InfoGANs (short for Information Maximizing Generative Adversarial Networks) are somewhat similar to CGANs in the sense that both generator networks take in an additional parameter and the conditional variable, c, such as label information. They both try to learn the same conditional distribution, P(X |z, c). InfoGANs differ from CGANs in the way they treat the conditional variable.
CGANs consider that the conditional variable is known. Hence, the conditional variable is explicitly fed into the discriminator during training. On the contrary, InfoGANs assume that the conditional variable is unknown and latent, which we need to infer based on the training data. The discriminator in an InfoGAN is responsible for deriving the posterior, P(c |X). The architecture of an InfoGAN is presented in the following diagram:
Since we do not need to supply the conditional...