What and how to pretend these artifacts in Training GAN

Question

currently I am training 3D Patches as you can see in the picture. At the moment I wonder where these artifacts on the edges came from? My intuition says that it could be the deconvolution, but I don't know how to prevent it.

one middle Slice of the 3D Patch

Answer 1

from my experience this is from your padding layers. I have seen similar effects during my projects. Deconvolutions might result in checkerboard effects all over the image.

I replaced zero padding layers with reflection padding layers. When padding with zeros, it usually has a large impact of the distribution of your layer's values. This impact is reduced when padding using values from the same feature map.