Interestingly, in lpips module, they are just normalized to [-1, 1].

Hi @borisdayma!
When you are using the pretrained weights (pretrained='imagenet'), the input should be between 0 and 1.
The torchvision normalization is applied to the input in the __call__ method so you don't have to do that, see here.
I am considering to add an additional argument normalization so that people can use the imagenet weights without using the torchvision normalization, what do you think about that?

For the training code I used the range [-1, 1] because that range has worked better for me in the past.

Yeah, for lpips the input is [-1, 1] and then the input is normalized using mean=[-.030,-.088,-.188] and std=[.458,.448,.450].

Thanks, this is much clearer.
I think it's a good idea to add the argument normalize as a parameter.

Thanks for the feedback! I added the argument:

import flaxmodels as fm

vgg16 = fm.VGG16(output='logits', pretrained='imagenet', normalize=False)

This way the imagenet weights are used but the images are not normalized internally.