Why your AI character changes between shots
Faces drift because every generation invents the character again. Here is what causes it and the workflow that stops it.
You wrote the same description twice and got two different people. That is not a bug in the prompt. It is what happens when a model is asked to invent a face it has never seen.
The face is being invented every time
A text prompt describes a person in adjectives. The model turns those adjectives into a face using its own idea of what those words mean. Ask twice and it interprets twice, so the cheekbones move.
No amount of extra detail fixes this. Longer descriptions narrow the range slightly, then the model still picks a different face inside that range.
Three things that cause drift
- Describing the face in words instead of showing it.
- Generating each shot from a fresh prompt with no reference image.
- Letting the video model re-cast the character when it moves from the first frame to the last.
Lock the frame, not the prompt
The reliable way to keep a face is to stop describing it and start reusing an image of it.
Build the actor once: a headshot, a couple of angles, and a note on wardrobe. Every shot after that is generated with that image as a reference, so the model is matching a picture rather than guessing at a paragraph.
For video, the same idea applies one level up. Generate the keyframes first, approve the ones where the face is right, then animate from those frames. The video model is now carrying the face forward instead of inventing it.
A workflow that holds
- Create the actor and keep the reference images in the project.
- Write the episode, then generate the frames for every shot.
- Fix or regenerate the frames that look wrong. This is cheap. Video is not.
- Render only from approved frames, so the face is already decided before any motion exists.
Wardrobe and extras are the next leak
Once the face holds, the next things to drift are clothing, hair length and background characters. Locking a look and a small set of locations solves most of it.
Extras are the hardest. Keep crowds out of close shots, and keep the number of recurring background characters low enough that you can reuse the same reference for each one.
What still needs a human eye
Hands, teeth, and fast motion still break. Look at every frame before you spend on video, because a bad frame becomes twenty bad seconds.
The workflow is not there to remove judgement. It is there to make your judgement cheap to apply.