Using GANs for style transfer

So far, we have only looked at DCGANs in detail. Hundreds of different types of GAN models exist already, and many more are in the making. Each of these GAN variants differs by either the application they are catering to, their underlying model architecture, or due to some tweaks in their optimization strategy, such as modifying the loss function. For example, Super-Resolution GAN (SRGAN) are used to enhance the resolution of a low-resolution image. The CycleGAN uses two generators instead of one, and the generators consist of ResNet-like blocks. The Least Squares GAN (LSGAN) uses the mean square error as the discriminator loss function instead of the usual cross-entropy loss used in most GANs.

It is impossible to discuss all of these GAN variants in a single chapter or even a book. However, in this section, we will explore one more type of GAN model that relates to both the DCGAN model discussed in the previous section and the neural style transfer...