CycleGAN is an approach for training generative adversarial networks (GANs) that can transform images as if they are rotating one type to another. This approach was first proposed in 2017 by Jun-Yan Zhu, et al. and demonstrated on a wide range of image-to-image translation tasks, such as converting photos of apples to oranges, summer to winter scenes, or horses to zebras.
This approach works by training two separate GANs, or generators, upfront: the first transforms images from one domain into the other; then, the second transforms those new images back into the original domain. This back-and-forth approach allows the network to learn from both domains, making it better at translating between them.
The CycleGAN approach has seen widespread success due to its ability to learn to translate between two different domains without the need for any labeled data. This means that users don’t need to annotate data sets, as would be required for supervised methods. Additionally, the two separate generators allow for improved image quality compared to single-generator GANs.
In the years since its introduction, CycleGAN has been used to produce a wide range of applications, from automatic colorization of grayscale photos to synthesizing convincing deepfakes of real people. As GANs become increasingly powerful, CycleGAN will remain an important tool in the development of cutting-edge image-processing technologies.