Remove background
Drop a photo in. A matting model runs on this device, cuts the subject out with a soft edge, and hands back a full-resolution file with a real alpha channel.
The short answer
To remove the background from an image, add the photo, wait for the AI model to load the first time, and press the button. The model separates the subject from everything behind it and writes the result as a PNG with a transparent background, or on a colour if you pick one. Everything runs in this browser tab on your own device, so the photo is never uploaded and the export comes back at the resolution you brought.
Drop a photo, click to browse, or paste one from the clipboard. The first run downloads the model and says how far along it is; after that it is loaded and the cut-out starts immediately. The checkerboard behind the result is what transparency looks like.
Presets
What goes behind
Cut-out quality
Save as
PNG and WebP carry a real alpha channel. JPG does not, so a cut-out saved as JPG lands on a solid colour.
AI model
Checking what this device can run.
Checking what this device can do. Nothing is uploaded.
Add a photo and run the tool. The download appears here.
Try another tool on the same photo
The short version
A background removal is two separate problems wearing one name. The first is finding the subject, which is what the model does, and it is now genuinely good: a person, a product, a car, a plant against a room will come out cleanly and the edge will look like an edge rather than a cut. The second is what happens in the half-transparent pixels along that edge, and that is where cut-outs still go wrong.
Those pixels contain a mixture of the subject and the old background, because that is what a camera recorded. Put them on white and the mixture reads as a grey halo. Put them on black and it reads as a bright fringe. There is no setting that makes the problem go away, only choices about which side of it to err on, and this tool errs towards cutting a fraction inside the fringe when the result is going onto a solid colour, and keeping every partial pixel when it is going out as transparency.
The first run fetches the weights and stores them in the browser's cache. Every run after that, and every other background tool on this site, reuses them with nothing to download. That is why the second cut-out feels like a different product from the first.
The model returns how much each pixel belongs to the subject, from nothing to all of it. Hair and fur live entirely in the middle of that range, so the values are kept rather than rounded to a hard yes or no.
JPG has no alpha channel. Saving a cut-out as JPG puts it on a solid colour, which is right for a marketplace listing and wrong for anything you intend to layer over something else.
The model reads a bounded copy because it resizes internally anyway. The mask is then scaled up and applied to the original, so a phone photo comes back at phone-photo resolution rather than at the model's.
A sharp subject against a plainly different background cuts out beautifully at two megapixels. A subject the same tone as the wall behind it is hard at forty. If you can choose the photo, choose the one with contrast at the edges.
Glass, water, chain-link, lace and hair over a matching background are the known limits. The tool will produce something; it will not produce magic, and knowing which of those you are looking at saves an afternoon.
Step by step
Five steps, all of them inside this page. The photo is read from disk into this tab and stays there.
Drag a file onto the panel, click it to browse, or paste an image from the clipboard. The tool reads the file, applies any camera rotation, and shows the real dimensions.
The first time, the AI model downloads and the panel counts the megabytes as they arrive. It then spends a few seconds setting itself up on your graphics card, which is silent but not stuck. This happens once per device.
Leave it on nothing for a transparent PNG, or pick a colour if the photo is heading somewhere that wants a plain background. Pure white is the first swatch, which is what most marketplace rules mean.
Press the button. The model finds the subject, the mask is softened by the amount set in the edge control, and the result is composited against the full-resolution original.
Look at the hair, the thin parts and anywhere the subject touched something the same colour. Adjust the edge softness or the quality setting and run it again if you need to, then download.
Two different networks, and the honest summary is that one is nine times the download for a better edge. The detailed model is the one to use on hair, fur, thin straps, plant leaves and anything with a complicated outline against a busy background. It is about 115 MB and it wants a graphics card.
The fast model is about 13 MB and was trained on portraits, so it is genuinely strong on people and noticeably weaker on objects with holes in them. On a phone, on a metered connection, or on a machine with no WebGPU support, it is the difference between a working tool and a tool you give up on, which is why it is what this page picks automatically in those cases.
Leaving the setting on pick for me reads the device and chooses. Overriding it is worth doing when you disagree with that reading: a desktop on a tethered phone might want the small model, and a difficult product shot on a laptop is worth the big one.
The model runs through WebGPU where the browser has it, which means the work happens on the graphics card and a cut-out takes a fraction of a second. Where WebGPU is missing or the driver refuses the model, the same weights run through WebAssembly on the processor instead. That path is slower, measured in seconds rather than fractions, and it produces the same picture.
The badge says which one you got because the speed difference is large enough to be confusing otherwise. If the model fails to start on the graphics card, the error offers the processor path directly, and that choice then sticks for the run as well as for the load.
Nothing about either path needs special headers on this site, which is deliberate: the isolation that multi-threaded WebAssembly requires would break embedding and cross-origin images, and the graphics path is faster than threads anyway.
Start with the detailed model. The gap between the two is widest exactly here, and no amount of adjustment afterwards recovers detail the smaller network never found.
Then use the edge softness rather than fighting it. One or two pixels of feather blends the cut-out into whatever it lands on and hides the disagreement between the mask and the individual strands. Zero is the right answer for a logo, a screenshot or anything with a genuinely hard outline, and the wrong answer for a person.
If a grey halo persists on white, the cause is the old background still present in the edge pixels. Cut tighter by setting softness to zero, or, when the photo allows it, go back and shoot or choose a frame where the subject is not the same tone as what is behind it. That single change does more than every control on this page put together.
Yes, and it is checkable. Open your browser’s network panel, run a cut-out, and watch what leaves. On the first run you will see the model weights arriving from a content delivery network. On every run after that you will see nothing at all, because there is nothing to send: the file is read from disk into this tab’s memory, the model runs in a background thread in the same tab, and the result is written back out as a download.
That is also why the tool keeps working when the connection drops after the page is open, and why there is no queue. Read how on-device processing works for the whole pipeline in five steps.
| Property | Detail |
|---|---|
| Input formats | JPG, PNG, WebP, AVIF, GIF, BMP. HEIC where the browser can decode it. |
| Output formats | PNG or WebP with an alpha channel, or JPG flattened onto a colour. |
| Model, detailed | BiRefNet lite, fp16, about 115 MB, MIT licensed. |
| Model, fast | MODNet, fp16, about 13 MB, Apache 2.0 licensed. |
| Download | Once per device, cached by the browser. Reused by every background tool here. |
| Backends | WebGPU where available, single-threaded WebAssembly everywhere else. |
| Resolution | Full. The model reads a bounded copy; the mask is applied to the original. |
| Edge control | 0 to 8 pixels of feather. The mask itself is continuous, never thresholded. |
| Where it runs | In this browser tab, on your device, in a background thread. |
| Export limits | Full resolution, no watermark, no credit counter. |
Slide the table sideways to see every column.
Worth avoiding
JPG has no alpha channel, so the transparent part becomes a colour. The tool picks white and says so rather than letting it come out black, but if you wanted a layerable file, choose PNG or WebP.
The bar is counting real bytes and it is going to finish. Reloading restarts the fetch, though the parts already cached are kept, so a reload is slow rather than fatal.
A subject the same tone as the wall behind it is the hardest input there is. Try one with clear separation before deciding the tool does not work, then go back to the hard one.
The small model is trained on portraits and it is honest work at 13 MB. On hair against foliage, or a plant, or a bicycle, the detailed model is a different result rather than a slightly better one.
Do the cut-out on the largest version of the file you have, and resize afterwards. A mask computed on a small copy and then enlarged loses exactly the edge detail you were trying to keep.
Anything see-through is partly transparent and partly refracting what is behind it. One alpha channel cannot express that, so glassware, bottles and windows need hand work no matter which tool you use.
Questions
Every answer is in the page source, so a reader and an answer engine get the same words.
The first run downloads a matting model, which is a neural network trained to separate a subject from everything behind it. The model then runs on your graphics card, or on the processor where there is no WebGPU support, and produces a greyscale map of how much each pixel belongs to the subject. That map becomes the alpha channel of the picture. No part of the photo is sent anywhere, and after the first download the tool keeps working with the connection off.
Once. The detailed model is about 115 MB and the fast one is about 13 MB, and whichever one your device loads is stored in the browser's own cache. Opening the tool again, or opening a different background tool on this site, reuses it with nothing to fetch. If your browser is set to save data, or the device reports little memory, nothing downloads until you press the button that says the size.
Yes. The model looks at a bounded copy of the picture, because it resizes everything to about a thousand pixels internally anyway, and then the mask is scaled back up and applied to the untouched original. A 24 megapixel photo comes out 24 megapixels. There is no watermark on the export and no credit counter.
Hair or fur against a busy background of a similar tone is the hardest case, and you will see the model guess. Clear glass, water and anything you can see through is a genuinely unsolved problem for a single mask, because the honest answer is partly transparent and partly refracted. Chain-link, wire mesh and lace lose their small holes. Motion blur at the edge of a limb tends to be cut rather than kept soft.
PNG and WebP both carry a real alpha channel. PNG is lossless and larger; WebP with alpha is usually 25 to 35 percent smaller at a quality nobody can see the difference in, and every current browser reads it. JPG has no alpha at all, so choosing it means the cut-out lands on a solid colour. That is exactly what a marketplace asking for a white background wants, and it is the wrong choice for anything you plan to composite later.
Because the pixels along the edge genuinely contain a mix of the subject and whatever was behind it, and putting that mix on white leaves a grey halo. This tool tightens the mask slightly when the result goes onto an opaque colour, which pulls the cut inside that fringe. If a halo is still visible, drop the edge softness to zero and cut a little tighter, or start from a photo with more separation between subject and background.
Not yet. The batch queue runs the resize, crop, compress and convert operations across a folder today, and the matting tools handle one photo at a time. The model stays loaded between runs, so working through a set by hand is faster than the first run suggests: the second cut-out starts immediately.
The same model and the same cut. This page also offers a background colour, because most people arriving at a background remover are heading for a marketplace that wants pure white. The transparent background page has no colour option at all, so a file cannot accidentally come out opaque.
Drop it onto a colour or a gradient, blur the room behind it instead of deleting it, or keep the alpha and take it into a layout. Every one of those runs on your device and keeps the resolution you brought.