Generative models have spread widely over the last few years and are now at the cores of powerful softwares such as Dall-E or Midjourney. The Generative Adversarial Networks (GANs) issued first in 2014 are a class of such models, used to create synthetic images from scratch. GANs rely on two functions optimized to achieve two conflicting objectives; one function G is "trained" to represent the distribution of "realistic" images in order to produce such images, while the other function D is trained to efficiently distinguish between actual images and those produced by G. Simply put, G strives to fool D and D strives not to be fooled by G, which translates mathematically in a minimax objective, which we will analyse in the presentation.
