Why does my AI character's face change between videos?
Because the model is not remembering your character: every generation draws a new face that fits whatever it was given. If it only got words, any face matching those words is a valid answer, so the face changes; attach the same reference picture every time and most of the drift stops.
The five causes we see most
- No reference image. A text description like "long blonde hair, grey eyes" fits millions of faces. The model picks a different one each run. In our experience it is the most common cause.
- The description changed. Small edits between videos ("blonde" one day, "platinum" the next, a new outfit) read as a redesign. Keep one master description and paste it unchanged.
- Frames made from frames. If each new still is generated from the previous video's last frame instead of the master picture, small errors add up over a series. Go back to the master each time.
- Other people in the shot. In group shots the model averages faces, so friends start to look alike and skin tones shift. That one has its own fixes: see group shots.
- Style words fighting the reference. A video prompt that also describes a new art style, lighting look or age pulls the face away from the picture. Put the look in the image you start from and keep the video prompt about motion.
Does a better prompt fix it?
Partly. A precise, unchanging description narrows the range of faces the model can pick, and restating hair, skin and expression helps secondary characters most. But in our renders no amount of prompt detail held a face across separate generations the way an attached picture did. Prompts set the outfit and behaviour; the reference image sets the face.
Do reference images fix it?
Mostly, if you use them the right way. Start each video from a still frame painted off the master picture, then animate that frame. For our explainer Penny we generated 31 separate clips this way and she reads as one character throughout. Two caveats from our own work:
- Feeding a previous video back in as a video reference, to continue a scene, made the model restart the scene in our test, with a visible jump. Continuing from the last frame as an image worked.
- If the model blocks a face because it resembles a famous actor, the only fix that worked for us was redesigning the face, and then every earlier video had to be re-rendered to match. Design an original face from the start.
A quick checklist before you render
- Same reference picture as last time, attached.
- Same description, pasted, not retyped.
- Outfit written out in full.
- Each extra character described and referenced separately.
- Video prompt says "No change to her face" in the avoid list; we keep that line in every prompt.
The longer version, with what broke along the way, is our character consistency guide.
Where Cast fits
- You want the reference handled for you: a Cast character's chosen look is attached automatically to every video, along with its fixed description.
- You want several recurring characters: each one's picture is sent as its own numbered reference, in the same order every time.
Where it does not:
- Fixing drift on a character from your own photo or drawing: Cast cannot take an uploaded picture; it draws characters from a text description.
- Re-rendering a long series shot by shot: Cast makes single 5 to 15 second clips.
People also ask
- What causes AI face drift?
- Mostly generating without a reference image, so each run samples a new face that fits the words. Edited descriptions, chained frames and group shots add more drift.
- Why does the face change halfway through one video?
- Often a cut the model invented, which restarts the face. In our prompts, "one uninterrupted take", "the only viewpoint" and "no cuts" together stopped invented cuts.
- Will a higher-quality model stop the drift?
- It helps detail but not identity. A cheaper model with a reference image keeps a face better than an expensive one with text only, in our experience.
- How do I fix drift in group shots?
- Restate each character's hair, skin and expression, pass each one's own picture, and give each a signature move or prop. Details are in group shots.
Related answers
- How do I keep an AI character consistent across videos?
- How do I keep two or three AI characters consistent in one video?
- What is the best AI video generator for consistent characters?
- Why did Seedance block my video?
- Why are AI videos only 8 to 15 seconds, and how do I make longer ones?
Part of the Cast answers library (Character consistency).
Updated