Latent Diffusion
A highly efficient generative model architecture that performs the diffusion process (adding/removing noise) in a compressed lower-dimensional latent space.
Think of It Like This
Like an artist sketching a complex scene as a rough, tiny thumbnail before meticulously painting the full high-resolution canvas.
By utilizing an autoencoder to compress images into latents, models like Stable Diffusion drastically reduce the computational complexity of image generation. The denoising U-Net operates entirely in this compressed space. A decoder then maps the final clean latent back into pixel space, making high-quality generation possible on consumer GPUs.