What is VAE (Variational Autoencoder)?
A neural network that compresses images into latent representations and decodes them back. The encoder/decoder component of latent diffusion models.
How it works
A VAE consists of an encoder (compresses images to latent space) and a decoder (reconstructs images from latent space). In latent diffusion models, the VAE is responsible for the compression that makes generation computationally feasible. The quality of the VAE directly affects output quality: a better VAE preserves more detail during compression and produces sharper, more accurate reconstructions. This is why some models produce sharper output than others even when using similar diffusion processes. The VAE is trained separately from the diffusion model and can sometimes be upgraded independently to improve output quality.
Tools that use vae (variational autoencoder)
Related terms
Frequently asked questions
What does vae (variational autoencoder) mean in AI video?▾
Need AI video produced by professionals, not generated by yourself?
Apostle is an AI-native video production studio. We use every tool on this page in real client work.
Get in touch