Diffusion models are the dominant paradigm behind modern image and video generation. Instead of constructing images directly, they learn how to systematically remove noise, transforming randomness into structure through a sequence of small, learned corrections. This lens explores diffusion as both a conceptual idea and a practical system. It covers how diffusion models are trained, why architectures like U-Nets and attention are effective for denoising, and how conditioning signals—such as text, images, or structural constraints—guide generation without explicitly defining rules.