Goodfellow et al. published Generative Adversarial Networks, formulating generative modeling as a two-player minimax game between a generator network and a discriminator trained end-to-end with Backpropagation—eliminating intractable likelihood bounds and Markov chain sampling.
Generative Adversarial Networks (GANs)
Also known as: GAN · GANs · StyleGAN · Wasserstein GAN · Adversarial Training
Plain English
A two-network generative architecture where a 'generator' creates fake images or videos and a 'discriminator' tries to spot them from real data—training both in a competitive game until outputs look realistic in a single pass.
Technical Overview
GANs pit a generator G(z) against a learned discriminator D(x) via an adversarial minimax or Wasserstein objective, enabling single-forward-pass synthesis without iterative denoising. Although Diffusion Models overtook pure GANs for broad text-to-image diversity after GigaGAN, stabilized R3GAN baselines (2025) and adversarial losses paired with score/flow Distillation & Synthetic Data (ADD, DMD2, Adversarial Post-Training, and Adversarial Flow Models) now power 1-to-4-step image and real-time Video Generation Models (Sora, Veo, Kling, Wan).
Timeline (6)
2014
Paper2017
PaperArjovsky, Chintala, and Bottou introduced Wasserstein GAN (WGAN), replacing Jensen–Shannon divergence with a continuous Earth Mover's critic objective that provides informative gradients even when generated and real distributions do not overlap, mitigating mode collapse and training instability.
2018
PaperKarras, Laine, and Aila at NVIDIA introduced StyleGAN, combining progressive convolutional growth with a latent mapping network and adaptive instance normalization (
AdaIN) across spatial scales—unlocking disentangled style control and photorealistic 1024×1024 face synthesis.2023
PaperKang et al. introduced GigaGAN, scaling a multi-scale StyleGAN-derived generator with cross-Attention to 1B parameters for text-to-image synthesis—producing 512px images in 0.13 seconds in one pass while highlighting the diversity tradeoffs that pushed frontier labs toward Diffusion Models.
2025
PaperHuang et al. published The GAN is dead; long live the GAN! (R3GAN), stripping away legacy StyleGAN2 heuristics in favor of a regularized relativistic GAN loss with local convergence guarantees and modern
ConvNeXt-style blocks—matching single-step diffusion baselines on ImageNet and FFHQ.2026
PaperLuo et al. introduced Adversarial Flow Distillation (AFD), pairing on-policy student rollouts with a prompt-conditioned video discriminator that translates teacher-student discrepancy on completed clips into dense forward-process velocity supervision—distilling black-box teachers into real-time causal autoregressive video generators.
See also (11)
Open full graphGenerative Adversarial Networks (GANs)