What is latent diffusion and why is it more efficient than pixel-space diffusion?

Latent diffusion operates in compressed latent space instead of pixel space, making generation faster and cheaper.

Latent diffusion models first encode images into a lower-dimensional latent space using a variational autoencoder. Diffusion is then applied in this compressed space, significantly reducing computational cost while preserving semantic structure. Finally, a decoder reconstructs the image. This enables high-resolution image generation efficiently.