Drop in a photo and see what a 21 MB vision model finds in it: objects, text and the scene, drawn over the image. It runs in this tab, so the photo never leaves your device.
The first run downloads the model (about 23 MB) and keeps it in this browser for next time.
Mirage is a small vision model from Cactus Compute. One pass over an image returns a scene summary, named objects with boxes and outlines, the text it can read, and how the objects relate. Scores run from 0 to 1.
The engine is Cactus's WebAssembly runtime (1.3 MB, served from this site). The weights (21.6 MB) come from Hugging Face on the first run, are checked against a known SHA-256, and are then kept in this browser. Your image is decoded, scaled to 1024 px on its long edge and handed to the model in a background worker; no part of it is sent anywhere.
Engine and weights are Apache-2.0, © Cactus Compute. This page is our own take on Joe Maddalone's mirage-demo.