[Core] Add Kolors

[asomoza]

What does this PR do?

Adds Kolors from the Kwai-Kolors team

Text to Image

importtorchfromdiffusersimportDPMSolverMultistepScheduler, KolorsPipelinepipe=KolorsPipeline.from_pretrained("Kwai-Kolors/Kolors-diffusers", torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.scheduler=DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)
prompt='一张瓢虫的照片，微距，变焦，高质量，电影，拿着一个牌子，写着"可图"'image=pipe(
prompt=prompt,
negative_prompt="",
guidance_scale=6.5,
num_inference_steps=25,
).images[0]

Image to Image

importtorchimportmathfromdiffusersimportDPMSolverMultistepScheduler, KolorsImg2ImgPipelinefromdiffusers.utilsimportload_imagepipe=KolorsImg2ImgPipeline.from_pretrained("Kwai-Kolors/Kolors-diffusers", torch_dtype=torch.float16, variant="fp16").to("cuda")
pipe.scheduler=DPMSolverMultistepScheduler.from_config(pipe.scheduler.config, use_karras_sigmas=True)
source_image=load_image(
"https://huggingface.co/datasets/OzzyGT/testing-resources/resolve/main/kolors/bunny_source.png?download=true"
)
prompt="high quality image of a capybara wearing sunglasses. In the background of the image there are trees, poles, grass and other objects. At the bottom of the object there is the road., 8k, highly detailed."strength=0.65steps=25image=pipe(
prompt=prompt,
image=source_image,
negative_prompt="",
guidance_scale=6.5,
num_inference_steps=math.ceil(steps/strength),
strength=strength
).images[0]

source	strength 0.65	strength 0.9

Fixes#8801

TODO

• Tests
• Docs

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.

@yiyixuxu@sayakpaul@JincanDeng

[yiyixuxu]

looks good to me overall!
I wonder if we should call it pipeline_stable_diffusion_xl_kolors? it is sort of a "variant" , let's ask the author for their preference!

[asomoza]

I'm ignoring for now the failed test with the docs until I finish with all the other stuff. Kind of curious why the docs are the only ones that fail with a circular import.

[s9anus98a]

u guys need made it under diffusionpipeline just like playground

[asomoza]

@s9anus98a We decided not to do that because this pipeline doesn't share the same defaults and also nothing from SDXL works with this model right now, like IP Adapters, LoRAs and TIs

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[DN6]

Looks good to me. Just some minor fixes related to docstrings. Could we also add fast tests please.

[a-r-r-o-w]

Looks absolutely amazing! Just a few missing copied-froms

[vladmandic]

it seems its not picking up the selected variant correctly?

importtorchimportdiffusersfromrich.tracebackimportinstallinstall()
pipe=diffusers.KolorsPipeline.from_pretrained(
'Kwai-Kolors/Kolors',
variant='fp16',
torch_dtype=torch.float16,
cache_dir='/mnt/models/Diffusers',
)

OSError: Error no file named diffusion_pytorch_model.bin found in directory /mnt/models/Diffusers/models--Kwai-Kolors--Kolors/snapshots/f0d70fcea007ce197eea9631a35507d827e4a72e/unet.

there are both fp32 and fp16 variants available in https://huggingface.co/Kwai-Kolors/Kolors/tree/main/unet, but they were not picked up.

this is the local snapshot state - no unet downloaded at all:

 └── f0d70fcea007ce197eea9631a35507d827e4a72e
├── model_index.json -> ../../blobs/d473050985693036d4be16253f0d12418b3c3378
├── scheduler
│ └── scheduler_config.json -> ../../../blobs/4cb73ebda6cfb0a97a2eefd289ef0bd27d8f712e
├── text_encoder
│ ├── config.json -> ../../../blobs/c6e19300822b25ae0a07125bbc171c6581dbeda4
│ ├── pytorch_model-00001-of-00007.bin -> ../../../blobs/b6a6388dae55b598efe76c704e7f017bd84e6f6213466b7686a8f8326f78ab05
│ ├── pytorch_model-00002-of-00007.bin -> ../../../blobs/2f96bef324acb5c3fe06b7a80f84272fe064d0327cbf14eddfae7af0d665a6ac
│ ├── pytorch_model-00003-of-00007.bin -> ../../../blobs/2400101255213250d9df716f778b7d2325f2fa4a8acaedee788338fceee5b27e
│ ├── pytorch_model-00004-of-00007.bin -> ../../../blobs/472567c1b0e448a19171fbb5b3dab5670426d0a5dfdfd2c3a87a60bb1f96037d
│ ├── pytorch_model-00005-of-00007.bin -> ../../../blobs/ef2aea78fa386168958e5ba42ecf09cbb567ed3e77ce2be990d556b84081e2b9
│ ├── pytorch_model-00006-of-00007.bin -> ../../../blobs/35191adf21a1ab632c2b175fcbb6c27601150026cb1ed5d602938d825954526f
│ ├── pytorch_model-00007-of-00007.bin -> ../../../blobs/b7cdaa9b8ed183284905c49d19bf42360037fdf2f95acb3093039d3c3a459261
│ └── pytorch_model.bin.index.json -> ../../../blobs/cd6a94a411934e8cc04e271f169a68754d14733d
├── tokenizer
│ ├── tokenization_chatglm.py -> ../../../blobs/50e44b05e4b3e54d2f1c3f0cab8247ea53a7d4e5
│ ├── tokenizer.model -> ../../../blobs/e7dc4c393423b76e4373e5157ddc34803a0189ba96b21ddbb40269d31468a6f2
│ ├── tokenizer_config.json -> ../../../blobs/f6f13c88707490cebd8023da86e8bf7a56fa21e3
│ └── vocab.txt -> ../../../blobs/e7dc4c393423b76e4373e5157ddc34803a0189ba96b21ddbb40269d31468a6f2
├── unet
│ └── config.json -> ../../../blobs/52ab049bc6d1490ead84272eb3736822e1d02d63
└── vae
├── config.json -> ../../../blobs/6e9694046afd2a944dd17a2390b98773cacf2f7c
└── diffusion_pytorch_model.fp16.bin -> ../../../blobs/2ce744db8ec41697eaecabe3508566aa76e53d71f79e595b0d0f56c9f07405ce

[asomoza]

@vladmandic it was the decision of the model authors to have a different repo for the diffusers compatible model, so the repo is Kwai-Kolors/Kolors-diffusers and not Kwai-Kolors/Kolors

[vladmandic]

thanks for pointing it out, it works
two notes:

• Kolors requires pipe.vae.config.force_upcast = True which is no longer needed on most modern models.
  any chance authors can use fp16-fixed vae?
• and option to quantize massive text_encoder during load would be a nice addition.
  right now, ChatGLMModel does not expose underlying transformers quantization options.

(yes, both can be monkey-patched in app, just would be nice to have out-of-the-box solution)

[lixiangcog]

https://huggingface.co/Kwai-Kolors/Kolors-diffusers/tree/main

Currently, only fp16 models are available under huggingface. Should fp32 models also be provided to allow control via the variant parameter in from_pretrained?

[asomoza]

@vladmandicthe vae will be changed to the fp16fix one, about the quantization we're still deciding about it, the original one has an external dependency and we want to probably use transformers for it.

edit: The vae won't be changed since there was a performance degradation detected, users will need to replace it by themselves.

[asomoza]

Currently, only fp16 models are available under huggingface. Should fp32 models also be provided to allow control via the variant parameter in from_pretrained

We can add them but AFAIK the model was trained in fp16, so you'll be just wasting VRAM with this. @JincanDeng can you confirm this?

[sayakpaul]

Thank you! Looks veryyyyy good. Left a couple of comments.

[sayakpaul]

Not a blocker for merging but did you have some time play around with these params? They are fun.

[asomoza]

Now that you mention it, I've never played with these, not even with the base SDXL, I'll do some experiments but for kolors they change the generation but doesn't seem to do what they should. Maybe because of the training?

[sayakpaul]

Could very well be. @JincanDeng do you have any suggestions here?

[JincanDeng]

Currently, only fp16 models are available under huggingface. Should fp32 models also be provided to allow control via the variant parameter in from_pretrained

We can add them but AFAIK the model was trained in fp16, so you'll be just wasting VRAM with this. @JincanDeng can you confirm this?

Yes, the model was trained in fp16.

[vladmandic]

edit: The vae won't be changed since there was a performance degradation detected, users will need to replace it by themselves.

can you share that info? fp16-fixed should be better than requiring to run vae with upcast?

[asomoza]

can you share that info? fp16-fixed should be better that requiring to run vae with upcast?

That was the answer I got, I don't have more info. I'm also curious about it so I intend to do some test when I have time.
Kolors could be more sensible to the vae since it's a lot more aesthetic.

Also probably the human eye doesn't see the difference between them but the numbers are more important to researchers (understandable) than the VRAM or if there's no visual difference.