https://yupuyao.github.io/ai-meditations/blog/Recent content in Home on AI MeditationsHugoen-US© 2026 Yupu Yao.Fri, 08 May 2026 00:00:00 +0000https://yupuyao.github.io/ai-meditations/meditations-on-diffusion/Fri, 08 May 2026 00:00:00 +0000https://yupuyao.github.io/ai-meditations/meditations-on-diffusion/<h1 id="diffusion-is-not-merely-denoising-an-intuitive-view-of-what-ddpm-learns-from-sampling">Diffusion Is Not Merely Denoising: An Intuitive View of What DDPM Learns from Sampling</h1> <p>Discussions of diffusion models often begin with ELBOs, score matching, stochastic differential equations, or probability flow ODEs. These derivations are essential, but they can obscure a more basic intuition: what does the model actually learn? Why can a model start from pure Gaussian noise and, after a finite number of sampling steps, gradually produce an image with coherent semantic structure?</p>