[Feature Request]: Add Würstchen / Stable Cascade support

[dome272]

Is there an existing issue for this?

• I have searched the existing issues and checked the recent builds/commits

What would your feature do ?

We released Würstchen 2 days ago in Diffusers. The blog post can be found here: https://huggingface.co/blog/wuerstchen
TLDR; It is much faster than SDXL, requires less memory and generates images at higher resolutions.
The tradeoff? Fine details are hard for the model due to its f42 spatial compression of images.
However, with future versions we are sure we can improve this a lot.
The GitHub with code is here and the Diffusers documentation is here.

Proposed workflow

It could just be loaded like any other model in the GUI.

Additional information

No response

[prp-e]

LGTM.

[paniq]

SD.Next (a A1111 fork) just implemented it, but had to roll back support as there's an undisclosed problem with the latest diffusers package

[operationairstrike]

I hope this gets added. But please try making your own fork until it's in the official version.

[XZVB12]

In case you're wondering, basic support for Würstchen has been implemented in the project DiffusionMagic . There's almost no customization, but it does allow for some experimentation.
I can't say that images are better than on SD XL in terms of quality, but resource consumption is really lower and processing is faster.
Maybe custom models will improve image quality (hopefully) like SD did, but when that will happen is not clear. The promise of faster learning gives hope that it will be soon, but many models will forever be SD-only.

[operationairstrike]

Yeah but it doesn't seem to load on Google colab

[rupeshs]

Try DiffusionMagic, diffuser-based diffusion workflows; supports wurstchen ( VRAM < 4GB).Google Colab is also available.
https://github.com/rupeshs/diffusionmagic

[paniq]

For reference, the stable cascade project is here https://github.com/Stability-AI/StableCascade

[CCpt5]

There's a "temporary" extension for A1111 that will allow you to generate w/ Stable Cascade (it an independent tab) here:

https://github.com/blue-pen5805/sdweb-easy-stablecascade-diffusers

Hopefully it's natively implemented sometime in the near future.

[VeteranXT]

AssertionError("Torch not compiled with CUDA enabled") dose not work with AMD. Even with directml argument wont make it work!

[Pevernow]

Looking forward to this feature. This is a great update no less than SDXL.

[VeteranXT]

Looking forward to this feature. This is a great update no less than SDXL.

There is extension for this. https://github.com/blue-pen5805/sdweb-easy-stablecascade-diffusers

But for me it does not work

[placeboyue]

hurry! Every day without this there's hordes of people flocking to comfyUI. I don't want to, i have all my shit set up for A1111.
I know devs don't owe me anything tho, but as an user of years i think i can at least ask, right? i can be ignored.

[PierrunoYT]

Is this added?

[paboum]

What's the release schedule for this feature? What is the top blocker? It really hurts my projects when the competition can use models that are half a year newer. I will start familiarizing myself with ComfyUI next week if this bottleneck persists.

[QuintessentialForms]

I'm being forced to focus ParrotLUX support on ComfyUI (despite its terrible API) because Stable Cascade is not supported in A1111. Is there any way to make this feature happen?