One thing very noticeable in both works is also the tendency to overshoot borders when coloring. In many images it can be seen that the color of the main subject bleeds into the background. A good example is e.g. the fifth picture in http://richzhang.github.io/colorization/resources/images/exs... , where parts of the sky get assigned the same color as the mesa. This looks like something that should be fixable if one tries to detect these kind of edges and gives less weight to information from the other side of such an edge.
Does anyone know what the differences between this and the previous neural net approach are on a technical level?