Mirza et al. in their paper, Conditional Generative Adversarial Nets, introduced a conditional version of the GAN framework. This modification is extremely easy to understand and is the foundation of amazing GAN applications that are widely used in today's world.
Some of the most astonishing GAN applications, such as the generation of a street scene from a semantic label to the colorization of an image given a grayscale input, pass through image super-resolution as specialized versions of the conditional GAN idea.
Conditional GANs are based on the idea that GANs can be extended to a conditional model if both G and D are conditioned on some additional information, y. This additional information can be any kind of additional information, from class labels to semantic maps, or data from other modalities. It is possible to perform this conditioning by feeding...