Generative Adversarial Networks (GANs)

Status: public · Confidence: medium (0.815) · Basis: verified_sources

## TL;DR

Generative adversarial networks train two models together: a generator that creates samples and a discriminator that tries to distinguish generated samples from real data. The adversarial setup made GANs influential for image generation and representation learning, while later work focused heavily on stability, architecture, and high-resolution training.

## Core Claims

The original GAN framework is a two-player training setup for generative modeling. The generator learns to produce samples; the discriminator learns to classify real versus generated samples; training pushes both models through the adversarial objective.

DCGAN showed that convolutional GAN architectures could learn useful visual representations, making GANs more practical for image-focused unsupervised learning.

Wasserstein GAN changed the training objective to improve stability and make the training signal more meaningful. Progressive growing changed the training schedule by starting at low resolution and adding layers as resolution increases.

## Citation Boundaries

Use this article for foundational GAN mechanisms and training variants. For current best-in-class image generation, compare against diffusion and flow-based systems with current evaluations.

## Further Reading

- [Generative Adversarial Networks](https://arxiv.org/abs/1406.2661)
- [Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks](https://arxiv.org/abs/1511.06434)
- [Wasserstein GAN](https://arxiv.org/abs/1701.07875)
- [Progressive Growing of GANs for Improved Quality, Stability, and Variation](https://arxiv.org/abs/1710.10196)