A denoising diffusion model samples an image — pure Gaussian noise iteratively refined over T steps as the U-Net predicts and subtracts noise, a blurry structure emerges, sharpens, and converges to a clean image.