Can't train model, only green samples generated

[furkan-celik]

Hi,

I am working on a project where we want to use Stable Diffusion. I have already used OpenAI's improved-diffusion and score-guided diffusion on my environment. However, when I try to use stable diffusion all I get is green samples and blank picture for class conditioned images. I have looked at this and it is the same from the initialization of the model but no weight is NA or INF. When I use extreme learning rates outputs get black but still no learning. Loss also doesn't change and stay at 1. Any idea how to resolve this?

GPU: Nvidia A40
Image size: 256x256
I have tried with CelebA and custom dataset with ImageFolder function from torchvision.
Environment is set by 'conda env create -f environment.yaml' and environment is clean besides these packages.
Tried running for 12 hours observed no change on outputs.
Inputs saved from ImageLogger looks fine.
This is also mentioned in #69 but no solution is provided.

Reconstruction result

Diffusion Row Result

[fujistoo]

Did you manage to solve this?

[furkan-celik]

Yes, I can't recall how I solved thıs though. Do you still need assistance?