toolsmith

No upload · everything runs in your browser

Cut out an object by clicking

Point at the thing you want and it gets separated from everything else. Add more points to refine it. The model runs on your device — the picture never leaves this tab.

Drop a photo here

PNG · JPG · WebP — one at a time

Preparing a photo downloads about 18.1 MB of runtime and model. It is fetched once and kept by your browser — the next photo only needs the reading.

How is this different from the background remover?

The background remover decides for you: it finds the one thing it judges most prominent and keeps that. Here you decide. If a photo has three objects and you want the second one, or you want the lamp rather than the person, this is the tool. If there is one obvious subject, the background remover is faster because it needs no clicks.

Where does my photo go?

Nowhere. It is opened inside this browser tab and never uploaded. The model comes down to you once and your browser keeps it.

Why does the first click take a few seconds?

It does not — the wait happens before it. The model reads the whole picture once and builds a summary of it, which is the slow part, and after that each click only consults that summary. We measured 6 seconds for the reading and a tenth of a second per click on a CPU. That is why the button says prepare, and why the picture has to be prepared again if you swap it.

The selection grabbed too much or too little.

Add another point. Switch to remove and click on a part that should not be included, or stay on add and click another part of the same object. The model returns three candidate shapes each time and we keep the one it scores highest, so a second point usually settles it.

Which model is this?

SlimSAM, a compressed version of Meta's Segment Anything, both under Apache-2.0. The compressed one is fifteen times smaller than the original — 9MB against 359MB — which is the difference between a tool that starts in seconds and one nobody waits for.

Other tools

Worth reading