Output is a garbled mess
I tried as you suggested with the INT8 loader set to flux2 and on-the-fly on true. Just artifacts. Works with Z-Image and Flux models and Wan. Haven't got it to work with Anima. Any suggestions?
Are you certain you are using the correct branch? QUIP only works with quip branch https://github.com/BobJohnson24/ComfyUI-Flux2-INT8/tree/Quip or this node https://github.com/ThunderFun/ComfyUI-Wan-INT8
Yes. I have done exactly as you asked. And reinstalled the repo. Cannot produce an image.
Oh wait, on the fly should be set to false, because this is a prequantized model.
I left it to false and tried three INT8 models, the QUIP, my own, and another one. All three give either black or heavily garbled outputs.
That seems very weird. Are you doing anything out of the usual, like sage attention? What's your gpu? Upload your workflow and I'll take a look.
RTX 3090, I tried without Sage and with. I tried 3 models and finally got 1 out of the 3 working. The regular INT8 version worked! The other two kept reproducing artifacts. For workflow I first tried the Comfyui sample template and used the INT8 Diffusion loader and kept "on the fly quantization" off like you said. Surprisingly, the regular INT8 version worked, but the QUIP one and Tensorwire one didn't.
I think in my testing, the regular int8 did have some artifacting issues for me. If you get sick of that, I'd be curious to hear if this node works any better. https://github.com/ThunderFun/ComfyUI-Wan-INT8
I tried that ThunderFun link too, but no solution. Always garbled. We really need a Cosmos2 selection in the dropdown menu.