Both the FID scores and visualization are pretty bad in FFHQ dataset

crowsonkb / k-diffusion

Karras et al. (2022) diffusion models for PyTorch

MIT License

2.33k stars 378 forks source link

Both the FID scores and visualization are pretty bad in FFHQ dataset #52

Open IrohCao-NYU opened 1 year ago

IrohCao-NYU commented 1 year ago

Hi authors:

Thanks for your work.

The original K-Diffusion paper report FID scores for FFHQ dataset (64x64) is about 2, but when we use your code repo, the FID is over 60 and the visualization results are very bad after 100 epoch with 256 batch size training. What do you think?

Best, Tutu Iroh Zhao

frangam commented 1 year ago

I think there is a bug in the use of "accelerate launch" for launching train.py with acceleration. If you do not use "accelerate" FID is close to 2 points.