Lesson 12 of 22

The clip3 min

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.

Next: The beat map

Every law on one page

The compiled reference — framing laws, the negative prompt block and the pre-ship checklist. One email, no sequence you can't leave.

The pack, and a note when a new teardown goes up. Unsubscribe whenever.