Sampler implementation

[andreaferretti]

Hi Lou, thank you for your great work, I think it goes in a very exciting direction!

If I understand correctly, it should be possible to package the reflected SDE technique as another inference time sampler, to be used with pretrained diffusion models, much like it has been done for the ones in k-diffusion, or DPM solver. Are there any theoretical obstacles? Or can your technique be used as is with any image diffusion model?

If you were to produce an implementation compatible with diffusers, I think it would greatly expand the reach of your approach. For instance, it would become immediately usable together with the many variants of Stable Diffusion.

Unfortunately my knowledge of SDEs is too primitive to do this myself, at least in the short term

[louaaron]

Hi Andrea,

Thanks for the comment and suggestion. Currently, the approach is a bit specialized to the training method (at least in theory). If the original sampling method used in the DDPM paper (which has thresholding) is implemented in diffusers, perhaps our advances could be ported over.