What Is a VAE?
7 entriesA variational autoencoder learns a smooth, probabilistic latent space. An encoder maps data to a distribution over latent variables, a decoder reconstructs data from samples of it, and training maximises the evidence lower bound (ELBO): reconstruction quality plus a KL term that keeps the latent space well-behaved. The reparameterisation trick makes sampling differentiable. VAEs train stably and support interpolation and anomaly detection, though their raw samples are blurrier than GANs or diffusion; today they are most important as the autoencoder that compresses images for latent diffusion models.
Landmark Papers
7 entriesFrom the original VAE to hierarchical and discrete models.
Code & Libraries
10 entriesImplementations to read and run.
Related Pages
10 entriesNeighbouring pages on this site.
Frequently Asked Questions
What is a variational autoencoder?
A generative model that encodes data into a probability distribution over a latent space and decodes samples back to data, trained by maximising the evidence lower bound.
What is the reparameterisation trick?
Writing a sample as a deterministic function of the distribution parameters plus noise, so gradients can flow through the sampling step.
How do VAEs relate to diffusion models?
Latent diffusion models such as Stable Diffusion use a VAE-style autoencoder to compress images, then run diffusion in the smaller latent space, which makes high-resolution generation affordable.
When should I use a VAE?
For stable training, learned latent representations, interpolation, anomaly detection and as a compression stage in larger generative systems; use diffusion or GANs when you need the sharpest samples.
How current is this list?
Links were checked in October 2026. Archived projects are labelled; sites that block automated checks are kept and may need a manual click.