Have you tried with your FullGenerator_SR model?

[anguoyang]

I found there is source code on FullGenerator_SR for those in_size!=out_size cases, have you trained on this model? for GPEN generator, the input/latent is [N, 512], and output is [N, 3, 512, 512], just curious how you design your mapping network for 4x SR for any input image size, for example, if the input image is 1024x768, and I want 4XSR to 4096x3072, if with GPEN model, the input image have to be resized to 512x512, then some information maybe lost.
BTW stylegan2/gpen is much clear than esrgan although with the same resolution, esrgan tend to smooth the image, especially, loss on some key details, so I don't want to use it, I want to do the 4xSR with stylegan2/gpen only