Add Karras Sigmas to all schedulers

[patrickvonplaten]

Is your feature request related to a problem? Please describe.

As stated in multiple issue:

• How can we use DPM++ 2M Karras sampler in Diffusers? #1887
• "DPMSolverMultistepScheduler "Is not" dpm++ 2M K ", "DPMSolverMultistepScheduler" = "dpm++ 2 m" Karras doesn't seem to be effective #2801
• "What Schedulers are supported on diffusers? I found some in the webUI that I couldn't find on diffusers, such as "DPM++ 2M Karras" and "DPM++ SDE Karras". Does anyone know their names on diffusers?" #2676

Describe the solution you'd like

It would be nice to be able to do:

scheduler = DPMSolverMultistepScheduler(use_karras_sigmas=True)

and

scheduler = DPMSolverMultistepScheduler.from_config(scheduler.config, use_karras_sigmas=True)

Describe alternatives you've considered
No alternative possible

Additional context
We have a nice PR: #2874 that solves this for the Karras stable diffusion pipeline, but it would be nice to have this for all native diffusion schedulers as well.

#2874 can be very nicely used for integration testing.

Also cc @sayakpaul @takuma104

[aandyw]

Seems fun. Currently working on the LoRA tests PR but I'd like to take a crack at this if it's not urgent.

[sayakpaul]

Seems like a requested feature. I will try to allocate some time next week. @patrickvonplaten, anything you have in mind regarding the design I should know? I think we should consider separate PRs for separate schedulers to help with the reviews and also to keep the PRs disentangled.

@pie31415 thanks so much for being willing to help! We deeply appreciate it. Let's see how we want to tackle this and based on that we can circle back :)

[aandyw]

Looking forward to it!

[patrickvonplaten]

Regarding design I think it'll be all about changing the sigmas here correctly:

diffusers/src/diffusers/schedulers/scheduling_euler_discrete.py

Line 140 in c433562

sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)

In my opinion some kind of helper function that we can easily copy-paste to other schedulels like:

if use_karras_sigmas:
    sigmas = self.convert_to_karras(sigmas)

would make sense

[sayakpaul]

#2956 could be a good reference PR to add use_karras_sigmas to the rest of the schedulers. Let's continue the discussion there.

[sayakpaul]

Related PR: #3001