Latent Space
The compressed, abstract representation a generative model works in before decoding its output into pixels.
Latent space is the compressed, abstract representation a generative model manipulates internally, rather than working with full-resolution pixels directly. A diffusion model typically denoises within this compressed space, and only a final decoding step expands the result into the pixels you actually see.
Generating in latent space is dramatically cheaper than doing the equivalent work at full pixel resolution, which is a big part of why modern image and video models are fast enough to be usable.
See also: Diffusion Model, Upscaling.
Put these ideas into motion.
DownloadFrequently asked questions
Latent space is a compressed mathematical representation of an image that a model manipulates internally. A separate decoder step turns that representation into the final pixels you see.
Working in a compressed representation is dramatically cheaper to compute than manipulating full-resolution pixels at every step, which is what made modern diffusion models fast enough to be practical.