denoising-diffusion-pytorch issues

change of beta_schedule leads to significantly worse results

I just altered the beta_schedule parameter in the class GaussianDiffusion from the default 'sigmoid' to 'linear', which was utilized in the original ddpm paper. However, the results were unreasonably much...

BillyChern

Loss on Unet1D

Dear author, I have tried to run default setting of Unet1D. However, the loss would not converge and the generated distribution is not a uniform distribution (torch.rand). Have you ever...

wanshuiyin

Sanity Check - Looking for a basic CIFAR10 hyperparameter set

6

I'm running the `denoising_diffusion_pytorch.py` script as-is on the CIFAR10 dataset, however the FID quickly plateaus to ~90, which is a far cry from both those reported in the DDIM/DDPM paper...

samuelemarro

training on CIFAR-10 performs not well, whether L2 loss or FID

hi developers: I am sorry to bother you! I am training on `CIFAR-10`, with setting `train_num_step=500k, batch=16, unet dim 64`. But get FID 21.57 not well. And when i increase...

SaulQiu

A question related to batch size and training speed

2

Hi! The code in this repository has helped me a lot! I found that as the batch size increases, the training time increases dramatically. When I set the batch size...

paidaxinbao

scale up UNet with different resolution

Hi Phil, First and foremost, I would like to express my gratitude for your DDPM implementation. Our recent publications ([DMHomo](https://github.com/lhaippp/DMHomo), [RecDiffusion](https://github.com/lhaippp/RecDiffusion)) have been based on your repository, and we are...

lhaippp