Every clip starts from a picked image (a plate) that the owner has already seen. The video model moves her; it never invents her.
Four jobs, four engines
| Job | Engine | Input | Clips |
|---|---|---|---|
| Sing to the lens | Wan 2.7, 720p, on Higgsfield | The plate and the vocal stem of its lines | Send It 26, Gold Rush 2049 7, La última moneda 7 |
| Act the story | Kling 3.0 pro, sound off | The plate as the first frame and a one-moment script | La última moneda 25 |
| Move without singing | Seedance 2.5 | The plate | Send It 3 |
| A still that moves | A 2.5D camera move in code | The plate | Gold Rush 2049 19, La última moneda 1 |
Higgsfield charged 1.5 credits per second for both Wan 2.7 and Kling 3.0.
A sung take
- The plate is already singing. The prompt put her facing the lens, mouth slightly open and clearly visible, nothing in front of her face.
- The audio is the vocal stem, cut to the window of the lines that shot sings, not the full mix.
- The sync is measured after the edit, per sung shot, at the offset the edit really plays, with SyncNet. In Gold Rush 2049 all seven sung shots landed within one frame. One shot's offset was moved from −80 ms to 0 by measuring the whole clip instead of the used part.
- A take that does not sync is cut, not forced. In an early cut of Send It, two takes followed the vocal only 19% to 53% of the time. Those shots cut away to a shot where she does not sing. The final version re-shot every sung shot on Wan 2.7.
An acting clip
- The first frame is the picked plate, so her face is the one the owner approved.
- The prompt is a script of one moment: who moves, what they do and how the camera moves. When a picked plate did not match the board's action, the motion was adapted to the plate: "she turns and walks out".
- Sound off. The song plays over the edit.
Retakes fix the prompt
The close-up where she stares at the chart (shot 8 of La última moneda) was retaken with a corrected prompt, without the chart reflected in her eyes, for 6 credits. The plate was not edited: a fix goes into the prompt, so the next take cannot bring the fault back.
The scan
Before the owner sees v1, every take the edit plays is scanned: six frames per take on a sheet, checked for safety and for canon (her face, her age, her clothes, stray text). A block is cleared only by remaking that take or cutting it from the edit, then scanning again.