VAE (variational autoencoder)

A VAE, or variational autoencoder, is the component that translates between the compact internal representation a diffusion model works in, called latent space, and the actual pixels you see in the final image. It compresses images for efficient processing during generation and then decodes the result back into a full-resolution picture, and it has a noticeable effect on color accuracy and fine detail.

In tendre.AI, the Krea 2 VAE is installed with the image engine, so images decode correctly without extra setup.