Free background remover for photos and artwork

Two kinds, and it picks the right mode for you. Photos: a person, a pet, a product or a car, cut out of its surroundings. Artwork on a plain background: logos, clipart, scanned line drawings and designs, where it lifts the design off the backdrop and keeps the gaps inside letters.

Yes, when the person stands out from the background in colour, as in most portraits, pet photos and product shots. It learns the background colours from the top and sides of the picture and the subject from the middle, then finds the edge between them. When it gets something wrong, paint over it with Keep or Remove and it runs again. That quick cut is not a trained AI model, so a subject that matches its background, touches the top of the picture or sits in front of a busy wall can come out wrong. So on a computer that can run it, a photo is cut by a trained model, Better cutout, by default: the quick cut shows first and the model replaces it, still on this computer. On a phone or a small device it stays a choice under the result. And when the cut keeps the wrong things in a particular picture, Tap what to keep lets you tap the things you want.

Better cutout is the default for photos on a computer that can run it: one with WebGPU, or one that reports 4 GB of memory or more, and is not a phone. The built-in engine makes a quick cut first; Better cutout runs a trained matting model on your picture and then gives the result the same edge softness, Keep and Remove brushes and downloads. The first time, nothing is downloaded until you press the button above the picture, Better cutout for photos: one-time download of about 169 MB, about 20 to 45 s (172 MB and about 20 s where the browser uses WebGPU; the button shows the figures for your browser, and the progress line after it uses the same size), and Not now is remembered in this browser. Once the model is in your browser, a photo is cut by it automatically: the quick cut shows at once, marked as a quick preview (the page says it is loading the model from this browser, with nothing to download), and Better cutout replaces it when it is ready, with a Show the quick cut switch back that is instant and keeps both cuts, so Show Better cutout is instant too. On a phone, or a device that reports under 4 GB of memory, it stays an option under the result, with its size and time warning first. Right after a Better result, Tap something to remove it (a hand, a cloth) and Only what I tap start the tap step, and Fix leftovers is under the result. Logos and flat artwork are never switched to it. The model is BiRefNet_lite (Peng Zheng and co-authors, MIT licence, github.com/ZhengPeng7/BiRefNet), in a 512 x 512 ONNX export by Studio Ludens (MIT). It is a one-time download of 169 MB (172 MB with WebGPU), and your browser keeps it so it is fetched once. It runs on your graphics chip when the browser has WebGPU (a few seconds a picture, 10 to 25 for the first one on a page) and on the processor otherwise (15 seconds to a minute); your picture never leaves this computer. It finds a subject that stands out as a whole object, and can find nothing in a picture with no clear subject, such as a plant filling the frame or a street scene; the page then keeps the first cut and says so. It is not run by default, and not pushed, for logos and flat artwork on a plain backdrop (a few flat colours and hard edges, no shading), because the standard cutout is better at those (it keeps fine lettering, dots and thin lines solid, which a photo model fades). A photograph is always offered it, including a very dark one or one on a plain or dark backdrop that the page opens as artwork: the page checks the picture for shading, not only the border, and the cutout of a photograph is left exactly as before. Very dark photos are shown to it as they are: brightening them first made it find less on lit pumpkins (licence and notice).

An optional step after any cutout, for the picture where the automatic cut keeps the wrong things. Press Tap what to keep, then tap or click each thing you want (each of three pumpkins), or choose Tap to remove and tap what should go (the wall behind a plant). To outline something long, such as a whole plant or a rope, choose Draw a short line (or just drag with a mouse) and draw along it. Tap the middle of the thing: when the model is surest of only a part (an eye, a face, a lit mouth), the first outline is the whole thing it belongs to, if the model scores that as a plausible object. Tap a dot again to grow its outline, and from the biggest it goes back to the smallest. A hole inside a kept thing that is joined to the outside by a thin gap (the glowing mouth of a pumpkin beside another) is filled; a real gap, like the space between an arm and a body, is not. A second model outlines exactly what you tapped, and the page adds each outline to the cut (or takes it away), then gives the result the same edge softness, Keep and Remove brushes and downloads; Only what I tap starts from nothing, for a picture whose first cut is no use. The model is SlimSAM-77 (Zigeng Chen, Gongfan Fang, Xinyin Ma and Xinchao Wang, Apache-2.0, github.com/czg1225/SlimSAM), a slimmed-down Segment Anything (Meta AI, Apache-2.0), in an ONNX export by Xenova (Apache-2.0). Nothing is downloaded until your first tap. Then about 40 MB (43 MB where the browser uses WebGPU) comes down once and your browser keeps it; the model reads your picture on this computer, which takes about 15 to 40 seconds the first time (longer on a phone), and after that each tap takes a second or less. Your picture never leaves this computer. It outlines whatever is under a tap, so it can pick a leaf where you meant the plant, or a wicker strand where you meant the basket: tap again to step to a bigger outline, or draw a line across the whole thing (licence and notice).

Two ways, both on this computer. First, the page looks for pieces that may be leftover background: after a photo cutout it checks for kept pieces that are cut off from the subject, namely a thin strip along the edge of the picture, a slab in a corner, or a small blob well away from the subject. If it finds any it outlines them on the picture and asks Found 1 piece that may be leftover background. Remove it? with a Remove button, and Undo after. It only suggests, because no rule can tell a stray strip of table from a tail, a hand or an accessory the cutout happened to cut loose: nothing is removed unless you click Remove, and a cutout is exactly what it was before until you do. It does not look at pieces that are joined to the subject, it leaves a second subject of comparable size alone, and anything you marked with a Keep brush stroke or a Keep tap is never offered. Second, Fix leftovers under the result: choose Draw to remove, draw a loop around the patch (it closes itself when you let go) or switch to Box and drag a box, and everything inside goes; Draw to keep does the same the other way and brings an area back from your original picture. A slab joined to the subject needs this second way, because the page cannot tell where one ends, and you decide what is leftover. The Remove brush, Keep brush and Tap to remove are in the same row. It works with a mouse or a finger (one finger draws, two fingers zoom), has undo and redo, needs no download, and every download uses the fixed result.

The old background bled into the edge pixels when the image was first saved. Nudge Trim the edge up by one; it gives up the outermost ring of pixels, which is usually exactly the halo. Lowering Edge softness also helps on hard-edged graphics.

Leave "only remove background around the artwork" switched on. With it on, only background connected to the outside is removed, so enclosed areas stay. Switch it off when you genuinely want every white area gone. One exception, which you can switch off: on a plain or lightly textured backdrop, the tiny gap inside a small letter (the middle of a small o) is backdrop, so it is cleared too; untick "Still clear the tiny gaps inside small letters" to keep it.

Yes. Press Pick a colour from the image and click the colour you want gone. That's the way to drop a coloured backdrop while keeping white parts of the design.

Drop the photo, then open Clean-up options under Download (it is closed until you open it). Choose the Background (transparent, white or a colour you pick), the Framing (as cut, or a square or 4:5 portrait canvas with the subject centred and a margin you set, 9% to begin with), the Size on the long side (the photo's own size, 1000, 1600, 2000 or 3000 px) and the File (JPG for a white or coloured background, PNG for a transparent one). The page says the size the file will have before you press the button. The cut-out is the one on screen, so fix it first with Keep and Remove if it needs it. The centring ignores faint stray pixels, so a speck of leftover haze far from the product does not push it off the middle. The white is exactly white (255, 255, 255), the dpi is written into the file, and the file is named after the photo with -clean on the end. With every option left as it is, the download is the same transparent PNG as always.

Only when it has to, and it says by how much. If a photo has fewer pixels than the size you chose needs for the product in it, the product is enlarged, and the line under the options says how many times (it warns above 3 times, and says most of the detail is guessed above 8). A photo that has enough pixels is not enlarged: a big photo is read again at its own size, up to 4000 px on the long side, so a 2000 px result is made from the real pixels, not from the smaller working copy the cut-out is made on. When enlarging is needed, the page uses the on-device enlarging model of the upscale tool (about 8 to 11 MB, downloaded once, and your picture still never leaves this computer) when it would not take long on your device, and a plain high-quality resize otherwise; the line under the options and the report say which. Enlarging never adds detail that was not photographed.

One picture at a time is free, with every Clean-up option, no watermark and no limit. Dropping a whole folder at once (up to 30 photos) is part of Pro: every photo gets the same steps, you see each one's progress, and you get ZIPs (split before they grow too big for the browser) with clean-photos-report.txt. The report says, for each photo, which cut-out engine was used, what size it came out and whether it was enlarged, and it lists first the photos whose cut-out may be wrong: the subject touches the edge of the picture, the subject is very small, almost nothing (or almost everything) was kept, or the picture does not look like a photo. Look at those first. The first time on a device, a visitor who is not Pro can try the folder run on up to 5 photos for free.

No. The image is decoded by your browser and the pixels are worked on by code running on this page. Open the Network tab in your browser's developer tools and watch while you use it: nothing leaves this computer. If you press the Better cutout download button, or tap in Tap what to keep, or ask Clean-up for a size that needs the enlarging model, the model file comes down to your browser, once, and once it is there it runs by itself on your photos; nothing goes up.