Currently our implementation of FID/256 relies on the torchmetrics FID, which includes image interpolation. For images of size 256, this doesn't cause a huge difference, but many standard methods don't include interpolation in calculating FID.
So, we should probably have an FID metric that has no interpolation and passes the images straight to the Inception model.
In my past experiments, this could cause a divergence of maybe 0.01-0.1 in the FID score on CLIC2020.
Currently our implementation of FID/256 relies on the
torchmetricsFID, which includes image interpolation. For images of size 256, this doesn't cause a huge difference, but many standard methods don't include interpolation in calculating FID.So, we should probably have an FID metric that has no interpolation and passes the images straight to the Inception model.
In my past experiments, this could cause a divergence of maybe 0.01-0.1 in the FID score on CLIC2020.