What is a GAN (Generative Adversarial Network)? — AI Glossary

A GAN (Generative Adversarial Network) is a type of AI system that generates realistic synthetic data — images, audio, video, or text — by pitting two neural networks against each other in a competition. One network creates fake content; the other tries to distinguish fakes from real. Through this adversarial process, both improve, and the generator eventually produces convincingly realistic outputs.

GANs were invented by Ian Goodfellow in 2014 and sparked a revolution in AI-generated media. The “this person does not exist” websites that show photorealistic faces of people who never lived? Those are GANs. The deepfake videos of public figures? GANs. AI-generated fashion photography and product renders? Often GANs. They introduced something genuinely new: AI that creates rather than just classifies.

How GANs Work

A GAN consists of two competing neural networks:

• Generator (G) — takes random noise as input and outputs synthetic data (e.g., a fake image). Its goal: fool the discriminator.
• Discriminator (D) — receives either real data or generator output and predicts “real” or “fake.” Its goal: catch the generator’s fakes.

Training alternates between the two:

• Train the discriminator on real examples (labeled “real”) and generator outputs (labeled “fake”)
• Train the generator to produce outputs the discriminator classifies as “real”

As the discriminator gets better at spotting fakes, the generator must improve to fool it. This feedback loop drives both networks toward higher quality. When the GAN reaches equilibrium — the generator’s outputs are indistinguishable from real data — the discriminator can only guess randomly. At that point, the generator is creating convincingly realistic content.

Why GANs Matter

GANs were the first practical approach to high-quality image generation and remained dominant until diffusion models emerged around 2020–2022. Their contributions include:

• Demonstrating that AI could generate photorealistic images
• Enabling synthetic data generation to augment real datasets
• Powering image-to-image translation (turning sketches into photos, converting day scenes to night)
• Video and audio synthesis
• Artistic style transfer and creative tools

They also raised important ethical concerns. Deepfakes — AI-generated videos that convincingly substitute one person’s face for another — are primarily GAN-based and have been used for misinformation and non-consensual intimate imagery. This is a core AI governance challenge.

GANs in Practice

Notable GAN variants include:

• StyleGAN / StyleGAN2/3 (NVIDIA) — produces photorealistic faces and is used in art, gaming, and fashion
• CycleGAN — unpaired image-to-image translation (horses to zebras without paired training examples)
• Pix2Pix — paired image translation (architectural sketches to photos)
• BigGAN — high-resolution class-conditional image synthesis

While diffusion models have largely surpassed GANs for image quality and training stability, GANs remain valuable for applications requiring fast inference — diffusion models generate images slowly through many denoising steps, while a trained GAN produces an image in a single forward pass.

Common Misconceptions

Misconception: All AI-generated images come from GANs. Since 2022, diffusion models (Stable Diffusion, DALL-E 3, Midjourney) have become the dominant approach for text-to-image generation. GANs are still used in many contexts, but they no longer dominate the landscape.

Misconception: GAN training is easy and stable. GAN training is notoriously unstable. Mode collapse (the generator learns to produce only a few types of output), vanishing gradients, and training divergence are constant challenges. This instability is one reason diffusion models gained traction.

---

Key Takeaways

• GANs consist of two competing networks: a generator and a discriminator.
• Adversarial training drives both networks to improve until the generator produces realistic outputs.
• GANs pioneered AI-generated images, deepfakes, and synthetic data creation.
• Diffusion models have largely replaced GANs for text-to-image generation but GANs remain fast at inference.
• Ethical concerns around deepfakes and synthetic media are a major AI governance challenge.

Frequently Asked Questions

What is the difference between a GAN and a diffusion model?

GANs generate images in a single forward pass through the generator. Diffusion models iteratively denoise a random image over many steps. Diffusion models produce higher quality and more diverse outputs but are slower at inference. GANs are faster but harder to train.

What is mode collapse in a GAN?

Mode collapse happens when the generator learns to produce only a narrow range of outputs — for example, always generating the same few types of faces instead of diverse ones. The generator has found a strategy that fools the discriminator but doesn’t reflect the full diversity of real data.

Can GANs generate text?

GANs struggle with discrete outputs like text because the generator can’t be differentiated through discrete tokens. Most text generation uses autoregressive LLMs or diffusion models. GANs are primarily used for continuous data like images, audio, and video.

Are deepfakes always made with GANs?

Historically yes, but newer deepfakes increasingly use diffusion models and other techniques. The key characteristic of a deepfake is manipulation of media to misrepresent a person, regardless of the specific AI method used.

What is a conditional GAN?

A conditional GAN (cGAN) feeds additional information — a class label, a text description, or another image — to both the generator and discriminator, conditioning the generation. This lets you control what kind of output is generated (e.g., “generate a cat image”) rather than getting random outputs.

---

Sources: Grokipedia — GAN · arXiv: Generative Adversarial Networks (Goodfellow et al., 2014) · This Person Does Not Exist (StyleGAN demo)

You May Also Like

• What Is Artificial Intelligence
• AI Glossary
• AI for Dummies
• How Does AI Work
• Essential AI Skills

---

Get free AI tips daily → Subscribe to Beginners in AI

Sources

This article draws on official documentation, product pages, and industry reporting. Specific sources are linked inline throughout the text.

Last reviewed: April 2026