Generative Adversarial Networks

A generative adversarial network (GAN) is a framework for training a generative model in which two neural networks are trained against each other: a generator that maps random noise to synthetic samples, and a discriminator that tries to tell synthetic samples from real training data. The generator is trained to make the discriminator err, so the generator's distribution is pushed toward the data distribution without an explicit likelihood being computed (Goodfellow et al., 2014).

Generative modelling aims to learn a distribution $p_{\text{data}}$ well enough to draw new samples from it. Many earlier approaches required an explicit, tractable density or approximate inference procedures (for example Markov chain sampling). Goodfellow et al. (2014) proposed an alternative: define a sampler directly as a network and train it using a second network's judgement as the learning signal. The paper notes that, when both networks are multilayer perceptrons, the system can be trained with backpropagation and needs no Markov chains or unrolled approximate inference networks (see [TECH-BACKPROP-001]).

GANs were the leading approach to high-fidelity image synthesis for several years and introduced an influential idea, learning from a trained critic rather than from a fixed loss. Diffusion models have since become prominent for large-scale image applications; Rombach et al. (2021) state in their abstract that diffusion models achieve state-of-the-art synthesis results on image data, and their own latent diffusion models are reported to set a new state of the art for image inpainting and to be highly competitive on other tasks (see [TECH-DIFF-001]). GANs remain relevant for speed at generation time and for specific tasks. The sources cited here do not establish whether diffusion has displaced GANs in practice. Their training is also a standard example of the difficulties of optimizing a game rather than a single loss.

Goodfellow, Pouget-Abadie, Mirza, Xu, Warde-Farley, Ozair, Courville and Bengio (2014) introduced the framework. Radford, Metz and Chintala (2015) proposed a class of convolutional GANs (DCGANs) with architectural constraints that made training more reliable on image datasets, and reported evidence that the networks learn a hierarchy of representations (see [TECH-CNN-001]).

Training stability then became a main research theme. Arjovsky, Chintala and Bottou (2017) proposed the Wasserstein GAN (WGAN), whose authors report improved learning stability and meaningful learning curves, and offer a theoretical account in terms of distances between distributions. Gulrajani et al. (2017) reported that WGAN's weight clipping could lead to undesired behaviour and proposed a gradient penalty on the critic as an alternative. Karras, Laine and Aila (2018) proposed a style-based generator architecture that, as reported in the abstract, yields automatic separation of high-level attributes from stochastic variation and improved distribution-quality metrics on the datasets studied (StyleGAN).

Diffusion models became prominent for image generation in the early 2020s (see [TECH-DIFF-001]; the cited Rombach et al. (2021) paper dates from December 2021, and earlier diffusion papers are not cited in this article), and pretrained autoencoders are components of latent diffusion systems (Rombach et al., 2021; see [TECH-VAE-001]).

#4.1 The minimax game

Let $z \sim p_z$ be noise, $G_\theta(z)$ the generator and $D_\phi(x) \in (0,1)$ the discriminator's estimate that $x$ came from the training data. The original objective is the two-player game (Goodfellow et al., 2014)

The discriminator is trained to maximize $V$ (classify real versus fake correctly), and the generator to minimize it.

#4.2 Optimal discriminator and the divergence interpretation

For a fixed generator with sample distribution $p_g$, the optimal discriminator is

Substituting $D^*$ into $V$ yields

where JSD is the Jensen–Shannon divergence. The global minimum is reached when $p_g = p_{\text{data}}$, at which point $D^*(x) = 1/2$ everywhere. This result holds in the space of arbitrary functions $G$ and $D$ (Goodfellow et al., 2014), which is an idealization: real networks have finite capacity, and training is by stochastic gradient steps, so the guarantee does not carry over directly to practice.

#4.3 Training in practice

Training alternates gradient steps on $D$ and $G$ using minibatches. Early in training, when $D$ easily rejects generated samples, the term $\log(1 - D(G(z)))$ saturates and gives the generator weak gradients. The original paper suggests having the generator instead maximize $\log D(G(z))$, the non-saturating loss, which has the same fixed point and stronger early gradients.

#4.4 Wasserstein formulation

WGAN replaces the discriminator with a critic that estimates a Wasserstein distance between distributions, with a Lipschitz constraint on the critic (Arjovsky et al., 2017). The original method enforced the constraint by clipping weights. Gulrajani et al. (2017) instead penalized the norm of the critic's gradient with respect to its input and reported stable training across a wide variety of architectures.

#4.5 Conditional and style-based generation

A conditional GAN feeds a label or other conditioning signal to both networks, so that generation can be directed. Style-based generators inject a learned intermediate latent code at multiple layers of the generator and support scale-specific control of the synthesis (Karras et al., 2018). Conditional GANs were introduced by feeding the conditioning data to both the generator and the discriminator (Mirza & Osindero, 2014).

• Image synthesis. Face and object image generation is the best documented use (Radford et al., 2015; Karras et al., 2018).
• Representation learning. DCGAN authors reported that the discriminator's learned features were useful for other tasks (Radford et al., 2015).
• Data augmentation, image-to-image translation, super-resolution. Conditional adversarial networks have been applied to image-to-image translation tasks such as synthesizing photos from label maps, reconstructing objects from edge maps and colorizing images (Isola et al., 2016), and a GAN-based method has been proposed for single-image super-resolution (Ledig et al., 2016). Data augmentation is a commonly described use for which no source is cited here.
• Synthetic media misuse. Realistic synthetic images and video raise concerns about deception and provenance; that topic belongs to the safety and governance volumes and is not analysed here (see [SAFE-IP-001] for copyright aspects and [GOV-LAW-001] for governance). No empirical source on the prevalence of misuse is cited here.

Performance comparisons, model sizes and benchmark scores are omitted because they change quickly; see MODEL-001.

#Advantages

• Sharp samples in the studied image settings. Authors report high-quality image generation for GAN variants (Karras et al., 2018).
• Fast sampling. Generation requires one forward pass of the generator, in contrast to the many sequential evaluations typical of diffusion models (compare the cost remarks in Rombach et al., 2021).
• No explicit likelihood or inference network required (Goodfellow et al., 2014).

#Limitations

• Training instability. Gulrajani et al. (2017) state that GANs "suffer from training instability", and WGAN was proposed to improve stability (Arjovsky et al., 2017).
• Mode collapse. The generator may cover only part of the data distribution. Arjovsky et al. (2017) claim their approach removes problems such as mode collapse, but the extent to which this holds in all settings is not established by the sources reviewed here.
• No likelihood. A GAN does not provide a density estimate for evaluating samples or comparing models by likelihood.
• Evaluation difficulty. Judging sample quality and diversity requires auxiliary metrics, each with its own weaknesses. No source on specific metrics is cited here.

#Unresolved questions

1. Convergence of the game. Whether simultaneous gradient training of $G$ and $D$ converges in practice, and under what conditions, remains a research question; the idealized theory does not cover finite-capacity networks (see section 4.2).
2. GANs versus diffusion models. Which family is better depends on the criterion: sample quality, diversity, sampling speed, training stability and compute. Rombach et al. (2021) report state-of-the-art results for diffusion models, but a comprehensive controlled comparison was not confirmed for this draft.

As of 2026-10-19, GANs are a mature family whose core theory is understood in the idealized setting (Established, Goodfellow et al., 2014) and whose practical training is supported by a body of stabilization methods (Well Supported: Arjovsky et al., 2017; Gulrajani et al., 2017). For large-scale text-conditioned image generation, diffusion-based systems have become prominent (Rombach et al., 2021 demonstrates text and other conditioning in latent diffusion models; prevalence in deployed systems is not documented in the sources cited here). The question of how GANs and other generative families compare is partly answered by benchmarks that change quickly (see MODEL-001).

This section describes plausible directions and does not register forecasts.

• Near term (0–3 years). Adversarial training is expected by some researchers to persist mainly as a component or fine-tuning signal within other generators rather than as a stand-alone approach; this is an observation about research direction, not a documented result. No source is cited for this trend.
• Medium term (3–10 years). The relative position of adversarial training in a landscape of diffusion, autoregressive and hybrid generators depends on sampling cost and quality comparisons that are not settled.
• Long term (10+ years). The adversarial idea of learning from a trained critic also appears outside image synthesis, for example in alignment techniques that use a learned judge (see [TECH-ALIGN-001]); this is an analogy and not an equivalence.

• Foundational Concepts:
  • TECH-NN-001 Artificial Neural Networks: the networks used as generator and discriminator.
  • TECH-BACKPROP-001 Backpropagation: gradients through the discriminator into the generator.
• Related Methods & Technologies:
  • TECH-VAE-001 Autoencoders and Variational Autoencoders: the likelihood-based latent-variable alternative.
  • TECH-DIFF-001 Diffusion Models: the family that has displaced GANs in many image applications.
  • TECH-CNN-001 Convolutional Neural Networks: the backbone of DCGAN-style generators.
  • TECH-ALIGN-001 Pretraining, Fine-Tuning and RLHF: learned-judge training in language models.
• Advanced Implications & Future Trajectories:
  • CAP-GENAI-001 Generative Artificial Intelligence: the capability landscape of generative systems.
  • SAFE-IP-001 Copyright and Intellectual Property: legal questions raised by generated media.