Diffusion Model Simulator

Generative diffusion models such as DDPM, DDIM and Stable Diffusion learn to reverse a simple noising process. This simulator visualizes that process directly on a grid of coloured points forming a clean pattern: the forward process progressively mixes Gaussian noise into every point using the closed-form formula x_t = √(ᾱ_t)·x₀ + √(1−ᾱ_t)·ε, dissolving the pattern into static, while the reverse process replays the same sequence backward, reconstructing the pattern from pure noise. Scrub the timestep slider to inspect any point in the process, switch direction, or let it play continuously in both directions while tuning the number of diffusion steps and the steepness of the β noise schedule.