SVDQuant support?

[0x4E69676874466F78]

https://www.reddit.com/r/StableDiffusion/comments/1h9d9xy/svdquant_now_has_comfyui_support/ https://github.com/mit-han-lab/nunchaku https://github.com/mit-han-lab/nunchaku/tree/main/comfyui https://huggingface.co/mit-han-lab/svdquant-models/tree/main

[syddharth]

@Acly This would be great! These models take up less VRAM and are much faster. Supports dev and fill versions both.

[syddharth]

Link to ComfyUI node : https://github.com/mit-han-lab/ComfyUI-nunchaku/ (I use them regularly in my workflows)

[tazztone]

by now the project might even be mature enough to use it? the 3x shorter wait times add a lot of joy. it works with the union pro 2 controlnet and turbo lora, as well as flux fill

[AlexYez]

Yea, it is ready for integration: Release v0.3.0 adds support for multiple-batch inference, ControlNet-Union-Pro 2.0 and initial integration of PuLID. You can now load Nunchaku FLUX models as a single file, and our upgraded 4-bit T5 encoder now matches FP8 T5 in quality! Support Pytorch 2.7 and Python 12

[angelika3377]

my krita doesn't see the nunchaku models, how to fix this?

[Acly]

supported in v1.38.0

[tazztone]

they recently portded nunchaku from c to python i heard. hope this doesn't cause issues for dev.