DPM++ SDE variant

[patrickvonplaten]

Model/Pipeline/Scheduler description

It looks like there is a variant of the very nice DPM++ scheduler on k-diffusion, see here: https://github.com/crowsonkb/k-diffusion/blob/5b3af030dd83e0297272d861c19477735d0317ec/k_diffusion/sampling.py#L543

It would be nice to add this one to diffusers as well.

Also gently pinging the author of DPM++ @LuChengTHU (have you played around with a stochastic variant of your scheduler as well?)

Open source status

• The model implementation is available
• The model weights are available (Only relevant if addition is not a scheduler).

Provide useful links for the implementation

No response

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[LuChengTHU]

Theoretically speaking, I don't think the implementation of the SDE-variant of DPM-Solver / DPM-Solver++ in k-diffusion is correct :( It is not corresponding to the reverse diffusion SDE, though it can provide beautiful images (because diffusion models are "tolerant" to SDE solvers).

I've implemented the DPM-type SDE solver in this PR: #3344

[patrickvonplaten]

Perfect! Much nicer that we have the "official" version from you now @LuChengTHU :-)

[a-r-r-o-w]

Marking as closed because this is supported now, thanks :)