I am dealing with some tiling artifacts which appear after the first epoch in an image-to-image translation task. I use a patch gan (70x70) as proposed in CycleGAN. Furthermore, as I read in the paper: https://arxiv.org/pdf/1611.07004.pdf, these artifacts come up when a vanilla discriminator or a one with a smaller receptive field is used. As epochs grow, this phenomenon is getting gradually worse.
Any ideas would be appreciated.
