Paper · 2023-08-30 · cs.LG, cs.CV
Ten Years of Generative Adversarial Nets (GANs): A survey of the state-of-the-art
Survey of GAN theory, variants, applications, evaluation metrics and training failure modes, with variant loss formulas. It proposes no new method and reports no experiments of its own. Its headline claims about variants are summarised, not measured.
Generative adversarial network (GAN)Generative model evaluationFréchet Inception Distance (FID)generative-modelsgansurvey
AI-compiled by claude-haiku-5-5 on 2026-11-12, not yet reviewed by a human. Every claim cites its source; check it when a detail matters.
In plain words
A generative adversarial network, or GAN, is a pair of neural networks trained against each other. One network, the generator, makes fake samples from random noise. The other, the discriminator, tries to tell the fake samples from real ones. Each network gets better by competing with the other, and in the end the generator should produce samples that look like the real data.
This paper is a review of that idea. It explains how the pair is trained, lists many named variants and what each one changes, shows where GANs have been used, and describes the usual scores for judging their output. It also covers the common ways training goes wrong and what researchers have tried about them. It does not add a new model or new experiments. Use it as a map of the field and check the original paper before you implement any variant.
TL;DR
A survey, not a method. It gives the GAN minimax objective, the loss functions of about thirty variants, a catalogue of applications, the Jensen-Shannon link and related statistical results, the usual evaluation metrics and the training pathologies. It has no experiments and no official code. Use it for the variant loss forms and the metric list; take hyperparameters and results from the original papers.
Problem
Generative models are hard to train, and GAN training is unstable. The survey says training can produce mode collapse or oscillation, that conventional criteria may not capture diversity and realism, and that GANs can reflect biases in their training data §1. It also notes that the GAN loss gives no direct measure of training progress, so evaluation needs separate metrics §7.
The survey does not set a benchmark task. Its scope is the GAN literature from 2014 onward, across image, video, text, music, medical, geoscience, urban, time-series and privacy applications §1, §4.
Method
This section gives the formulas the survey states. It describes no procedure of its own, and no algorithm box appears in the paper.
Core objective. The generator G maps noise z from p_z to samples. The discriminator D outputs the probability that its input is real. The two play a minimax game with the value function below. D maximises it and G minimises it §3.
\underset{G}{\min}\;\underset{D}{\max}\; L = \mathbb{E}_{x\sim p_{\text{data}}}\left[\log D(x)\right] + \mathbb{E}_{z\sim p_{z}}\left[\log(1 - D(G(z)))\right] eq. 1 §3
Training loop (as described in text). The survey describes the two networks being trained in alternation, a continual back-and-forth process §3. It gives no batch size, step ratio, optimiser or learning rate, so these are not stated.
# Minimax GAN training, as described in the text [eq. 1] [§3]
# Input: real batch x ~ p_data; noise batch z ~ p_z
# Output: updated D and G
def gan_step(real, noise, D, G):
# Score real samples; D wants this large [eq. 1]
score_real = log(D(real)) # (B,)
# Score fake samples; D wants this large, G wants it small [eq. 1]
score_fake = log(1 - D(G(noise))) # (B,)
# Value of the minimax objective for this batch [eq. 1]
value = mean(score_real) + mean(score_fake)
# No parameter update is shown: the survey gives no optimiser [§3]
return valueVariant losses. The survey lists the following, with its own formulas where it gives them. Each is cited to its variant paragraph in §5.
- Conditional GAN. Both networks receive a condition y. The value function keeps the form of eq. 1 with D(x,y) and G(z,y) §5 CGAN.
- InfoGAN. The generator objective subtracts \lambda\,\mathcal{I}(c;G(z)), the mutual information between a latent code c and the output, weighted by a trade-off \lambda §5 InfoGAN.
- Least-squares GAN (LSGAN). Squared losses replace the log losses, with target labels a, b and c §5 LSGAN:
# LSGAN losses [§5 LSGAN]
# Labels: a for fake and b for real (D side)
# c is the value G wants D to give fakes
def lsgan_losses(D, G, real, noise, a, b, c):
# Generator pushes D's output on fakes towards c [§5 LSGAN]
loss_G = 0.5 * mean((D(G(noise)) - c) ** 2)
# Discriminator pushes real towards b and fakes towards a [§5 LSGAN]
loss_D = 0.5 * mean((D(real) - b) ** 2)
loss_D = loss_D + 0.5 * mean((D(G(noise)) - a) ** 2)
return loss_G, loss_D- Wasserstein GAN (WGAN). The critic estimates the Wasserstein-1 distance. The survey gives the distance as an infimum over couplings \Pi(\mathbb{P}, \mathbb{Q}) of \mathbb{E}\,\|\tilde{X}-\tilde{Y}\| §5 WGAN. Lipschitz continuity is enforced by clipping the discriminator weights §5 WGAN.
- CycleGAN. Two generators G_{XY} and G_{YX} and two discriminators translate between domains without paired data. The cycle-consistency term is \mathbb{E}_x\|G_{YX}(G_{XY}(x))-x\|_1 + \mathbb{E}_y\|G_{XY}(G_{YX}(y))-y\|_1 §5 CycleGAN. The survey gives no weight for this term in the full objective.
- SRGAN. The perceptual loss is l^{\text{SR}} = l^{\text{SR}}_x + 10^{-3}\, l^{\text{SR}}_{\text{Gen}}, the content loss plus a weighted adversarial loss. The content loss uses a pre-trained VGG-19 §5 SRGAN.
- StarGAN. One generator maps an image and a target-domain label c to an output. The discriminator has an added domain classifier. The generator loss combines the adversarial, classification and reconstruction terms, weighted by \lambda_1 and \lambda_2 §5 StarGAN.
- SeismoGen. The text says the formulation follows the Wasserstein GAN framework §5 SeismoGen. The loss is written as L_G = -\mathbb{E}_z D(G(z)) and L_D = \mathbb{E}_z D(G(z)) - \mathbb{E}_x D(x) + \lambda\,\mathbb{E}_z[(\|D(G(z))\|_2 - 1)^2] §5 SeismoGen.
The remaining variants in §5 are described in prose without a full loss: DCGAN §5 DCGAN, AAE §5 AAE, SAD-GAN §5 SAD-GAN, ProGAN §5 ProGAN, MidiNet §5 MidiNet, SN-GAN §5 SN-GAN, RGAN §5 RGAN, BigGAN §5 BigGAN, MI-GAN §5 MI-GAN, AttGAN §5 AttGAN, DM-GAN §5 DM-GAN, SinGAN §5 SinGAN, PATE-GAN §5 PATE-GAN, POLY-GAN §5 POLY-GAN, MIEGAN §5 MIEGAN, VQGAN §5 VQGAN, DALL-E §5 DALL-E, CEGAN §5 CEGAN, MetroGAN, M3GAN, CNTS and RidgeGAN §5.
Theory. The survey summarises results by Biau et al. linking GAN training to the Jensen-Shannon divergence, with convergence-rate statements in JS divergence §6. It also cites work on statistical inference for GANs and PAC-Bayesian generalisation bounds for WGANs §6. These are reported as cited, not derived.
Evaluation metrics §7:
- Inception Score (IS): uses Inception v3 class probabilities. The text gives no formula §7.1.
- Fréchet Inception Distance (FID): \operatorname{FID} = |\mu-\mu_w|^2 + \operatorname{tr}(\Sigma + \Sigma_w - 2(\Sigma\Sigma_w)^{1/2}), with (\mu, \Sigma) for real and (\mu_w, \Sigma_w) for generated images §7.2.
- MS-SSIM, C2ST, maximum mean discrepancy, time-series metrics and uncertainty quantification are described in prose §7.3–§7.8.
Hyperparameters
The survey runs no experiments, so it has no training configuration of its own. Values below are those stated for specific variants in the text.
| Name | Value | Where | Notes |
|---|---|---|---|
| SRGAN upscaling factor | 4 | §5 SRGAN | |
| SRGAN adversarial-term weight | 10^{-3} | §5 SRGAN | multiplies the adversarial loss in the perceptual loss |
| ProGAN resolution range | 4×4 up to 1024×1024 | §5 ProGAN | given as an example; schedule not stated |
| BigGAN output size | up to 512×512 | §5 BigGAN | stated in text |
| IS classifier | Inception v3, trained on ImageNet's 1000 classes | §7.1 | |
| InfoGAN \lambda | not stated | §5 InfoGAN | trade-off with the mutual-information term |
| LSGAN labels a, b, c | not stated | §5 LSGAN | "a-b encoding scheme" only |
| CycleGAN cycle-loss weight | not stated | §5 CycleGAN | |
| StarGAN \lambda_1, \lambda_2 | not stated | §5 StarGAN | classification and reconstruction weights |
| AttGAN \lambda_{\text{Rec}}, \lambda_{\text{Cls}_G}, \lambda_{\text{Cls}_D} | not stated | §5 AttGAN | |
| WGAN weight-clipping limit | not stated | §5 WGAN | |
| SeismoGen \lambda | not stated | §5 SeismoGen | |
| Optimiser, learning rate, batch size, D:G step ratio | not stated | §3, §5 | not given for any variant |
Results
No results are reported. The survey has no experiments and no benchmark table. Its quantitative content is limited to:
- Survey coverage tab. 1. The survey compares itself with 13 earlier surveys on nine criteria, using full, partial or no coverage. It marks itself "fully covered" on all nine. This is the authors' self-assessment, not independently checked.
- Software links tab. 2. Thirty variants with a language, backend and repository. Twenty-seven rows give a GitHub link. Rows 5 (SAD-GAN), 22 (MIEGAN) and 25 (CEGAN) give "--" for all fields. See Paper vs code.
- Headline claims in the text, with no numbers given in the survey:
- LSGAN is said to beat regular GANs on Inception score on CIFAR-10, among other datasets §5 LSGAN. The survey gives no score.
- SeqGAN is said to beat earlier text-generation methods §2. No score given.
- Diffusion models are said to beat GANs on image synthesis, which the survey credits to their greater stability and lack of mode collapse, citing Dhariwal et al. §9. The survey states this but does not measure it.
Paper vs code
No official code. The paper has no code link of its own, and the arXiv metadata records no comments field. Table 2 tab. 2 lists third-party repositories for the variants. None is presented as the authors' code, and none is used here. Search: no project page is linked in the paper; a GitHub repository search for the paper's title returned no results (WebFetch, 2026-11-12).
Discrepancies inside the paper, each checked against its text:
- ProGAN is described as training several generator and discriminator networks one step at a time §2. The ProGAN paragraph describes one generator and one discriminator whose size grows §5 ProGAN. The two passages disagree.
- The survey says Goodfellow et al. "proved the existence of a unique solution" of the minimax loss §3. The survey gives no proof and does not check this claim against the original.
- The TeX source labels the timeline figure
fig_no3, but it appears before the application-domain figure, so the rendered PDF numbers it Figure 2 fig. 2. This card uses rendered numbers: fig. 1 architecture, fig. 2 timeline, fig. 3 application domains. Only one equation is numbered in the paper, eq. 1; the others are unnumbered and are cited by section.
Pitfalls
- Only the minimax form is given eq. 1. The survey does not discuss alternative generator losses or say which one a given variant uses.
- The SeismoGen penalty is written on the discriminator output norm, (\|D(G(z))\|_2 - 1)^2 §5 SeismoGen, not on an input-gradient norm. Check the original Seismogen paper before implementing this as a gradient penalty.
- Several variant formulas use \lambda with no value. InfoGAN, StarGAN, AttGAN, SeismoGen and the WGAN clipping threshold have no stated value §5. Any implementation needs values from the original paper or its code.
- InfoGAN's generator loss has a minus sign on the mutual-information term §5 InfoGAN. Whether that matches the original is not checked here.
- Benefit claims are asserted, not measured. Examples: AAEs "exhibit noise resistance" §5 AAE; CGANs show "superior noise resistance" §5 CGAN; RGAN has "improved resilience against adversarial attacks" §5 RGAN. The survey cites no experiment for these.
- Mode collapse, vanishing gradients and instability are covered in §8 with named remedies: unrolled GAN, minibatch discrimination, least-squares and Wasserstein losses, feature matching, historical averaging, and batch normalization §8.1–§8.5. The survey gives no numbers for any of them.
- The IS formula is missing §7.1. The FID formula is given §7.2, but the text does not state the sample count, feature layer or preprocessing a reliable FID needs. Those belong to the FID concept page and to the papers that define the metric.
- Some variant descriptions have no link. SAD-GAN, MIEGAN and CEGAN have "--" in tab. 2; the other 27 links are not vetted.
Lineage
- Builds on: Goodfellow et al. (2014), which introduced the GAN objective [§1, §3]. The bibliography entry has no arXiv id, and that paper is not a card in this wiki.
- Compared against: 13 earlier surveys, including Kulkarni et al. (2019) and Jabbar et al. (2021), in tab. 1 §2.
- Superseded by: not known from content already in this wiki.
Source and license
Adapted from “Ten Years of Generative Adversarial Nets (GANs): A survey of the state-of-the-art” by Tanujit Chakraborty, Ujjwal Reddy K S, Shraddha M. Naik, Madhurima Panja, Bayapureddy Manvitha (arXiv:2308.16316), used under CC BY 4.0. Changes: condensed and restructured for implementation, with tables, equations and algorithms taken from the paper and its code, each cited to its source. Not affiliated with or endorsed by the authors.