In diffusion models, the Latent (latent space) performs noise diffusion and denoising in a compressed low-dimensional space, which is then decoded into pixel output by the VAE, making it orders of magnitude faster than operating in the original pixel space.