macOS
Nodebook
AI chat on an infinite canvas. Branch a conversation instead of losing it.
CLI
Mosaic
Splits an image into tiles so a small vision model can actually review it.
CLI
Splits an image into tiles so a small vision model can actually review it.
Requirements
Install
pip install git+https://github.com/mfleurival/mosaic.git
Links
What it does
Splits the image into a grid so a small model reviews one piece at a time.
A context pass, a focal pass per tile, then a pass that stitches the findings together.
Each tile is read alongside its neighbors, so nothing falls between the cracks.
Dial coverage from a quick look to a dense, every-tile sweep.
Structured reports you can read or hand to another tool.
Overview
Hand a 4B vision model a full screenshot and it describes the gist. Details go missing, because the whole image gets squeezed into a handful of tokens.
Mosaic cuts the image into a grid and walks it: one global pass for context, a focal pass on each tile plus its neighbors, then a synthesis pass that puts the findings back together. The model never has to see everything at once, so it stops skimming.
Usage
mosaic review screenshot.png --model qwen/qwen3-vl-4b --intensity dense
Points at LM Studio on localhost:1234 by default. Any vision-capable model works.
