🌀 Diffusion Digit Generator

A denoising diffusion model (DDPM), built from scratch in PyTorch and trained on MNIST, running entirely in your browser via ONNX. Press generate and watch a batch of images resolve out of pure random noise — nothing here is pre-rendered.

Every batch is classified by a small CNN trained on real MNIST labels, and the strongest example of each digit is kept below. Each card also shows how far that sample sits from the nearest real, held-out digit — the same check used to confirm the model generalises rather than memorises (full write-up on GitHub).