Fix doesn't work for GTX 1650
I copied the exact code of demo, provided in README, however it still produces black images.
Hmm, it works fine on Colab, so there's probably some issue with your specific software or hardware setup.
- You could try using the XFormers implementation in case that bypasses the issue.
- You could try running the UNet (non-VAE) portion in full float32 precision like some people have suggested here.
- You could also try updating various components (CUDA, PyTorch, Diffusers) just in case something was fixed.
If you just want to confirm that this VAE isn't the issue, you can using TAESDXL (cheap preview VAE). If TAESDXL also produces black images, then the NaN/black image problem is coming from the UNet (input to the VAE) and not the VAE itself.
Hmm, it works fine on Colab, so there's probably some issue with your specific software or hardware setup.
- You could try using the XFormers implementation in case that bypasses the issue.
- You could try running the UNet (non-VAE) portion in full float32 precision like some people have suggested here.
- You could also try updating various components (CUDA, PyTorch, Diffusers) just in case something was fixed.
If you just want to confirm that this VAE isn't the issue, you can using TAESDXL (cheap preview VAE). If TAESDXL also produces black images, then the NaN/black image problem is coming from the UNet (input to the VAE) and not the VAE itself.
Hello! First method doesn't work for me because I have CUDA 12.6 and I was unable to compile for my version of PyTorch.
I can't find anything about UNet on the second linked page. Are you sure you posted the correct link?
Thank you.
@GulgDev The reddit thread doesn't explicitly mention UNet, but I think all of their suggested precision changes apply to the UNet (not just VAE portion)