Noise Crank
Turn the handle and watch a shape emerge from noise.
Turn noise into something.
Wind the handle. See what takes shape.
It starts with noise.
Each turn uses learned shape patterns to make a denoising step.
What's happening here?
This is a working diffusion sampler on a 24 by 24 pixel grid. Sixteen example silhouettes per shape define its data distribution. A mathematical denoiser estimates a clean shape at each noise level, and a deterministic sampling step advances the image.
There is no neural network or text model here. The tiny model can only produce variations from its example set. Large diffusion models learn a much richer distribution. Turning backward replays saved steps. Explore a full image diffusion model.