Generative adversarial networks

A training scheme was described in which a generator and a discriminator learn against each other.

Why it matters

GANs made realistic image generation a practical problem and set a research direction that preceded diffusion models.

Training such models is unstable, and much of the later work addresses exactly that problem.

Event date

June 10, 2014

Timeline date

Event date

Verification

Sources gathered automatically · September 16, 2026

Lines

ID

evt-0051

primary Generative Adversarial Networks
arXiv · Published June 10, 2014
primary Generative Adversarial Nets
Advances in Neural Information Processing Systems 27 (NIPS 2014), NeurIPS proceedings · Published 2014

Contrasts The variational autoencoder 2013links to this record
A different answer to the same generation problem, arriving six months later.
- Auto-Encoding Variational Bayes · arXiv

Related Diffusion models 2015
Both describe image generation learned from data; the training mechanisms differ.
Builds on pix2pix: one model for image translation 2016links to this record
pix2pix is a conditional adversarial network: the generator learns to fool the discriminator, as in the GAN of 2014, which the paper cites.
- Image-to-Image Translation with Conditional Adversarial Networks (arXiv:1611.07004v1) · arXiv
Builds on StyleGAN: a style-controlled face generator 2018links to this record
StyleGAN changes only the generator of an adversarial network, leaving the discriminator and the loss as in a GAN.
- A Style-Based Generator Architecture for Generative Adversarial Networks (arXiv:1812.04948v1) · arXiv
Contrasts Denoising diffusion probabilistic models 2020links to this record
A different route to the same task: sample quality is measured against adversarial networks, without adversarial training.
- Denoising Diffusion Probabilistic Models · arXiv