Lesson 12 of 22
When the model pads and repeats
A short line in a long clip doesn't get slower. It gets repeated.
Video models that generate speech fill the clip length they were given. Hand one a five-second line and a nine-second clip and it does not deliver the line slowly with a pause — it pads. Usually by repeating words or part of a sentence.
This shows up as clips where the person says something, then says a piece of it again, slightly differently. It's uncanny in a way that's hard to place if you don't know what caused it.
Fix it upstream
- Match the requested duration to the line, snapping to the nearest available tier rather than rounding up.
- Give the prompt an explicit timing cue so the model knows how the words map onto the seconds.
- Keep a repeat-detection pass downstream as a safety net — but treat it as a net, not the fix.
The general principle
When a generative model produces something strange, check whether you handed it a shape it had to fill. A lot of "model weirdness" is really the model solving a constraint you didn't realise you'd set.
Takeaway
Padding is a duration mismatch, not a model defect. Fit the clip to the line.
Course outline
Progress
0 of 22 lessons completed
Chapter 1
Start here
Chapter 2
The script
Chapter 3
The frame
Chapter 4
The clip
Chapter 5
The edit
Chapter 6
Judging it honestly