amige.
blog & research · glossary · GAN

what’s a GAN?

by Tuan Hoang · detection lead · last reviewed 2026-05-15
⤹ not extinct. out of fashion.

A GAN (generative adversarial network) is an architecture where two neural networks compete: a generator producing synthetic data, a discriminator trying to tell real from fake.

GANs were introduced by Ian Goodfellow and collaborators in 2014 and dominated AI image generation for roughly the next seven years. the setup is a game: a generator network takes random noise and tries to produce images that look real; a discriminator network looks at images and tries to classify them as real (from the training set) or fake (from the generator). each network's loss feeds the other's improvement.

GENERATOR VS DISCRIMINATORrandom noisegeneratorreal imagefakediscriminatorreal? or fake?each mistake trains the other. ↑
a GAN: two networks locked in a forgery contest

at training equilibrium, the generator produces images the discriminator can no longer reliably tell from real data. that adversarial pressure is what made GANs so good at narrow tasks like face generation. NVIDIA's StyleGAN family (2018-2021) produced the synthetic faces of thispersondoesnotexist.com and powered the first wave of consumer-facing deepfake apps.

diffusion models have displaced GANs for general-purpose image generation. GAN training is unstable and its output diversity is narrower than diffusion on the same dataset. face-swap deepfake video in the wild still runs on GAN-derived pipelines, and dedicated face-synthesis applications ship GAN backbones for their speed and identity-fidelity tradeoffs.

for detection: GAN outputs leave characteristic frequency-domain fingerprints from their upsampling layers. checkerboard artifacts, periodic spectral peaks, and noise-residual signatures all betray a GAN at the pixel level, often after compression and re-encoding. detectors trained on these signatures still catch modern face-swap content.

most in-the-wild deepfake face-swap video still uses GAN-based pipelines, even where state-of-the-art new generators are diffusion. a detection panel that drops GAN coverage misses that category.

questions

a GAN, or generative adversarial network, runs two neural networks against each other. a generator builds synthetic images from random noise, and a discriminator tries to separate real images from fake ones. each network’s mistakes train the other. at equilibrium the generator makes images the discriminator can no longer tell apart from real data.

diffusion models have taken over general-purpose image generation, since GAN training is unstable and its output variety is narrower. GANs still matter in one corner. most face-swap deepfake video in circulation runs on GAN-derived pipelines, and face-synthesis apps keep GAN backbones for their speed and identity accuracy.

both produce synthetic images through different machinery. a GAN trains two networks in one competitive game, while a diffusion model learns to reverse a step-by-step noising process. diffusion gives more varied output and now dominates new general-purpose generators, though GANs stay common in face-swap work. the two leave different detection fingerprints, so a detector built for one can miss the other.

GAN outputs carry frequency-domain fingerprints from their upsampling layers, including checkerboard artifacts, periodic spectral peaks, and noise-residual signatures that often survive compression and re-encoding. detectors trained on these traces still catch modern face-swap content. a detection panel that drops GAN coverage misses a category that still circulates widely.

sources.

  1. 01
    Goodfellow et al., Generative Adversarial Networks (arXiv:1406.2661)
    The foundational 2014 paper that introduced the architecture.
  2. 02
    Karras et al., StyleGAN — A Style-Based Generator Architecture (arXiv:1812.04948)
    The face-generation result that powered thispersondoesnotexist.com.
  3. 03
related terms
put one through amige →is this AI? →