DinoDreamer

DinoDreamer

DINOv3 is a vision model from Meta. It cuts a picture into patches and describes each one in 768 numbers. DinoDreamer reads those descriptions with a basis collection: 16,384 small learned spaces, three numbers wide, called basis groups. For a basis group it trains a cellular automaton, a grid of cells that all run the same small network, until the image the cells grow from noise is one DINOv3 reads as that group. Browse the groups in the Index, watch the Automata and draw on them, Train one on your own device, or read How it works. Everything runs in your browser with WebGPU.