Latent Diffusion

Vanilla diffusion denoises full images — 512×512×3 values, a thousand times over. Most of that work is wasted: neighboring pixels are nearly redundant, and the model burns compute resynthesizing texture it already knows.

Latent diffusion — the architecture of Stable Diffusion — moves the whole process into a VAE's compressed latent space: encode once, denoise a 64×64 latent instead of a 512×512 image (roughly 48× less data per step), then decode once at the end.