Why characters drift
Text-to-video has no memory. Even image-to-video re-renders the source frame with new motion, and small changes in light or pose can cascade into a different face. The more the prompt changes between shots, the more the model has to invent.
Reference images reduce the search space, but they do not eliminate it. You still need a stable description, a clean set of angles, and a shot plan that does not ask the model to solve identity and action at the same time.
The core rule: Keep identity constant and change one production variable at a time — camera, action, or light. If all three move, you will not know which one broke the character.


