This chapter transitions from analyzing image content to synthesizing new images using Generative Adversarial Networks (GANs). GANs provide a framework for training models to produce realistic outputs, often images, through a competitive process between two networks: a generator $G$ that creates data and a discriminator $D$ that tries to distinguish real data from generated data.

We will begin by reviewing the fundamental concepts behind GANs and their adversarial objective function. We then address common training challenges like mode collapse and instability, along with potential solutions. You will study important GAN architectures, including Deep Convolutional GANs (DCGAN), Conditional GANs (cGAN) for controlled generation, and the style-based approach of StyleGAN. Methods for evaluating the quality and diversity of generated images, such as the Fréchet Inception Distance (FID) and Inception Score (IS), will also be covered. The chapter includes a practical exercise on implementing a DCGAN.