Feed a variational autoencoder thousands of face photographs and it will learn a compressed map of what faces look like, then generate new ones that never existed. The model belongs to the generative family, and its distinguishing feature is a probabilistic twist on the ordinary autoencoder.

A standard autoencoder squeezes input into a fixed vector and reconstructs it. A VAE instead encodes each input as a distribution, typically a mean and a variance, then samples from that distribution before decoding. This forced randomness smooths the latent space so that nearby points decode into similar outputs, which makes interpolation and sampling meaningful.

Key ingredients

• An encoder network that outputs parameters of a latent distribution
• A sampling step using the reparameterization trick to keep gradients flowing
• A decoder network that reconstructs the input from a sampled latent vector
• A loss combining reconstruction error with a KL divergence term that regularizes the latent space

VAEs produce blurrier images than GANs, but their latent spaces are more structured and easier to navigate. That trade-off keeps them popular for drug discovery, anomaly detection, and any task where a smooth, interpretable representation matters more than pixel-perfect sharpness.