EXPLAINER

Generative Adversarial Networks (GANs)

conceptfreshness: stablesources not yet verified

GANs train a generator and a discriminator against each other so the generator learns to produce realistic samples.

Short answer: A GAN pits a generator (makes fakes) against a discriminator (spots fakes). It yields sharp samples in one pass but training can be unstable and prone to mode collapse; diffusion has taken over many tasks.

What is it?

Two networks play a game: a forger tries to produce convincing fakes while a detective tries to tell fakes from real examples. Each gets better because of the other, until the forger's output is realistic.

Also known as: GAN, GANs, generator discriminator, StyleGAN, mode collapse, adversarial training, deepfake generator.

Technical definition

Goodfellow et al. (2014) formulate a minimax game between G and D. Practical issues: non-convergence, mode collapse, vanishing gradients; mitigations include Wasserstein losses, spectral normalisation and progressive growing (StyleGAN family). GANs generate in a single forward pass (fast sampling), whereas Diffusion Models now dominate many text-conditioned generation tasks.

Entity type: concept · Freshness class: stable.

When should I use it?

Consider for real-time or low-latency generation, super-resolution and legacy pipelines.

Implementation patterns

  • Conditional GANs for class- or image-conditioned generation.
  • Spectral normalisation and two-timescale updates for stability.
  • Use for fast real-time generation or super-resolution.

Technologies and frameworks

Commonly associated: StyleGAN, PyTorch, Wasserstein GAN. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

  1. Check whether diffusion fits better.
  2. Start from a published architecture.
  3. Track FID and visual samples.
  4. Plan misuse mitigations and provenance.

Common failure modes

  • Mode collapse.
  • Training oscillation or divergence.
  • Hard-to-evaluate quality (FID has caveats).
  • Misuse for synthetic media and fraud.

Troubleshooting

  • Generator outputs lack variety — Use stabilised losses, minibatch discrimination, or move to a diffusion model.
  • Training diverges — Lower learning rates, balance G/D updates, use proven architectures.

Contrasted with

Questions this page answers

  • What is a GAN?
  • What is mode collapse?
  • GAN vs diffusion model: what is the difference?
  • How are deepfakes made?

Typical goals: understand GANs and their limits; compare GANs with diffusion models.

Sources and verification

Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.

Freshness class: stable — Core ideas that change slowly (maths, classic algorithms). Re-review yearly.

Canonical sources (unverified leads):

  • Goodfellow et al., Generative Adversarial Nets (2014) — arXiv:1406.2661
  • Karras et al., A Style-Based Generator Architecture for GANs (2018) — arXiv:1812.04948