One 3-second span was cut out, sent to seedance-2.0-mini/video-edit with our packshot
and a still from the span itself, and spliced back. $0.26, 97.6 s of render. The audio never passed through the
model.
| Checked | Result |
|---|---|
| the product in the reveal, 6.0 / 6.5 / 7.0 s | the teal competitor bottle reading MONEY is now our white bottle with the blue trigger head and the Instant Slip+Shine label, in the same hand, same pose, same framing |
| the joins at 5.0 s and 8.0 s | hair, room, light and clothing continuous across both |
| length and frame size | 627 frames at 496x864, frame-identical to the source, audio stream-copied |
| everything the reviewer still flags | each one also appears in a review of the untouched original, confirmed frame by frame |
All three runs are in qa/ on this page. Runs A and B are the same bytes, the
unrepaired original, reviewed hours apart.
| Run | File | Fatal | Major | Its fatal was |
|---|---|---|---|---|
| A | before | 1 | 0 | the teal competitor bottle at 6.0 s |
| B | before, identical bytes | 1 | 1 | the hair going dry to wet at 5.0 s |
| C | after | 1 | 4 | the blue bottle in the spraying shot at 8.0 s |
Two passes over one unchanged file named two different single fatals, so the count is not a score. Every defect run C raises against the repaired cut also appears in a review of the unrepaired cut, and each one was confirmed by pulling the same timestamps out of both videos: the dry-to-wet hair is the donor's own hard cut, the stray hand at 4 s is in the original, the bottle in the spraying shot is in the original, and the bracelets sit in identical frames either side of the join. The teal competitor bottle is the one thing that changed, and it is gone.
before
after
Two things the review could not see, both caught by measuring. The endpoint returned 4.042 s for a 3.000 s input, with the extra second appended at the end, which pushed the tail late against the audio and cut 1.2 s off the end of the ad until the splice started trimming. And the splice was downscaling the whole video to 854 high when the source is 864. Both are fixed and the length is now asserted against the source on every run.