Topic · Diffusion
A score network learns the data field, walkers walk back from noise.
Training and sampling start the moment the page loads. A small MLP keeps improving its noise prediction in the background; rose walkers continuously denoise themselves from a Gaussian cloud onto the data manifold, restart, and do it again. Switch dataset, schedule or T at any moment — the chain reorganises live.
Reading the canvas
Cyan dots: the toy data distribution. The denoiser sees only noised copies of these during training — it never memorises positions; it learns the local 'which way is in?' field.
Rose dots: walkers drawn from pure 2D Gaussian noise and pushed T steps backward through the learned reverse chain, leaving fading trails behind them. Faint blue under-glow: the magnitude of −ε_θ at mid-noise — the score field the walkers ride down.
