Blog ·

Mirage online: try Cactus' 22 MB vision model on your photos

Mirage describes a photo, outlines every object, reads text and explains relations, in 22 MB. What it returns, its Apache-2.0 license, how to try it in your browser and how to run it locally.

Mirage from Cactus Compute is a vision model small enough for a phone: 51 million parameters, about 22 MB. In a single pass it describes the scene, finds and outlines every object, reads the text in the picture and explains how things relate to each other. The weights are open (Apache-2.0), and on LiveWeights you can try it on your own photos in a few seconds, on a plain CPU.

MirageCactus Compute · Apache-2.0A 22 MB vision model that describes a photo, finds and outlines every object, reads text and explains how things relate, on a plain CPU.Try it online →

What Mirage returns

  • The scene: what kind of place it is, the lighting, the background and the style of the photo.
  • Objects: a name, a box, an outline and attributes such as colour, material and shape. Names come from the language model, not from a fixed list of classes.
  • Text: the writing in the image, with a box for each line.
  • Relations: which object sits on, holds or is attached to another, plus a short reasoning trace you can check.
  • Seven languages: descriptions in English, Italian, German, French, Spanish, Dutch and Polish.

License: can I use it commercially?

Mirage and the Cactus runtime are released under Apache-2.0, so commercial use is allowed. The annotated images and descriptions you get on LiveWeights are yours.

How to try it online

  1. Open the Mirage page and sign in, then pick a small credit pack (no subscription).
  2. Upload a photo: a desk, a street, a shop shelf, a product shot.
  3. Pick the language of the description and press Describe.
  4. You get the photo with every object outlined and labelled, and the description next to it. Copy JSON gives you the full structured result.

Tips for better results

  • Mirage is built for speed on small devices: it is great at everyday objects and scenes, less precise on tiny details.
  • Detection scores are shown next to each label: low scores are guesses worth double-checking.
  • For long documents or dense text, a dedicated OCR model like Unlimited OCR 2 reads more accurately.

Run it locally

Install the Python package with pip install cactus-compute and call cactus.Mirage().describe("photo.jpg"): the engine and weights download on first use. Native runtimes are available for Linux, macOS, Windows, Android and iOS, and the model even runs in the browser through WebAssembly. Cactus reports about 0.1 seconds per image on a recent phone.

LiveWeights is not affiliated with the model authors or with the AI Search channel. Facts come from the official pages linked on each model page.