search

genmoai / mochi

The best OSS video generation models
Apache License 2.0
2.13k stars 214 forks source link

Turn Mochi into a library #34

Closed ved-genmo closed 4 weeks ago

ved-genmo commented 4 weeks ago

Credits to https://x.com/kijaidesign?lang=en for a tiled VAE decoding method, and other OSS collaborators for implementing things like sage attention, Pytorch SDPA, etc.

Ednaordinary commented 1 week ago

Hey! I'm curious if num_inference_steps should always be a number squared, i.e for the scheduler to better align denoising steps. Is this correct, or am I reading into the wrong pattern?