GAN (Generative Adversarial Network)
Two networks trained against each other — one faking, one detecting — until the fakes pass. The technique diffusion largely replaced.
When not to use it
- For general image generation today. Diffusion models are better, more diverse, and vastly less painful to train. Choose a GAN only for a specific reason.
- When you need output diversity. Mode collapse is not an edge case; it's the characteristic failure, and it produces confident sameness.
- When you can't evaluate by looking. There's no loss value that means "good," and the automated metrics are proxies you shouldn't trust alone.
Reach for something else instead
- Diffusion models — the default for image generation now: stable training, better coverage of the data distribution.
- VAEs when you want a well-behaved latent space and can accept blurrier output.
- Distilled diffusion if what you actually wanted was GAN-like speed with diffusion quality.
Read more on the blog
- What is generative AI? The complete guideGenerative AI is the technology behind the chatbots, image makers, and video tools of the last few years, and it is usually explained as if each were a separate trick. Underneath, they share one idea: learn the probability distribution of some kind of data, then sample from it to make new examples. This guide explains what generative AI is, how it works, the model families, and where it really falls short.
- How AI generates images: from noise to a pictureType a sentence, get an image that never existed. The technology behind it, diffusion, is one of the most simple ideas in modern AI: teach a model to remove noise, then hand it pure static and let it sculpt. Here's how it actually works, how text steers it, and the 2026 rivalry reshaping the field.
A GAN pits a generator against a discriminator until fakes look real.
Two networks train in opposition: the generator makes fake samples from noise, and the discriminator tries to tell real from fake. Each pushes the other — the generator improves to fool the discriminator, the discriminator improves to catch it — and at equilibrium the fakes are indistinguishable from real data. That adversarial game is what produced the first convincingly realistic AI-generated images.
Further reading
- Goodfellow et al. (2014), Generative Adversarial Nets — the original, and unusually readable.
- Karras et al. (2018), A Style-Based Generator Architecture for Generative Adversarial Networks — StyleGAN, the peak of GAN image quality.
- Arjovsky, Chintala & Bottou (2017), Wasserstein GAN — the most influential attempt to make training stable, and a clear account of why it wasn't.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reading a falling generator loss as progress. In an adversarial game, loss values are relative to an opponent that's also moving. They mean much less than they appear to.
- Fighting mode collapse with more training. It's a failure of the objective, not of patience.
- Trusting FID as ground truth. It's sensitive to implementation details and rewards things human viewers don't care about.
At a glance
Often compared with
Where this sits
11 concepts come first. Understanding it opens up 1 more.
Computed from the prerequisite graph, not assigned. How this works