Comment by vunderba

From the paper:

> a variable-length training approach is adopted, with training durations ranging from 2 to 8 seconds. This strategy enables our model to generate 720p video clips with flexible lengths between 2 and 8 seconds.

I'd like to see it benched against FramePack which in my experience also handles 2d animation pretty well and doesn't suffer from the usual duration limitations of other models.

https://lllyasviel.github.io/frame_pack_gitpage