Winnougan

/

Wan2.2-INT8-Convrot

like 33

Hi, could you make an INT8 Convrot quantization of the models from this repo?

https://huggingface.co/lightx2v/Wan2.2-Distill-Models

The repo already provides INT8 versions, but they don’t work in ComfyUI.
From my testing, the Distill models seem to be the proper way to achieve good low step generation. Using the base Wan2.2 with a Turbo lora doesn’t seem to produce the same good results, the Distill models generate noticeably more motion at low step counts and more crisp image.

If possible, could you quantize both the high-noise and low-noise models? Specifically, the High 1030 and Low model.

Thanks.