VAEs are a type of generative model that learn to encode input data into a probabilistic latent space and reconstruct it back. Unlike standard autoencoders, which map inputs to fixed vectors, VAEs model the latent space as a distribution—typically Gaussian—and learn both the mean and variance for each input.

Training is done by maximizing the evidence lower bound (ELBO), which balances two objectives: minimizing reconstruction loss (how well the output matches the input) and minimizing the KL divergence between the learned latent distribution and a prior (usually a standard normal). This regularization ensures that the latent space is smooth and allows meaningful sampling.

VAEs are widely used for tasks like image generation, anomaly detection, and representation learning. They offer more interpretable and structured latent spaces compared to GANs, though their outputs are typically blurrier.

Key components include an encoder network that outputs parameters of the latent distribution, and a decoder network that reconstructs the input from samples. The reparameterization trick is used to enable gradient-based training.