GANs

Gold definitionUpdated Apr 2, 2026

Definition

Generative Adversarial Networks (GANs) are a class of generative models that learn to produce new data instances through an adversarial process between a generator and a discriminator network. They are widely used for synthesizing realistic content like images and videos.

At a glance

Executive summary

Generative Adversarial Networks (GANs) are a type of AI that learns to create new, realistic data like images or videos by having two neural networks compete against each other. One network generates content, and the other tries to tell if it's real or fake, leading to increasingly convincing fakes.

TL;DR

GANs are AI models that generate realistic fake data, like images or videos, by having two parts of the model compete to get better at generating and detecting fakes.

Key points

Employs an adversarial training framework with a generator and a discriminator network
Enables the creation of highly realistic synthetic data, useful for data augmentation, content generation, and simulation
Used by researchers in computer vision, video processing (e.g., summarization), and deepfake generation/detection
Unlike VAEs which learn a probabilistic mapping and often produce blurrier outputs, GANs aim for sharper, more realistic samples by directly optimizing for indistinguishability from real data
Evolving towards more stable training methods, higher resolution generation, and integration with other generative models like diffusion models

Use cases

Deepfake Generation: Creating highly realistic synthetic videos and images for various purposes, including entertainment or malicious use.
Video Summarization: Enhancing keyframe selection and generating coherent video summaries to efficiently browse large multimedia content.
Image Synthesis: Generating photorealistic images from text descriptions or creating images of non-existent objects or people.
Data Augmentation: Expanding limited datasets by generating synthetic training examples, especially in medical imaging or rare event detection.
Style Transfer: Applying the artistic style of one image to the content of another, creating new stylized artworks.

Also known as

DCGAN, WGAN, StyleGAN, Conditional GAN (cGAN), CycleGAN, BigGAN, InfoGAN