?Since a GAN network can distinguish between real and fake images. It ?should be able to tell and improve a black and white photo to color photo.
I was thinking you use a pair of images. Take a BW version of the image and you get a color BW pair. Then somehow you generate a color version of the target BW photo and let a discriminator tell which of the pair is fake or real.
The idea is that this will result in a photo realistic color image of the BW photo.