Standard Generative Adversarial Networks produce samples reflecting the training data distribution, but offer limited direct control over the specific output. This chapter introduces techniques for guiding the generation process. You will learn to implement Conditional GANs (cGANs), which generate data matching specific attributes or labels, often denoted as $y$. We cover how to integrate this conditional information $y$ into both the generator $G(z, y)$ and discriminator $D(x, y)$.

We then study Information Maximizing GANs (InfoGAN), a method for learning interpretable latent codes $c$ in an unsupervised fashion by maximizing mutual information $I(c; G(z, c))$ between these codes and the generator output. The chapter also addresses methods for manipulating the learned latent space $z$ to modify output characteristics post-generation and discusses approaches like StackGAN for tasks such as text-to-image synthesis. Finally, we consider the concept and measurement of disentanglement in learned representations. Hands-on sections provide practice in building these conditional models.