The first two seconds look great. By the end of the clip, the person has a different jaw, different eyes, or a hairstyle that never appeared in the uploaded photo. This is usually called identity drift.
Unlike a single-image edit, a video model must carry one identity through a sequence of changing frames. Head turns, shadows, cuts, motion blur, and additional faces all make that memory harder to maintain.
Find the exact moment of the drift
Don't judge the video only by its thumbnail. Scrub from beginning to end and find the first frame where the appearance changes. Did it happen after a cut? When the face became small? During a side profile? When another person entered the shot?
The trigger tells you what to fix. Drift after a cut often benefits from explicit continuity instructions. Drift during a profile turn points toward incomplete reference geometry. Drift when another face appears may be a target-selection failure rather than an identity issue.
Reduce the number of competing instructions
A prompt that asks for a face replacement, dramatic lighting, wardrobe changes, a camera zoom, and entirely new scenery is difficult to obey consistently. When the source video already contains the action you want, ask the model to preserve it.
Try: “Keep the uploaded subject's facial identity consistent through all shots, including head turns and re-entries. Retain the original scene timing, clothing and camera motion. Change only the intended person's identity.”
This does not guarantee perfect continuity, but it establishes a clearer priority.
Compare frames side by side
Capture three screenshots: opening, midpoint and closing. Resize them to similar face sizes before comparing. Inconsistent lighting can fool the eye, so focus first on stable features such as brow distance, nose shape, ear placement and jaw contour.
If you are using a one-image reference, try a more informative portrait rather than simply a larger one. If the workflow permits multiple images of the same person, include a complementary angle. A second person’s photo is not a substitute for another angle of the first.
The goal is not flawless skin in every frame. It is a person who remains recognizably themselves throughout the video. That is the standard worth optimizing for.
