At the beginning of the 2018, the Berkeley AI research lab published a paper entitled Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks (arXiv link: https://arxiv.org/pdf/1703.10593.pdf). This paper is special not only because it proposed a new architecture, CycleGAN, with improved stability, but also because they demonstrated that such an architecture can be used for complex image transformations. The following diagram shows the architecture of a cycle GAN; the two sections highlight the Generator and Discriminators playing a role in calculating the two adversarial losses:
![](/api/v2/epubs/9781788836067/files/assets/d5036aa6-77cf-41c1-8828-11436977198e.png)