Is it better to use BatchNorm or GroupNorm in the generator of my conditional WGAN-GP model? Although BatchNorm gave me good results, my small batch size of 10 meant the eval mode gives bad images. I decided to use GroupNorm, and it allowed me to use eval mode. Yet GroupNorm caused my generator outputs to look very similar even with very different inputs. Is this a normal problem with GroupNorm and how do I fix this?