How close does the clip keep your face?
Close enough to recognise, not close enough to guarantee. The model carries your face, hair and outfit into the reference performance, and then the performance decides the rest — the angle, the expression, the moment your face turns away. Knowing which half is yours is what makes the result predictable.
What carries over
Face shape, features, hair, skin tone and clothing come from the upload. The camera move, timing, head turns and the expression in each beat come from the template. This is why a clip can look exactly like you and still have a performance you did not choose.
Where likeness drifts
Three places, mostly. Angles the photo did not show: a single front-facing shot has no side of the head to invent, so a profile beat in the template gets a guess. Small faces: detail that is not in the photo cannot be in the clip. Accessories that hide features — sunglasses, a cap pulled low, a hand — remove the eyes the model relies on.
Hands and instruments are the other common soft spot; they come from the template, so they stay plausible rather than exact.
Why the clip always looks like the template
Reference-to-video keeps the source performance. You are not describing a new scene, you are swapping the lead in an existing one, and that is the point: the band, the room and the camera stay consistent from render to render. The price is that the clip will always look like the template you picked, not like a scene you imagined.
Getting the sharpest result
Start with a sharp, evenly lit, front-facing portrait. Choose 2K when the file will be edited or shown large; 768P is fine for a phone screen. Keep the head in frame with a little room above it, and avoid filters that smooth skin — texture is part of what makes a face recognisable.
The honest limit
AI output varies between runs. The example clips on this site show what a template does with a good portrait; they are not a promise that your render will match frame for frame. A render that finishes is a finished file, so it is not refunded on the grounds of likeness — failed renders are, and those return automatically.