Improved version with edited out watermarks and more focus on fdg in training data, should give better results.
Trained on 832 resolution, higher initial resolution e.g. 768 x 1024 should give better results.
Too lazy to test, I'll just drop 1 image :3
