Video 07, head-swapped — before and after

The man is the right man now in the frontal shots, and the QA count did not improve. One call replaced his head across all 24 seconds and left his hands, the bottle, the woman and every movement where they were. In profile, around 20 seconds, the identity holds less well, and the reviewer marks a face change at that cut which it never marked on the old version. Before: 1 fatal, 2 major. After: 1 fatal, 3 major.

The two versions

before
1 fatal · 2 major · 1 minor
after the head swap — $0.12
1 fatal · 3 major · 1 minor

The same frame, four ways

At 23.0 seconds, the moment the face was worst.

the cast reference, the shipped frame, the still swap and the head swap

The face through the finished video

Cast reference first, then the finished clip at 1, 5, 12, 20 and 23.5 seconds. The 20-second tile is the profile shot where the swap is weakest.

the face across the finished video

Every finding, side by side

what the reviewer sawbeforeafter
The label on the closing hero shot does not read as our label fatal at 23sfatal at 22.9s
The bottle turns into a solid light blue unbranded one while spraying major at 5smajor at 4.4s
The bottle vanishes from his hand across a cut major at 15s—
A table lamp appears on the left wall that was absent earlier —major at 9.1s
The man's jaw and head proportions change at this cut —major at 19.7s
Faint blurred overlay text over the woman minor at 2sminor at 1.7s

Read honestly: the reviewer grades whether one man stays one man inside the video, and it never had a quarrel with the old version, because the old version was one wrong man from end to end. The head swap installs the man we asked for and does it unevenly, which is why a new identity note appears. The bottle findings are the separate product defect measured across both Seedance tiers earlier today, and the head swap was never going to touch them.

What the endpoint did to the file, and what was done about it

The head-swap endpoint returns neither the frame rate, the frame size nor the length it is given. Twenty-four frames a second at 496 by 864 and 24.10 seconds went in; thirty frames a second at 464 by 832 and 23.77 seconds came back. The missing third of a second falls exactly where the closing line ends, so the input was padded with 0.8 seconds of frozen tail and run again, and the finishing step forces the house frame rate and size back on, trims to 24.097 seconds, muxes the audio and refuses to write the file at all if the length is wrong.

Checked after the fix: 578 frames at 24/1, 496 by 864, 24.097 seconds of video against 24.097 seconds of AAC audio. The picture is still moving at 23.9 seconds, so the closing line plays over live footage.