Weird edge pattern in generated images during training

[manuelmaior29]

So I am trying to perform a training of the CUT model in order to translate images captured from CARLA Simulator domain to Cityscapes images domain. During training I observed that many of the Fake_B generated images have some weird edges forming at the transition from a ground object to the sky.


Example 1:

Example 2:

Some training details:

• 3000 examples from CARLA Simulator domain and 4376 examples from Cityscapes domain.
• Parameters --CUT_mode CUT --n_epochs 25 --n_epochs_decay 0 --batch_size 1 --preprocess scale_width_and_crop --load_size 1024 --crop_size 512

Is there something related to the training setup that I should consider doing? Should I just wait for more training iterations? Also I noticed that the losses, after the second epoch, started not to really decrease anymore.

Thank you!