FID for CIFAR10 checkpoint about ddpm-torch HOT 2 CLOSED

Hi there! It seems that the example command you were using is the one with a DDIM sampler, which is a fast (10x faster in this case; 100 sampling steps vs. 1000) sampler proposed by another work Denoising Diffusion Implicit Models (DDIM) ¹. If you would like to reproduce the results of DDPM reported in this repo, please consider removing the use-ddim flag.

By the way, your DDIM result ($\approx$ 4.35) is actually within the reasonable range from the reported CIFAR-10 FID (with 100-steps and quadratic skips) of the DDIM paper. See below

and also the author's reply (implying a FID of 4.26 with a different seed in the same setting) on GitHub
https://github.com/ermongroup/ddim/issues/3#issuecomment-960431667

from ddpm-torch.

Thank you for your clarification!

Edit: I reran with DDPM at 1000 subseq-size and got an FID of ~3.2 as reported in the table

from ddpm-torch.