The UNSB project is quite impressive. During the actual code execution process, we used the code for underwater image translation experiments and noticed that several results exhibited flaws resembling "holes." Could you suggest possible improvements to address this issue? Thank you.
We have also experienced similar issues in natural scene dataset.
We do not know the exact reason for the artifact, but we can recommend several solutions
Longer training time
More downsampling & upsampling layer : If the network have to make large changes, using the bottleneck layer with lower resolution sometimes makes better results. As our basic network uses 2 down and upsampling layers, this might make some artifacts
Using different weight for regularization loss
Use patching sampling on vgg features instead of direct network features.
DZRRRRRR
commented
10 months ago 
cyclomon