I think if you trained it on a few hundred more images the results would look much better, though I understand that's a lot of effort. 400 images is very little for a network like this and it's probably why the images have so many artifacts.
Overall a very impressive project though, making it run realtime must have been a challenge. Respect.