🌫️ Diffusion Model Lab: Noise Schedules, Reverse Sampling and DDPM vs DDIM

A 2-D toy (ring of Gaussians, two moons, swirl, overlapping blobs) that shows the maths of diffusion generators: forward noising with linear vs cosine schedules and SNR curves, reverse sampling with DDPM and DDIM (eta, step count, horizon T), the probability-flow ODE view and classifier-free guidance, using the exact analytic score of a Gaussian mixture (plus an optional small MLP trained in the browser), with energy-distance quality checks. It is not an image generator. Unlike the existing Diffusion Model Simulator (a noised/denoised point-cloud picture with one beta knob), this lab compares noise schedules, DDIM vs DDPM and guidance against the true density.