Diffusion Models ☆

Diffusion models learn to reverse noise: start from static, subtract noise step by step, arrive at an image (or video, or protein pose) matching the prompt. Stable Diffusion, FLUX, Sora-class video and AlphaFold 3's structure module all use it. Inference takes many denoising steps, which is why image/video generation is so much more compute-hungry per output than text.