Some questions about your code

[vtddggg]

Thank you for sharing!!

I found that you get color_image through bluring :
batch_colors = np.array([self.imageblur(ba,True) for ba in batch]) / 255.0

Bluring is a good way to get color prior when we have groundtruth images. But if I only have a line map (without groundtruth), bluring can not work because bluring a line image can not get any color informations. As your paper said, a color_predict network can predict the color from a line image. I think it is a very nice idea while I found main.py does not realize this part of functions. Maybe you have tried in guess_color.py.

Good job!!

[vtddggg]

I have understood what your code mean. It seems that a Variational Autoencoder is trained to guess the color information.

[zjgirl]

Do you really think the code matchs with the paper? I think the code doesn't contain two networks (although the two networks are the same). It seems it achieves the direct network which has L1 loss and adversial loss. But it doesn't achieve what the paper says.