[WIP] Add Latent Consistency Model Img2Img Pipeline

[dg845]

What does this PR do?

This PR adds LatentConsistencyModelImg2ImgPipeline, a text-guided image-to-image version of LatentConsistencyModelPipeline based on the community pipeline by @nagolinc.

Based on PR #5477, which was originally in response to issue #5462.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@patrickvonplaten
@sayakpaul
@nagolinc
@luosiallen

[nagolinc]

Did you do any experiments before changing this line?

https://github.com/huggingface/diffusers/blob/main/examples/community/latent_consistency_img2img.py#L326

https://github.com/huggingface/diffusers/pull/5559/files#diff-205b2ad5eac653752af8760192d8315263a5aec4203451d38f3b1db2b20551b9R672

The reason why I am passing the strength to set_timesteps (and using a custom scheduler) is that I found for very small number of timesteps (n=4) the method here (https://gist.github.com/nagolinc/6e0d3960a81714d13b821e3694d673d6) does not work well.

[dg845]

Thanks for the info @nagolinc! I haven't done any experiments yet. I will try to see if I can handle the strength logic within the pipeline code itself to avoid passing it to set_timesteps (which I think would modify the scheduler API). I guess this might require changes to how LatentConsistencyModelImg2ImgPipeline.get_timesteps and LCMScheduler.set_timesteps are implemented.

The reason why I am passing the strength to set_timesteps (and using a custom scheduler) is that I found for very small number of timesteps (n=4) the method here (https://gist.github.com/nagolinc/6e0d3960a81714d13b821e3694d673d6) does not work well.

Just to clarify, does "not work well" mean that the code runs into errors, or that the generated images are low quality? I'm running into some errors when testing on different timesteps with the code in 86adad0, but it might be because there are still some other bugs that I haven't found yet.

[HuggingFaceDocBuilderDev]

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

[nagolinc]

The images look bad:

using 1st method (https://gist.github.com/nagolinc/6e0d3960a81714d13b821e3694d673d6)

using 2nd method (https://gist.github.com/nagolinc/53c395d8cd70a852b03fca4eb273923f)

here's a notebook I was using to test out the two methods

https://github.com/nagolinc/notebooks/blob/main/lcm%20img2img.ipynb

[dg845]

Hi @patrickvonplaten, since #5632 has been merged, does this PR still make sense?

[sayakpaul]

Yeah, I think we can safely close it.

[dg845]

Closing due to #5632 being merged.