README

How CutItOut works, what it's built on, and how it's licensed.

CutItOut is a privacy-first, in-browser background remover, cutout tool and collage maker. It is 100% client-side — images never leave your device. It's a static site with no build step and no server.

Open the tool →

Privacy model

There is no upload. When you open an image it is decoded into a canvas in your browser's memory and discarded when you close the tab. The AI model is downloaded to your device and runs there; your pixels are never transmitted anywhere. The full detail is in the privacy policy.

How the cutting works

Editing happens on stacked canvases in image-pixel space, so zoom never affects output quality.

Magic Erase

A 4-connected flood fill over squared RGB distance from the pixel you click. Pixels comfortably inside the tolerance become fully transparent; pixels just past the threshold get partial transparency, so edges come out anti-aliased instead of jagged.

Smart Lasso

The interesting one. Draw a rough loop and it snaps to the object's real edges:

• Crop the loop's padded bounding box and run matting on just that crop — the loop tells the AI which object you mean, the AI supplies the edge.
• Multiply the AI's alpha by a feathered version of your drawn loop, so the final edge follows the object's contour rather than your shaky hand.
• Flood-fill outward through AI-foreground pixels to recover parts the loop accidentally clipped — a chin, hair, an ear.
• If the model can't load, fall back to colour-based region growing rather than the raw polygon.

Undo history

History stores the alpha channel only as a Uint8Array. Every edit touches alpha and nothing else, and the pristine RGB is kept separately, so alpha alone fully describes a state — at a quarter of the memory of full RGBA snapshots.

The first time you use an AI tool

Because the AI runs on your device rather than on a server, your browser has to fetch it once before the first cutout. It's a sizeable one-time download, after which it's cached — every cutout after that is fast and works offline.

Credits

Subject detection uses BiRefNet (MIT) running via Transformers.js (Apache-2.0). Both are open source and run entirely on your device.