Annealed Langevin Sampler

[natolambert]

Closes #84

Maybe @MinkaiXu can give a review on this?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[patrickvonplaten]

@anton-l could you review here? :-)

[Suhail]

It would useful to compare this sampler to other samplers with the same prompt/seed/etc with images to see the difference it might have on images.

[anton-l]

@natolambert would it be possible to add an integration test with one of the existing models? Both for a usage example to understand the API (see Patrick's comment above), and for test coverage

[natolambert]

Yeah @anton-l and @Suhail I agree with both points. I'll do some experiments and share the results + tests

[patrickvonplaten]

Leaving open for now :-) Let me know if you require another review @natolambert - but I agree with @anton-l. We'd need some tests and also I think we might have to change the API regarding timesteps rather being indices.

[natolambert]

Yeah I'll come back to this when I have time. Mostly this was a learning side-project for me, so leaving open is preferred.

[natolambert]

What do people think. Is it worth spending time fleshing this out? SDE VE is the improved version of this scheduler, so unless we want to include it for completeness, it's not adding much.

Really it goes from NCSN to NCSNv2 to SDE VE/P.

We don't have the samplers for NCSN or NCSNv2 exactly. I could work to add them, but it may be better for a community member to pick this up?

[patrickvonplaten]

I think pretty much no-one uses SDE VE and SDE VP or Karras VE (<- this scheduler has a design that doesn't fit really anyways)

Think a community contributor could def take this over :-)

[github-actions[bot]]

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.