search

lucidrains / denoising-diffusion-pytorch

Implementation of Denoising Diffusion Probabilistic Model in Pytorch
MIT License
8.09k stars 1.01k forks source link

Output resolution issues in CIFAR-10? #88

Open VidushB opened 2 years ago

VidushB commented 2 years ago

Hey. Since images are relatively small for CIFAR-10, I tried to run the following 

 model = Unet(
    dim = 64, #Just affects how many nodes are in our blocks
    dim_mults = (1, 2, 4)
).cuda()

This model seems to be better then the model where we go down to 1/8. But my output resolution is just terrible. I also fixed sampling based on edits suggested in https://github.com/lucidrains/denoising-diffusion-pytorch/issues/81. After training, I set sampling time steps to 1000 for evaluation for best possible image quality

The losses look perfectly fine to me. They drop drastically before stabilising. But even then my output images are just very poor in terms of resolution. Even in my best images, the output is purely a vague outline and nothing close to what was observed in the original paper.

yiyixuxu commented 2 years ago

I saw many people complain about cifar10 output being blurry here - just curious did you change 'image_size' to 32? by default it resize the training data to 128 and the result will naturally look blurry