loupe
Give your AI a magnifying glass. When you send a big image to an AI it gets downscaled on the way in — so fine print and small details turn into an unreadable blur. Loupe splits the image into tiles small enough to keep every detail sharp, and writes a manifest that tells the AI how to read them.
Why
Vision models have a resolution ceiling. Any image you send is scaled down to fit it before the model ever sees it — so a 4K screenshot, a scanned document, or a dense UI panel loses exactly the small text you wanted read.
The problem
Send an 8000×8000 page of small print as one image and it collapses into a grey smudge — the AI can't read a single line. The detail is gone before the model can look at it.
The fix
Cut the image into tiles that each stay under the downscaling threshold. Every tile is read at full sharpness, the AI stitches them back together, and fine print down to a few pixels tall becomes legible.
How it works
Drop an image. Loupe measures it and, if it's larger than the tile size, cuts it into
an overlapping grid plus a README.md written for the AI.
| Step | What Loupe does | Output |
|---|---|---|
| Measure | Checks the longest edge against the tile size (default 800 px) | tile or pass-through |
| Tile | Cuts a grid of cols × rows, each ≤ tile size, with ~10% overlap so text on a seam survives |
tile_NN_rXcY.png |
| Manifest | Writes a README.md describing the grid, order and overlap — for the AI, not for you |
README.md |
| Hand off | Copies the folder path to your clipboard | path on clipboard |
If the image already fits, Loupe makes no folder — it just tells you to drag the original straight into the chat. The overlap means text sitting on a tile boundary appears whole in at least one neighbour, so nothing is lost at the seams.
Using it
Because the tiles and the manifest live in a folder on disk, you point your AI at that folder — no need to explain anything to it.
- Drop an image onto Loupe (or click to choose one).
- Loupe tiles it and copies the folder path to your clipboard.
- In your AI chat, paste the path and ask your question:
What's the fine print in the bottom corner? /Users/you/Downloads/receipt_loupe - The AI reads the
README.md, opens each tile sharp, and answers.
Loupe ships a live view: as it tiles, you watch the grid light up over a preview of your image. Tile size and overlap are adjustable in Settings (⌘,) and remembered between runs.
Requirements
- macOS on Apple Silicon (arm64). No universal binary at the moment.
- Works with AI that can read files on your computer — Claude Code, Cursor, Claude Desktop with file access, and similar local agents.
- Does not work with web chats (claude.ai, ChatGPT.com): they can't open a folder on your disk, so a pasted path means nothing to them.
- No dependencies — tiling is built in (Rust
imagecrate). Nothing to install. - Signed (Developer ID) and notarized — no Gatekeeper override needed.