Why it happens
- Video was generated before checking keyframes, so face, hand and text errors surface only in the paid stage.
- Every test runs at maximum resolution and maximum duration.
- Character, set and camera are changed together, so you can't tell what worked and end up guessing.
- Shots that already passed get regenerated along with the rest.
How to fix it
Budget per shot: duration × current per-second rate × expected attempts, plus voice and post costs.
Solve character, hands, text, composition and props on stills before moving to video.
Test motion at short duration or low resolution, and render full resolution only for the pick.
Change one variable per attempt and log model, seed, prompt and asset versions.
Redo only failed shots, and reserve extra attempts for high-risk shots.
Prevent it next time：Draw up a per-shot budget before starting, and spend nothing on video until the keyframe passes.