Unpaired image-to-image translation, exemplified by CycleGAN, enables the transformation of images between different styles or domains without the need for paired input-output data. Unlike traditional methods like Pix2Pix, which require corresponding pairs, CycleGAN employs two Generative Adversarial Networks (GANs) with cycle consistency loss to maintain the original content while altering the style. Each GAN features a generator for image translation and a discriminator for distinguishing real from fake images, often using PatchGANs for fine detail assessment.

The cycle consistency loss ensures that translating an image from one domain to another and back retains the original image, promoting content preservation. CycleGAN uses Least Squares loss for adversarial training to enhance stability and avoid vanishing gradients. An optional identity loss can also be included to preserve color consistency. Overall, CycleGAN significantly advances unpaired image translation by effectively learning to separate content from style.