Generative adversarial network
In short
A generative adversarial network (GAN) is an AI system made of two neural networks that compete: one generates fake data, such as faces, and the other tries to tell the fakes from real examples. Each improves by trying to beat the other.
- Also called
- GAN, GANs, Adversarial network
- Topic
- How AI models work
- Level
- Some background helps
- First described
- Ian Goodfellow and colleagues at the Université de Montréal, June 2014, in "Generative Adversarial Networks"1.
Ian Goodfellow and colleagues at the Université de Montréal introduced GANs in 20141. A generator tries to capture what the training data looks like, while a discriminator estimates whether a sample came from the real data or from the generator1. The generator is trained to make the discriminator wrong1.
The paper's own analogy is counterfeiters and police: competition drives both teams to improve until the counterfeits cannot be told apart from genuine currency1. In theory, training ends with the generator reproducing the real data and the discriminator reduced to a coin toss, scoring one half everywhere1.
GANs can generate convincing photographs of things that do not exist. Nvidia's StyleGAN, published in 2018, learned to separate high-level features such as pose and identity from fine details such as freckles and hair, and came with a new, high-quality data set of human faces2.
Diffusion models have since matched GANs on image quality: in 2020 Ho and colleagues reported a diffusion model with sample quality similar to ProgressiveGAN on one benchmark3. GAN ideas still appear inside newer tools. The High Court judgment in Getty Images v Stability AI notes that SD XL Turbo uses adversarial diffusion distillation, which combines aspects of GANs and diffusion models to create high-quality images quickly4.
Adversarial here refers to the two networks training against each other. It is a different idea from adversarial attacks, which try to trick a finished model.
Example
A GAN trained on photos of faces learns to produce faces of people who do not exist. StyleGAN, trained on human faces, lets users vary pose and identity separately from details such as freckles and hair2.
Why it matters
GANs showed early that AI could produce convincing synthetic faces, and their adversarial trick still appears inside newer image generators4.
Questions people ask
Who invented GANs?
Ian Goodfellow and seven co-authors at the Université de Montréal, including Yoshua Bengio, in a paper posted in June 20141.
What is the difference between a GAN and a diffusion model?
A GAN pits a generator against a discriminator and makes an image in one pass1. A diffusion model starts from noise and removes it over many steps, 1,000 in the original DDPM setup3.
Are GANs still used?
Less often as the main image generator, but their ideas persist. The High Court noted in 2025 that Stability AI's SD XL Turbo combines aspects of GANs and diffusion models to generate images quickly4.
Related terms
Sources
- Goodfellow et al. (2014), Generative Adversarial Networks, arXiv:1406.2661
- Karras, Laine and Aila (2018), A Style-Based Generator Architecture for Generative Adversarial Networks, arXiv:1812.04948
- Ho, Jain and Abbeel (2020), Denoising Diffusion Probabilistic Models, arXiv:2006.11239
- Getty Images (US) Inc v Stability AI Ltd [2025] EWHC 2863 (Ch), paras 4, 6, 41, 145, 554-558 and 600