The problem is that the machines are not learning. They are taking a statistical amalgamation of the input images and outputting a selected part of each relevant input (not how people learn or make inferences BTW)

It’s more like a complex selective compression and decompression method than generation of a new image.