Why different fid_stats are available for ImageNet dataset for different resolutions? How to generate fid_stats for Celeba dataset for resolutions other than 64x64?

[LakshmikarPolamreddy]

When I computed FID value for generated images of 256x256 resolution corresponding CelebA dataset, using fid_stats_celeba64_train_50000_ddim.npz, I am getting very high value (85.56). How to generate fid_stats for different resolutions? In general, fid_stats are pre-computed based on real images, right.

[forever208]

@LakshmikarPolamreddy long story short, using pytorch_fid repo