Diffusion on a 2D dataset
Dots are samples. Blue shading is the noisy data density at time t; arrows show its score, the direction toward higher density.
Time t1.000
Signal kept, sqrt(abar)0.007
Noise, sqrt(1 - abar)1.000
Try this
- Press Denoise and watch when the shape actually appears. Most of the structure arrives in the last part of the run.
- Switch between Deterministic (ODE) and Random (SDE) and replay. Both match the data; only the paths differ.
- Choose Add noise and scrub: every dataset becomes the same round cloud.
The data here is a mixture of Gaussian blobs, so the score of every noisy version is known exactly. Real diffusion models learn an approximation of it with a neural network.