## TL;DR

GANs (Goodfellow, 2014) pit two neural networks against each other: Generator creates fake data, Discriminator distinguishes real from fake. Both improve through adversarial training — Generator learns to produce increasingly realistic outputs. Used for image generation, style transfer, super-resolution.

## Core Explanation

Training instability: if discriminator is too strong, generator learns nothing. Mode collapse: generator produces limited variety. Key variants: DCGAN (deep convolutional), StyleGAN (progressive growing, controls style), CycleGAN (unpaired image translation), Pix2Pix (paired). GANs largely superseded by diffusion models (2022+) for image generation, but remain important for other adversarial tasks.

## Further Reading

- [Generative Adversarial Networks (Goodfellow et al., 2014)](https://arxiv.org/abs/1406.2661)