Questions regarding evaluation

[Seth-Park]

Hi,

Thanks for sharing the codebase and the pretrained models! Your work is very impressive :)
I just wanted to ask a few questions regarding quantitative evaluation:

1. How many images were used for computing the FID scores in Table 1 of the main paper?
2. How are the reference images that guide the generation process chosen?
3. What codebase did you use to compute the FID scores?

Regards,
Seth

[jychoi118]

Hi, thank you for enjoying our paper.

1. We used 50k generated samples and 50k random real images from the training set. (Generating 50k images with full 1000 steps takes a long time..)
2. For FID score, 50k reference images are randomly chosen from the training set.
3. We followed https://github.com/mseitzer/pytorch-fid for FID scores. I would like to note that our experiments in the paper are conducted with this codebase: https://github.com/rosinality/denoising-diffusion-pytorch . So, we measured FID scores with data pre-processing from this codebase.