In this chapter, we are going to generate handbags from sketched outlines without using explicit pair matching using DiscoGAN. We denote the sketch images as belonging to domain A, while the natural handbag images to belong to domain B. There will be two generators: one that takes the images of domain A and maps them to images that would look realistic under domain B, and another that does the opposite: one that maps handbag images from domain B to images that will look realistic under domain A. The discriminators would try to identify the generator generated fake images from those of the authentic images in each domain. The generators and the discriminator would play a minimax zero- sum game against each other.
To train this network, we will require two sets of images, sketches, or outlines of handbags and natural images...