Glossary
VAE (Variational Autoencoder)
The component that converts between pixels and the latent space the diffusion model actually works in.
Diffusion models don't operate on raw pixels — they work in a compressed 'latent' space. The VAE encodes images into that latent space and decodes the model's output back to pixels. A bad or mismatched VAE produces washed-out colours or grey casts; the right VAE is baked into modern checkpoints and you rarely have to think about it.