GAN

GAN (Generative Adversarial Network) is a type of generative machine learning model, introduced by Ian Goodfellow and colleagues in 2014, built from two neural networks set against each other. The generator starts from a random vector in latent space and tries to produce realistic fakes (say, images). The discriminator is shown a mix of real and fake samples and tries to tell them apart. They train as rivals (the generator keeps getting better at fooling the discriminator while the discriminator keeps getting better at catching it) until, ideally, the fakes are indistinguishable from the real thing. GANs drove an era of photorealistic face generation (StyleGAN) and image-to-image translation, but they’re notoriously finicky to train (a common failure is mode collapse, where the generator churns out the same few outputs). For general image generation they’ve largely been overtaken by diffusion models, though they still shine where fast, one-shot sampling matters.