The generative model introduced in the previous chapter essentially defines a data probability distribution function. Probability distribution is not the only way to define a generative model. We describe generative adversarial networks (GAN) (Goodfellow et al., Generative adversarial nets. In: Advances in neural information processing systems. Curran Associates, 2014) in this chapter. It is defined through a data generation process, rather than a probability distribution. The motivation is it is often difficult to estimate the probability distribution of data, and even the likelihood function cannot be explicitly expressed. Defining the data generation process and its constraints directly does not depend on the specific likelihood function formula therefore it provides an alternative formulation of generative models. We will further describe an extension, Wasserstein GAN (WGAN).