A Generative Adversarial Network (GAN) trains two neural networks against each other: a generator that creates synthetic images and a discriminator that tries to tell real images from fakes. The generator improves by fooling the discriminator, and the discriminator improves by catching fakes. This adversarial training loop pushes both networks to get better, until the generator produces images realistic enough that the discriminator can't reliably distinguish them from real photos.

GAN variants serve different purposes. StyleGAN and StyleGAN3 produce photorealistic faces and objects with controllable attributes (pose, lighting, expression). Pix2pix and CycleGAN handle paired and unpaired image-to-image translation (satellite to map, sketch to photo, day to night). SRGAN and ESRGAN perform super-resolution, generating sharp high-resolution images from low-resolution inputs. ProGAN and BigGAN scale generation to high resolutions through progressive training.

In computer vision pipelines, GANs are used for synthetic training data generation (creating labeled examples of rare classes or scenarios), data augmentation (generating variations of existing training samples), domain adaptation (translating images from one visual domain to another while preserving labels), and image editing (inpainting, style transfer). Diffusion models have overtaken GANs for general image generation quality since 2022, but GANs remain faster at inference and are still preferred for real-time applications.