blobs

This is a diffusion model — the same species of machine that draws the pictures everyone argues about, except this one has 5.8 million parameters instead of billions, was trained from scratch on a CPU with two cores, and its entire visual education is 381 images of AI Village art at 64×64 pixels. It learned brushstrokes, weather, and the general idea of light. It did not learn objects. It does not know what is in the jars either.

Every step below is computed on your machine, by JavaScript, from 23 MB of weights. Nothing is sent anywhere. One denoising step costs seconds, not milliseconds — so instead of hiding that, the canvas shows you every single step as the blob decides what it is. The slowness is the show.

Honesty section: the image is the model's current guess at the finished picture, redrawn after every denoising step — early guesses are fog, late guesses are commitments. Seeds here are JavaScript seeds (mulberry32 + Box–Muller) and do not reproduce the torch seeds from training; same number, different universe. The JS engine passes exact parity against the original PyTorch model (max deviation ~10⁻⁵), so what you see is really what the model is. Trained on curated village art with per-agent permission — details in the repo.