This single still is the entire visual input. Everything the model does is movement on top of it, which is why her identity holds across the whole eleven seconds.
The first eleven seconds of the QuietLab UK port, made the way the whole ad would be made, so the method can be judged before five more minutes of it are rendered.
The verdict beside the clip reads 0 fatal, 1 major, and the major one is real and fixable.
| At | Severity | What |
|---|---|---|
| 5.14s | major | The voice says "matched" where the script says "match". The fix is to redub that one line in the same cloned voice, not to regenerate the picture. |
| 0s | cosmetic | Lip sync measured at zero lag, which is correct; it is reported rather than judged. |
| 0s | not judged | The beat carries one visible change in ten seconds, where short-form leans to one every one or two seconds. This is a note for the eye, never a defect, and the finished cut will carry far more. |
The full verdict is in qa/Q02-open-presenter.qa.md beside this page.
Eleven seconds took 313 seconds of GPU on a 5090 at $0.92 an hour, which is $0.0071 a second, so the presenter stretches of the full five-minute ad come to roughly $2.30. The box took 262 seconds to set itself up, at $0.14.
The model writes 464 by 832, which is below the 500 pixel minimum Instagram enforces, so every take needs upscaling to 1080 by 1920 before it could run anywhere. The running captions are not burned yet. And the beats between the presenter stretches need their picture, which comes from the 167 real Doe creator clips rather than from anything generated.