IMPORTANT

This model utilizes INT6 packing and cannot run natively on ComfyUI, clone the below custom node into your ComfyUI's custom_nodes directory to enable it

https://github.com/bakapotatolord/ComfyUI-PotatoForge-INT6

And yes, INT6, not a typo.

It is an experiment to find a balance between INT4 and INT8 in terms of storage and quality. It's not a mixed model, quantized layers are INT6 Convrot. Generation speed is almost the same as INT8 Convrot.

Tested on GTX 1660 Super so I cannot guarantee how it works on other GPUs.