Bad output on Intel A770
crispysilicon opened this issue · 6 comments
You wanted to know how it ran on Intel.
SD 2.1 large, base fp-16 top, base fp-32 bottom.
I've tried converting an earlier version of SD, changing the steps and guidance scale, always bad output though.
Yes, another user was having the same issue. It seems like DirectML is currently not working with Intel drivers.
Unfortunately ONNX Runtime 1.14 has not released for OpenVino either. I hence fear there's no Arc support there either.
Your best bet may be to use Intel extension for PyTorch which in theory gives native support in torch and hence you could run nearly every Stable Diffusion implementation if you find your way around Python. Hopefully the driver issues get fixed soon. If I bump into a cheap way to get an Arc card I might try to figure out what's happening.
I was so excited to see this, and actually got it running with Arc A770... and then found that the output is broken.
Not sure if this is any help, but this fork currently works with Arc:
https://github.com/bes-dev/stable_diffusion.openvino
However, it is very much lacking in features and options, and you'd need to figure your own way around getting the SD 2.1 models into the right format and supported by that project. I'd love to see something merge whatever is going on with that OpenVINO branch into a functional webui with SD 2.1!
(FWIW, I got around 7.9 iterations per second on A770 with 512x512 generation and 50 steps.)
Would be interesting to know whether the latest drivers changed things for any of you with Arc. Given they did some Direct X fixes, they may have fixed DirectML ...
I am happy to report that the driver team has resolved this in the latest public release and Intel users are now in business.
Nice, thanks for letting me know!