denoising-diffusion-pytorch

denoising-diffusion-pytorch copied to clipboard

Hello, I downloaded your latest commit code. I feel that the reconstruction effect is much worse than the code downloaded in March. Have you observed the same phenomenon? Thanks a lot!

@XDUWQ Hi Wang, so i thought the culprit could be migrating to the next ConvNext blocks, but that was done in January https://github.com/lucidrains/denoising-diffusion-pytorch/commit/91cff459394b784f870f3998023439796c95b7ca

You could always try checking out the specific commit in the repository and doing a comparison

Would be really interested if you figure out what it was

The only other thing I changed was moving linear attention to full attention, but that should lead to an improvement

@XDUWQ there was also this bug fix https://github.com/lucidrains/denoising-diffusion-pytorch/commit/782c904d3b5240b79fb66929ca8ab9cf6d2496e5 , but i think the patch is the correct way to do it

@XDUWQ i brought back the old resnet blocks, so you can try turning on and off ConvNext using 'use_convnext`

curious what you find is better!

@XDUWQ i brought back the old resnet blocks, so you can try turning on and off ConvNext using 'use_convnext`

curious what you find is better!

Thank you for your contribution! I'm trying to use your suggestion…

So, should we be using a commit from March? Has anyone else verified the change in behavior from March until now? @XDUWQ do you have any examples of the difference in reconstruction quality?