[Schedulers] Add methods for conversion between noise and denoised predictions

[hlky]

Guidance methods like APG (#9585 #9626) require conversion between noise and denoised predictions. The computation of these predictions varies across schedulers and depends on prediction_type. Currently, this is embedded in step functions of the schedulers which complicates implementation of APG.

Adding separate methods for conversion between noise and denoised predictions would simplify implementation and allow easier experimentation.

[yiyixuxu]

Do you need a function like this we have for dpm?

diffusers/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

Line 594 in 07bd2fa

def convert_model_output(

[hlky]

Yes like that, there's currently also something similar in UniPC, they'd need to be controllable with a parameter rather than config or internal variables, and handle different sigmas for 2nd order schedulers like Heun and KDPM2

[yiyixuxu]

Make sense!

Do you want to open up a PR for UniPC? We'd need to keep this method consistent across all schedulers that have them, so we need to use the same name and signature and return the same output etc

For the rest of the schedulers, we can also open this task up to communities if you don't have much time: we would need to run slow tests for each scheduler we refactor here, so it might be a bit time-consuming for one person to work on!

[hlky]

Sure, I'll start with UniPC. I don't mind doing the rest of the schedulers, I don't have much else going on, more community contributions in schedulers would be good though!

[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.