CycleGAN

CycleGAN is a framework for unpaired image-to-image translation introduced by Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros in 2017. It enables learning mappings between two visual domains without the need for paired training examples, making it suitable for tasks where corresponding images are difficult to obtain.

The model comprises two generators and two discriminators. Let X and Y be the two domains. G maps X to Y, F maps Y to X, D_Y discriminates real from translated images in Y, and D_X does the same in X. During training, each generator is trained with an adversarial loss to produce realistic images in the target domain, while each discriminator learns to distinguish real from fake samples. A key feature is the cycle-consistency loss: for x in X, F(G(x)) should be close to x, and for y in Y, G(F(y)) should be close to y. This cycle loss is combined with the adversarial losses to form the overall objective. An optional identity loss can be used to encourage color and intensity preservation when the input appears already in the target domain.

CycleGAN typically employs convolutional neural networks with PatchGAN discriminators and can be trained with variants of GAN objectives. It has been applied to tasks such as style transfer between photos and paintings, season or weather changes in imagery, and unpaired translations like maps to aerial photographs or horses to zebras.

Limitations include potential artifacts, imperfect preservation of content, and reliance on sufficient overlap between domains. The approach remains influential, establishing a widely used baseline for unpaired image-to-image translation and inspiring numerous extensions and related models.