A class of generative models primarily used for generating high-quality images by progressively adding and then removing noise from an image.
Diffusion models learn to reverse noise: training corrupts images (or audio/video) step-by-step into static while the network learns each denoising step; generation then starts from pure noise and iteratively denoises into a coherent sample, guided by text or image conditioning. Latent diffusion runs this in a compressed space for speed, and modern samplers cut steps from hundreds toward a handful.
Diffusion is the engine of the visual generative wave, images, video, audio, even molecular design, and increasingly hybridizes with transformers (DiT backbones) in frontier media models. Understanding its iterative, guidance-driven nature explains both its quality ceiling and its controllability levers.
Generation starts from random noise; different seeds yield different valid samples. Fixing the seed (plus prompt and settings) reproduces an image: the basis of iteration workflows.
Classifier-free guidance scales how strongly generation follows the prompt versus the model's prior. Higher values increase adherence at the cost of naturalness; it's the main fidelity/creativity dial.
Transformers generate sequentially token-by-token; diffusion refines a whole sample in parallel over denoising steps. Media frontier models increasingly fuse them, using transformer backbones inside diffusion processes.