Skip to main content
All terms
Multimodal

Latent Diffusion

Running the image-building process on a compact summary of the image to cut computing cost.

Definition

Latent diffusion builds images from random noise while working on a compact numeric summary of the picture, learned by a small companion network, rather than on every pixel—which sharply reduces computing cost. A cleanup network gradually removes noise from this summary, guided by text or other input, and a decoder then expands the finished summary back into a full image. Stable Diffusion is the most widely adopted example.