>>28017
Forgot to mention, another finding I have is that compute time scales approximately quadratically with clip length, so it's better to think in terms of 3-6s gestures that stringed together compose the scene.

Actual clip length Recorded generation time
5.17 seconds 12.3 minutes
10.13 seconds 35.2 minutes
15.08 seconds 68.8–86.6 minutes

There's definitely more experimenting I want to do before I upload my first clip, but shorter clips should also allow for character refresh from reference images. I found that if the pussy of a character went offscreen then back on screen across different generations, it was horribly malformed.