ComfyUI is not deterministic

[asd417]

All these images are generated with exact same seed and node structure. I believe that the images above maintain their exif data but here is the node structure:

You can see that the two seed inputs are linked together. the ui does not even let me generate this twice in a row because it's the same prompt.
I think it's a bit of a problem for this to be not deterministic. Will this issue be fixed?

[comfyanonymous]

Yeah that's an issue I need to fix.

[wyrde]

didja fix it?

[jexom]

It's an xformers problem. Xformers optimizations make the generations undeterministic, hence why I stopped using it in auto webui long ago. Haven't looked into it but there should be a way to disable xformers if there's none.

[ill13]

To [mostly] address this issue, one can run ComfyUI with

python main.py --use-pytorch-cross-attention

This should disable Xformers and allow for repeatiblity across generations

[catboxanon]

xformers is deterministic as of 0.0.19. https://github.com/facebookresearch/xformers/releases/tag/v0.0.19

[ill13]

Good to know! Maybe I'm complelety offbase with the Xformers stuff?

Either way, until I used the above command all of my generations with ComfyUI were unrepeatable and random - directly after starting ComfyUI with --use-pytorch-cross-attention changing seeds back and forth generated what looks to be the same image.

[gonggp]

“Has this issue been fixed? I used --use-pytorch-cross-attention and adjusted the sampler, but even with the same seed and parameters, the generated images still have slight differences. Additionally, when used in conjunction with adapters, this difference is amplified.” ---> for example: there are some differences near the ears.

[gonggp]

used --use-pytorch-cross-attention --deterministic