Architecture
aka variational autoencoder
VAE
Probabilistic autoencoder that learns a smooth latent space with a Gaussian prior.
Definition
Variational autoencoders parameterise the latent distribution as a Gaussian and optimise the evidence lower bound. The resulting latent space is smooth and samplable, which is why nearly every latent-diffusion image model pairs a VAE with a denoising U-Net.
Common use cases
- Image compression
- Stable Diffusion latent space
- Generative modelling