Generative Adversarial Network (GAN)
Expanded definition
A generative adversarial network (GAN) is a generative modeling framework introduced by Goodfellow and collaborators in 2014. It trains two models at the same time. The generator maps a latent input, often sampled noise, into synthetic data. The discriminator receives real and generated samples and learns to distinguish between them. Training is adversarial because each model changes the other model's learning problem. The generator improves by producing samples that the discriminator is more likely to treat as real, while the discriminator improves by detecting generated samples. Conditional GANs add information such as a class label or another image so generation can be directed toward a requested output. GANs remain useful for understanding generative modeling and for workflows that benefit from fast generator inference or specialized image transformations. They can also be difficult to train: instability and mode collapse can leave the generator producing a narrow subset of the desired distribution. Diffusion models are a separate family with a different training and sampling path.
Common GAN variants
GAN variants change the architecture or the conditioning signal to fit different generation tasks. The generator-discriminator training relationship remains the core idea.
| Variant | What changes | Representative paper | Common use |
|---|---|---|---|
| DCGAN | Uses convolutional architectural constraints for unsupervised image generation. | Unsupervised Representation Learning with Deep Convolutional GANs | Image generation and representation-learning baselines |
| Conditional GAN | Conditions the generator and discriminator on additional information such as class labels. | Conditional Generative Adversarial Nets | Directed generation by class or other context |
| CycleGAN | Learns image-to-image translation between domains without requiring paired examples. | Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks | Unpaired image translation |
| StyleGAN | Introduces a style-based generator architecture that controls image synthesis at different scales. | A Style-Based Generator Architecture for Generative Adversarial Networks | High-resolution image synthesis research |
GANs vs diffusion models
| Dimension | GANs | Diffusion models |
|---|---|---|
| Core training idea | Train a generator against a discriminator that distinguishes real from generated samples. | Learn to reverse a gradual noising process. |
| Sampling path | Generate a sample with a forward pass through the trained generator. | Generate through an iterative denoising process, with the exact step count determined by the implementation. |
| Common engineering concern | Training stability and mode collapse require careful evaluation. | Iterative sampling adds inference work and requires scheduler and step-count decisions. |
| Representative image systems | DCGAN, CycleGAN, StyleGAN | DDPM, Stable Diffusion, Stable Diffusion XL |
| Decision rule | Use when the trained generator''s sampling path or a specialized GAN architecture matches the task. | Use when the diffusion ecosystem, conditioning controls, and measured output quality are the stronger fit. |
Choose with an evaluation set, not a family label. Compare output quality, diversity, controllability, latency, compute cost, safety controls, and licensing for the exact checkpoints and deployment path.
Related terms
Explore adjacent ideas in the knowledge graph.
Related
Comparisons, tools, and models that connect to this idea.