@genekogan Have you tried optimizing Z? I gave this a try, but did not get good results. Previously, I was able to successfully optimize the input noise vector for ProgGan and get a close reproduction, so I'm curious if there is something about BigGAN that makes this less successful.