Managed to get the code to run locally with my RXT-3060 however, it is running on CPU only (???)
I tried installing torch for my specific version of windows however could not generate anything as CUDA ran out of space.
I can generate a mesh in about 8 minutes at 15 steps but, wondering if anyone has gotten this running with their GPU locally.
JFYI, the GPU memory usage is about 9GB when I run this Space on my AWS environment with T4. RTX 3060 seems to have 12 GB VRAM, so I guess it should work. Maybe it's something to do with Windows environment? I can't check it myself as I don't have a Windows environment, though.
Ah, yeah 9GB is probably too much. While I don't mind waiting 8 minutes for a mesh (considering it would take me far longer to do by hand) was wondering if something like VAE would help performance and GPU size but, I don't have to coding experience in this domain to know if that is even possible.
@pjonesdotca Hmm, I don't have much knowledge on optimization and am not sure if it's possible either. FYI, currently, diffusers team is working on adding Shap-E to diffusers library in this PR, and their implementation might be more optimized than the original one if it's possible.
BTW, this Space can be run on T4, so I think you should be able to run this Space on free Colab as well.
you can speed up sampling speed 2 times by changing use_fp16=False in settings.py. it also speed ups sampling on cpu 64 times