VAE (Autoencoders)

← Back to the article

A VAE squeezes an image into a small code, adds a little randomness, and rebuilds a new image.

What it does

A VAE squeezes an image into a small code, adds a little randomness, and rebuilds a new image.

Hidden code: A small number-map for an image • Encoder: The squasher

How it works

  • Compress the image into a small code.
  • Add a small random change.
  • Decode the code into a new image.

Uses and applications

  • A design tool can create new face-like examples.
  • A picture app can explore smooth changes between images.
  • A research tool can learn the main shape of a group of images.

In short

It is like describing a face to a sketch artist.

Related methods

  • GANs: Generates with an adversarial judge instead of rebuilding inputs.
  • PCA (Data Shrinker): A linear way to compress data into fewer numbers.
  • Stable Diffusion: Uses a VAE to move between images and its latent space.

Back to Image and content generation · All topics

The interactive demonstration and the Python walkthrough for this topic need JavaScript.