Diffusion: Watch Noise Become a Shape
My roleIndependent build · the DDPM sampler, the closed-form score, and the visualization
The reverse process behind image generators, shrunk to two dimensions: a cloud of pure noise denoised step by step onto a target, live in the browser.
Overview
An interactive, in-browser demonstration of denoising diffusion, the method behind Stable Diffusion, DALL-E, and Imagen. A cloud of Gaussian-noise particles is denoised one step at a time using the exact ancestral sampling rule of Ho, Jain and Abbeel (2020): at each noise level the demo estimates the clean data as a posterior mean over the target points, steps toward it, and adds back a precise amount of noise. Because the target is a known set of points, the noised distribution is a mixture of Gaussians and its score has a closed form, so no training is needed. That is the one honest simplification: real image models cannot write the score down and instead train a large U-Net to estimate it, while the schedule, the posterior step, and the added noise you are watching are identical. A live cloud-spread trace shows the sample converging rather than replaying a canned animation.
Key Features
- ✓Exact DDPM ancestral sampling (Ho et al., 2020) run per particle in real time
- ✓Closed-form denoising score for a known target mixture, no training required
- ✓Honest framing: the same sampling math as real image diffusion, only in 2D
- ✓Live cloud-spread readout is direct evidence the sample is converging