Top left is the man the job asked for. Top right is what we shipped. The bottom row is that same frame through each swap tier.
Fraction of pixels in the face box moving more than 12 grey levels against the untouched frame.
| frame | $0.010 swap | $0.025 pro swap |
|---|---|---|
| 22.5s | 0.008 | 0.070 |
| 23.0s | 0.025 | 0.149 |
| 23.5s | 0.018 | 0.288 |
One frame from each of the six, with the cast reference first. None of them is him.
This is the one file every one of the six jobs named as its identity reference, the male stylist in the cast folder. It is a 720 by 1280 screengrab of somebody else's ad, with their caption burned across it and their bottles in his hands.
The prep-kit check has been printing MISSING cast for the male presenter on every
generation in this session. A proper three-view sheet on white, plus the woman, the brush and the product cut-outs,
is quoted at $0.10 for the set.
A video-edit model will not install an identity. It keeps what its input already shows, and when the job asks for a different person it invents one of its own. Seedance did change the man here — the donor's presenter is older and grey-haired — it simply invented a young man instead of using ours. The face-swap pass that ran afterwards used the same weak reference and the cheap tier, so it moved almost nothing.
Swapping the face on a single still turns the identity from something a model has to be persuaded into, into a fixed input it starts from, and a still can be inspected and redone for a cent before any video is generated. That holds here, with the pro tier and a real head-and-shoulders reference. What is still unmeasured is whether the identity survives the motion once that approved still is animated.