Models

Collagic can use a model on your machine, or one you name. A check tells you whether it answered and how long it took. It does not rank models.

What each one is for

• Magic select. A cut-out on this machine (MobileSAM). It is downloaded into the app. It does not send the picture out.
• Captions, tags, and search. Ollama on this machine, if it is running. Different model files do different jobs. A vision model reads a picture. An embedding model is for search. They are not interchangeable.
• Sharpen and a short edge fill. Not a downloaded model. They are small local passes. The original file stays.
• A model you bring. Optional. An endpoint and a key you set in the Mac app. The key stays in preferences on that machine. Collagic does not receive it.

A check, not a leaderboard

In the Mac app, Settings → AI → Check the models on this machine. You get a yes or no, a place (this machine, or the endpoint you set), and a time. A fast answer is not a better collage. A refused key means the endpoint said no. That is the calibration Collagic will stand behind.

Published benchmarks on the web measure other tasks, on other hardware, with other prompts. Copying those numbers here would pretend Collagic ran them. It did not. Read a model's own card if you want that literature. Then run the check on your machine before you trust it with a picture.

Choosing

• Leave the endpoint empty if you want nothing to leave the machine.
• Use a vision model for captions. Use an embedding model for search.
• Use magic select for a cut-out. Do not expect a caption model to cut.
• If a check fails, the collage tools that do not need that model still work.