I’m working on a research project using GANs and I’m trying to get a baseline by training a basic DCGAN on the CelebA dataset. I know we’re spoiled by StyleGAN and BigGAN, but without any esoteric optimizations the results are already looking decent by the 7th epoch or so:
However, as training goes on the image diversity and fidelity improves but the image contrast decreases substantially:
This is a very strange pathology because you would expect the discriminator to pick up on something like this and immediately exploit it. Any ideas?