Diffusion

Diffusion models generate images or other media by iteratively denoising from noise toward a sample that matches a prompt or condition.

A diffusion process trains a network to reverse noise added to data. At inference, you start from noise (or a partially noised image for img2img) and step toward a clean sample. Samplers, step counts, and guidance scales change quality and speed. Latent diffusion runs this process in a compressed space for efficiency. Diffusion is the usual engine behind SDXL-class image models; it is not the onl...