Models

GAN (Generative Adversarial Network)

A generative model architecture where two networks — a generator and discriminator — compete to produce realistic synthetic data.

Definition

Introduced by Ian Goodfellow in 2014, GANs consist of two neural networks trained in opposition: a Generator that creates synthetic data, and a Discriminator that tries to distinguish real from fake. The generator learns to fool the discriminator; the discriminator learns to catch fakes. This adversarial game drives both networks toward high-quality synthesis.

GANs were the dominant image generation approach from ~2015–2022, producing photorealistic faces (StyleGAN), deepfakes, data augmentation, and image-to-image translation (pix2pix, CycleGAN). They are difficult to train (mode collapse, training instability) and have largely been superseded by diffusion models for most tasks.

GANs remain useful for real-time inference (they generate in a single forward pass vs. iterative diffusion) and medical imaging where training data is limited.

Examples

  • StyleGAN (face synthesis)
  • CycleGAN (image translation)
  • BigGAN
  • Deepfake tools