Menu
GANsConcepts

Generative adversarial networks

3 min readintermediateUpdated 28 Sept 2026
1 · In one line

A GAN trains a generator to make samples and a discriminator to distinguish generated samples from training data.

1 · What it is

A GAN trains a generator to make samples and a discriminator to distinguish generated samples from training data. The generator maps a noise vector to a generated sample. The discriminator estimates whether an input came from the training data rather than the generator.

Training alternates discriminator updates with a generator update. The discriminator learns from training examples and samples from the generator. The generator is trained to increase the probability that the discriminator makes a mistake. The original paper describes this framework as a minimax two-player game.

A conditional GAN feeds extra information to both the generator and discriminator. DCGAN used four fractionally-strided convolutions to turn a representation into a 64 by 64 image. Wasserstein GAN was proposed in part to address learning instability and mode collapse. StyleGAN reported scale-specific control of image synthesis.

2 · Why it exists

The generator is trained to increase the probability that the discriminator makes a mistake.

No direct answerThe generator has no explicit representation of the probability distribution it is trying to reproduce.
Coupled trainingThe discriminator and generator must be kept in step during training.
Weak early signalThe original minimax objective may provide too little gradient for the generator early in learning.
3 · How it works

Follow one generated sample through the two-player game.

Generator and discriminator training loop Noise enters a generator to create a generated sample. A real example and the generated sample enter a discriminator. Its response to the generated sample feeds back to update the generator.
The discriminator learns from real and generated examples. Its response to a generated example supplies the generator's training signal.
  1. 1 · sampleDraw an input vector from a fixed noise distribution.
  2. 2 · generateThe generator maps that vector to a generated sample.
  3. 3 · judgeThe discriminator estimates whether each input came from the training data rather than the generator.
  4. 4 · alternateTraining alternates discriminator updates with a generator update.

The original paper describes this framework as a minimax two-player game.

4 · Where it's used
WhoWhat they askWhat it works with
Image researcher“Can the model synthesize another sample from this image distribution?”Noise vectors and training images
Conditional-generation team“Can the output be tied to a requested class?”Labels supplied to both networks
Representation researcher“Which high-level attributes can the synthesis process separate?”Intermediate controls in a trained generator
5 · What it solves, and what it doesn't
solves
  • After training, a generator can produce samples by forward propagation through the generative model.
  • A conditional GAN feeds extra information to both the generator and discriminator.
  • DCGAN used four fractionally-strided convolutions to turn a representation into a 64 by 64 image.
  • StyleGAN reported scale-specific control of image synthesis.
doesn't solve
  • The original GAN formulation has no explicit representation of the generator distribution.
  • The original paper says the discriminator and generator must be synchronized well during training.
  • Wasserstein GAN was proposed in part to address learning instability and mode collapse.
6 · Go deeper
Generative AINeural networksTrainingLoss functionComputer vision

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. paperGenerative Adversarial Nets, Goodfellow et al. · read 28 Sept 2026
  2. paperUnsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks, Radford, Metz and Chintala · read 28 Sept 2026
  3. paperWasserstein GAN, Arjovsky, Chintala and Bottou · read 28 Sept 2026
  4. paperA Style-Based Generator Architecture for Generative Adversarial Networks, Karras, Laine and Aila · read 28 Sept 2026
  5. paperConditional Generative Adversarial Nets, Mirza and Osindero · read 28 Sept 2026
Suggest a correction