jolibrain / joliGEN

Generative AI Image Toolset with GANs and Diffusion for Real-World Applications
https://www.joligen.com
Other
236 stars 31 forks source link

feat: allow to resume training with a different number of timesteps #627

Closed Bycob closed 5 months ago

Bycob commented 5 months ago

When we resume a diffusion model, maybe we want to change the number of sampling steps or the scheduling during inference. This was not possible because the gammas were reloaded from the checkpoint, and not recomputed according to the user options.

Changing test sampling steps does not have any impact on training, only the visuals are affected.