Background Remover
Cut the background out of any photo, without uploading it anywhere. The model runs in your browser, so the image never leaves your device.
Why "no upload" is the whole point
Nearly every background remover on the web works the same way: you hand your photograph to a company, their server processes it, and you get a result back. That is a reasonable engineering choice and a genuinely awkward privacy one. The images people most want to cut out are the ones they are least casual about — staff headshots, product shots before launch, photographs of children, identity documents, someone's face for a slide.
Here the model comes to your image instead. The network is downloaded to your browser once and executed by your own processor. Your photograph is read from disk into a canvas, turned into numbers, and the numbers stay in the tab. Traffic goes one way only: weights come down, the picture never goes up.
It is worth being exact, because "private" is a word people use loosely. The model files are fetched from Hugging Face, which means they see an IP address, exactly as any site you visit does. What is never transmitted is the picture, and no server here ever holds it — so there is nothing for anyone to log, retain, train on, or leak.
The remaining trade-off is real and worth stating. A server can run a model ten times larger than anything sensible to download, and larger models cut better hair. What you get here is a model that fits down a wire in under a minute, is honest about where its limits are, and does not ask you to decide whether you trust a stranger with the file.
How the cutout is actually made
The model is a salient-object detector. It does not know what a person or a chair is; it was trained to answer a narrower question — which pixels belong to the thing this picture is about? That framing explains most of its behaviour. Give it one clear subject and it is confident. Give it a group photo, a shelf of similar products or a landscape with no focal point, and it has no basis on which to choose. That is what the magic brush is for.
Your image is scaled to the 512-pixel square the model was trained at and passed through the network, which returns a score for every pixel. Those scores become a greyscale matte: white where it is certain this is the subject, black where it is certain it is not, and grey along every boundary where it is genuinely unsure. That band of grey is the interesting part, and it is what the two edge controls act on.
Cut tightness decides how much of the uncertain band to keep. Push it up and only confident pixels survive, giving the crisp, decisive edge you want on a bottle or a box. Pull it down and more of the doubtful boundary stays, which is right for hair, fur and motion blur, where the true edge really is partly transparent. Edge softness then blurs the matte slightly, so the cutout has a believable transition instead of the aliased staircase that makes a composite look pasted together.
The matte is finally scaled back up to your original resolution and written into the alpha channel of your own pixels. The model only ever saw a small square, but the pixels you download are the ones you supplied — which is why the output is full resolution even though inference happens at 512.
Getting a better result
Separate your subject before you shoot, if you can. The single biggest factor is contrast between the subject and what is behind it. A dark jacket against a dark wall is hard for any model, including expensive ones. Half a metre of distance between subject and background, or a slightly brighter backdrop, does more for the cutout than any slider.
Reach for the magic brush before the sliders. If a whole region is wrong rather than the edge, one click on that object fixes it properly, where an edge slider can only trade one kind of error for another.
Crop to one subject. If two people are in the frame and you only want one, crop first. Saliency detection will usually try to keep both, and no amount of edge tuning will make it choose.
Check the cutout on the colour you will actually use. Edges that look clean on the checkerboard can show a bright fringe once composited onto a dark background, because the original background colour is still faintly present in the semi- transparent pixels. Switch the background to your real one and judge it there — that is why the colour presets sit beside the edge controls rather than after the download button.
Trim only at the end. Trimming crops away the transparent margin, which is useful for a logo or an asset going into a layout, and unhelpful if you need the subject to stay in position relative to the original frame.
Frequently asked questions
No. The model is downloaded to your browser and runs there, on your own processor. Your image is read from disk into a canvas and is never sent over the network. What does leave your machine is a request for the model file itself — nothing more.
Both, and the distinction is the whole point. Traffic goes one way: weights come down, your picture never goes up. The model files are fetched from Hugging Face, so they see your IP address the way any website you visit would. Nothing else about you is sent, and your photograph never is.
Running a model locally means having the model locally. The background model is about 94 MB, the runtime that executes it about 13 MB, and the magic brush a further 13 MB if you use it. None of it is fetched when the page loads — only when you actually remove a background, so nobody who is merely reading the page pays for it. Your browser then caches all of it, and every later image starts immediately, including on future visits.
A second, much smaller model that answers a different question. The background model looks at a picture and gives you its own opinion, which you cannot argue with. The magic brush takes your click as an input: point at an object and it works out that object’s edges, so you can add something it dropped or cut away something it kept. It costs a 13 MB download the first time and a few seconds to read each picture, after which every click takes well under a second.
Because the background model has no way to hear the correction. It is a saliency detector: it looks at a picture and restates its own opinion, and your corrections are precisely the places where that opinion was wrong. Feeding it its own corrected output simply reproduces the original answer and throws your edits away — we measured exactly that, which is why the magic brush uses a promptable model instead.
Lower the cut tightness and raise the edge softness a little. Tightness controls how much of the uncertain boundary survives, and hair lives almost entirely inside that uncertain band, so a tight cut removes it. If whole clumps are missing rather than fine strands, the magic brush will usually bring them back in one click.
None of the three, and there never will be. This is exactly the part of background removal that is normally used to force an account: a free preview at low resolution, then a paywall for the full-size file. You get the full resolution of whatever you put in, every time.
The same resolution you supplied, up to 4096 pixels on the longest edge. The model itself only ever looks at a 512-pixel square, because that is what it was trained at, but the resulting matte is scaled back up and applied to your original pixels. The output is not a 512-pixel image.
Yes. Choose one of the presets or pick any colour, and the cutout is composited onto it before download. Leave it on transparent and you get a PNG with a real alpha channel, ready for a design tool, a slide or a website.
Yes, but the first run is a real cost on mobile data: about 94 MB before anything happens. Once cached it stays usable and later pictures start immediately. If a very large photo makes the tab run out of memory, scale it down before you start.
Yes. The background model is BiRefNet-lite under the MIT licence, the magic brush is SlimSAM under Apache 2.0, and this site is MIT — so there are no rights for anyone here to assert over your output. Whatever rights you held in the original photograph, you still hold.
No, deliberately. Every other tool here is callable over REST and MCP because the logic is a pure function we can run on a server. This one is a neural network, and running it for you would mean you uploading your image to us — which is precisely what this tool exists to avoid.
The Zeal Promise
Non-negotiable, on every tool we ship.
- No sign-up — ever
- No watermarks on any output
- No ads near the download button
- Works fully on mobile
- Privacy-first: processing stays in your browser
- Open source (MIT) — audit it, contribute, self-host
- Zero runtime dependencies in tool logic