[Sol-]

As amazingly plausible the pictures look, I personally have some dislike towards such applications (nothing against the author of course, just about ML in general) because I always feel a bit as if I'm being duped by the neural net. When I see image restoration, I'm subconsciously expecting historical fidelity even if I'm just marveling at the nice colorization. But of course such historical accuracy is not the primary goal of the GAN.

Maybe another cool avenue to explore would be combining models like this with some NLP approach that parses a historian's rough description of how the scene should be colored and biases the generator with prior information that way. (Maybe related to visual question answering or something.)