Models
Generative Adversarial NetworkGAN
Also known as: adversarial network
A generative model made of two networks competing, a generator that fakes data and a critic that tries to catch it, which together learn to produce realistic output.
A GAN is a generative AI built as a contest between two networks. One (the "generator") tries to produce fake data, say, realistic faces, while the other (the "critic") tries to tell fakes from real ones. As they compete, the generator gets better and better at fooling the critic, until its fakes look convincingly real. GANs powered an earlier wave of AI image generation and early deepfakes.
A generative adversarial network (GAN) pairs two neural networks trained against each other: - Generator: turns random noise into fake samples (e.g. images). - Discriminator: judges whether a sample is real (from the dataset) or fake (from the generator).
Training is a game: the generator improves to fool the discriminator, the discriminator improves to catch it. At equilibrium the generator produces realistic data.
Legacy and limits
GANs drove major progress in image generation (2014 onward) and early deepfakes. They can be unstable to train and prone to "mode collapse" (producing limited variety). For frontier image generation they have largely been overtaken by diffusion models, but the adversarial idea remains influential.
GANs frame generation as a minimax game between a generator G and discriminator D, with G minimising and D maximising a value function such that G learns to match the data distribution while D estimates the probability a sample is real (Goodfellow et al. 2014). At the theoretical optimum G reproduces the data distribution.
Practically, adversarial training is notoriously unstable, mode collapse, non-convergence, vanishing gradients, spawning many stabilisations (WGAN, spectral normalisation, progressive growing, StyleGAN). GANs offer fast single-step sampling and sharp outputs but weaker mode coverage and likelihood guarantees than alternatives. They dominated high-fidelity image synthesis until diffusion models surpassed them on quality/diversity and training stability. The adversarial principle persists in evaluation, domain adaptation, and as a component in hybrid systems, and GANs contributed to the early deepfake landscape.