Why it happens
- No approved character master image, so each shot re-invents the character from text.
- The appearance description is worded differently in every prompt, so the adjectives drift.
- The shot includes a turn, an occlusion or a big camera move, and the model rebuilds the face mid-motion.
- The clip runs too long and identity drifts further from the first frame toward the end.
How to fix it
Approve one front-facing half-body master image and use it as the reference for every keyframe.
Lock face shape, features and hairstyle into a fixed text block and paste it verbatim into every shot.
Use image-to-video instead of text-to-video, and check the face in the first frame before generating.
End the video prompt with explicit holds: keep face shape, features and hairstyle unchanged.
Split shots with turns or occlusions into two shorter shots, or cut them to 3–5 seconds.
Before and after
Fails
A pretty young girl stands in a convenience store, looking nervous.
Works
25-year-old Chinese woman, narrow oval face, almond eyes, shoulder-length straight black hair, grey-blue store uniform, thin orange tie, silver name badge; standing behind the counter, looking nervous. Keep face shape, hair and uniform unchanged.
Prevent it next time：Approve a master image and a fixed description before any shot; every shot references them and never rewrites them.
Worked examples
Keep one character's face stable across 6 shots
A character's face and features change as shots move between wide, close-up and profile.
→Emotional close-ups: expression changes, face stays
In close-ups the expression either freezes or changes so much the face changes with it.
→Crowd scenes: a clear lead, a crowd that stays put
In crowd scenes the lead's face drifts, extras move chaotically, and duplicate or extra faces pop up.
→Keep period costume and hairstyles consistent
A period character's topknot, hairpin, collar and embroidery differ from shot to shot.
→