Generative Adversarial Network (GAN)
A Generative Adversarial Network (GAN) is a dual-network architecture where a generator produces synthetic data and a discriminator attempts to classify it as real or fake; adversarial training between the two drives progressively higher output quality.
A Generative Adversarial Network (GAN) is a class of neural network architecture introduced by Ian Goodfellow and colleagues in 2014. It consists of two networks trained in opposition: a generator G that maps random noise vectors to synthetic data samples, and a discriminator D that outputs a probability estimate of whether a given sample is real or generated. Neither network has direct access to the other's parameters; each learns only through the feedback of the adversarial loss.
Training alternates between two objectives. The discriminator is updated to maximize its ability to distinguish real training samples from generated ones. The generator is simultaneously updated to minimize the probability that the discriminator correctly identifies its outputs as fake — learning to fool D. At theoretical Nash equilibrium, G produces samples indistinguishable from the training distribution and D outputs 0.5 for all inputs. In practice, stabilizing this minimax game requires techniques such as Wasserstein loss with gradient penalty (WGAN-GP), spectral normalization, and progressive growing of resolution, as in NVIDIA's ProGAN and StyleGAN series.
GANs established that a model could learn complex, high-dimensional data distributions without explicit likelihood estimation, enabling photorealistic face synthesis, image-to-image translation (Pix2Pix, CycleGAN, 2017), super-resolution, and early video generation. NVIDIA's StyleGAN2 (2020) produced human face images largely indistinguishable from photographs at high resolution, driving both creative applications and concern about synthetic media misuse.
By the mid-2020s, diffusion models displaced GANs as the dominant paradigm for image and video generation, offering more stable training and greater output diversity. GANs remain in active use where inference latency is critical — diffusion models are substantially slower — and for data augmentation in medical imaging, where synthetic samples supplement scarce annotated datasets. StyleGAN-based architectures continue to appear in avatar generation and artistic tools.