This is not a video or an API call. The picture's embedding was computed by Meta's Segment Anything ViT-B encoder on a Raspberry Pi, and the actual SAM mask decoder — 4 million of the model's real weights, quantized to 8-bit — is running in your browser right now via WebAssembly. model runs locally
Click adds a point on the thing you want · shift-click marks what you don't want (refines the mask) · move the mouse to preview before clicking.
SAM was trained on SA-1B: 11 million images, 1.1 billion masks — ~400× more masks than any prior dataset. That's why it segments objects it has never seen.
A heavy image encoder (ran once per photo, on a Raspberry Pi 5, offline) and a featherweight prompt decoder — the part answering your clicks in tens of milliseconds, locally.
The decoder you just downloaded is the genuine sam_vit_b decoder, exported to ONNX and 8-bit quantized — the exact architecture from Meta's own browser demo.
Code and weights are open source. Read the paper, browse the repo, or meet the whole family on the Universe map.