts from a face you already approved. Pinning both is where it goes wrong: if the two frames are visually far apart, the model morphs the character on the way between them.
It will, sometimes. Work through it in this order:
1. Swap the reference for the crop that matches the shot's angle. A profile shot wants the profile panel. 2. Simplify the prompt. Remove every description of the face. 3. Calm the camera. Fast motion and quick turns give the model room to re-invent. 4. Shorten the shot. Drift grows with length. 5. Regenerate. These models are not deterministic, and the second attempt is often simply right.
Never repair drift by feeding a drifted output back in as the new reference. That is how a character slowly becomes someone else over twenty shots. Always go back to the folder.
The specific failures, and what each one means:
| What happens | Why | Fix | |---|---|---| | Face morphs when the head turns | Your references only show the front | Use the turnaround crops, matched per shot | | Clothing changes between clips | The outfit lives only in the prompt text | Add a separate outfit reference and repeat the outfit sentence every time | | A different person each scene | Text-to-video, no reference | Anchor every clip to the same crops | | Close-up emotion looks like someone else | No expression reference for that emotion | Generate an Expressions sheet and crop per emotion | | Character looks older or younger | Lighting and grade changed | Hold the light direction constant, grade at the end | | Character warps mid-shot | First and last frame too far apart | Use the first frame only | | Two characters blend into each other | Too many people, competing references | Test each character alone first, then combine |
None of this makes a character permanent. There is no saved identity inside these models — every clip re-anchors from your images. That is exactly why the kit matters more than the prompt, and why keeping it in one folder matters more than anything else you do.
Build the sheet once. Crop it once. Reuse it for the whole production.