GAN

AI concepts
About 1 min read

A deep learning architecture where a generator and a discriminator compete to generate realistic data. Because its generation speed is significantly faster than diffusion models, it is mainly used for real-time video translation and high-resolution restoration.

Also known as
Generative Adversarial NetworkGenerative Adversarial Network

Detailed explanation

Proposed by Ian Goodfellow in 2014, GAN is a technology that refines the sophistication of data through a process in which two neural networks, a Generator and a Discriminator, attempt to deceive and detect each other. The generator creates realistic fake data, while the discriminator learns to detect it. While the text-to-image generation field is currently dominated by diffusion models, GANs still play a key role in real-time face filters, virtual human creation, super-resolution, and unpaired data translation (CycleGAN) due to their fast inference speed that yields results in a single operation. However, managing 'Mode Collapse'—a phenomenon where the model loses diversity during the training process—remains a major criterion for tool selection and application.

Why it matters in tool selection

If users prioritize 'real-time performance' and 'low computational cost', they should choose GAN-based tools. While modern diffusion models offer high quality, they take seconds to dozens of seconds to generate results, whereas GANs allow inference in milliseconds (ms), making them suitable for real-time filters in mobile apps or interactive services.

Key considerations

  • Is it an environment requiring real-time generation (low latency)?
  • Is there a lack of training data, or is translation between unpaired data (e.g., horse photos to zebra photos) required?
  • Does the generation tool feature algorithms to prevent mode collapse (repeatedly generating only specific images)?

Examples

Representative examples include real-time face translation filters on SNOW or TikTok, security solutions that upscale low-resolution CCTV footage to 4K, and virtual human solutions that generate high-resolution portraits of non-existent people.

Easily confused terms

Diffusion Model

Generates data by step-by-step noise removal. Though slower than GANs, it offers far superior quality and diversity.

VAE (Variational Autoencoder)

Generates data by compressing and then reconstructing data features. The structure is more stable than GANs, but the output tends to be somewhat blurry.

Related terms

Diffusion Model