AWS Certified AI Practitioner AIF-C01 Practice Question

Diffusion models are trained by gradually adding Gaussian noise to real images and learning how to reverse that noising process. At inference time they do the opposite: starting from pure random noise, the model applies a sequence of learned denoising steps that progressively remove noise and reveal structure until a coherent image emerges. The other options describe mechanisms used by GANs, autoregressive pixel transformers, or simple latent-space interpolation, none of which rely on step-wise denoising of an initially random input.