Background remover
Cut the background off a photo without sending it anywhere. The model is downloaded once and runs inside this tab, so your picture stays on your own machine and comes back at the size you started with.
- Drop the photo inNothing is uploaded, so nothing waits on a queue.
- Say what is in itA person, anything else, or one flat colour to erase.
- Tidy the edge, saveA transparent PNG at your photo's own resolution.
Drop a photo here, or click to choose
PNG, JPG, WebP, AVIF or GIF, at any size up to 40 megapixels. You get the cutout back at the resolution you dropped in.
Free · no sign-up · nothing uploaded. The model runs in this tab, so the photo stays on your device. The weights are fetched once from a CDN and then cached, and after that the tool works offline.
How to remove the background from a photo
Drop your photo in
One file, dragged onto the box or picked from it. There is no upload step, so a 20 MB photo opens as fast as a small one.
Say what is in the picture
A person, or anything else. On a flat white or grey backdrop, switch to the colour mode and click the background instead.
Tidy the edge, then save
Edge pulls the boundary in if there is a halo, Feather softens it. Save a transparent PNG, or drop a colour in behind it first.
Why this one does not upload your photo
Almost every background remover works the same way. You hand it your picture, it travels to a machine you do not control, a model there produces the cutout, and the result travels back. That is not a criticism of the maths. It is a description of where your photograph goes, and for a headshot, a passport photo, a child, a piece of unreleased product artwork or a client's file under NDA, it is the only fact about the tool that matters.
Here the model comes to the picture instead. A segmentation network is a file, and a browser can download a file and run it. So the weights arrive once from a CDN, your browser caches them, and every cutout after that is arithmetic happening in the same tab your photo is already open in. There is no endpoint for the image to be posted to, because none was written. That is the difference between a promise not to look and not having the picture.
The honest footnote: the weights themselves are a real download, from Hugging Face's CDN, and that is a request to a third party. It carries the model's name and nothing about you or your file. After it lands you can turn the network off entirely and the tool still works.
The 500-pixel free tier
The tool ranking first for this search will cut out your photo for free and then offer you the result at 500 by 500 pixels. Your actual resolution sits behind a paid unlock. This is not a trick, it is the business: a server did the work and somebody has to pay for the server, so the free tier gives you enough to see that it worked and not enough to use.
There is no equivalent here because there is no server bill to recover. A 4000 by 3000 photo saves as a 4000 by 3000 PNG. The model runs at a fixed smaller size, and the mask it produces is then scaled up and applied to your original pixels, so what you save is your own file with an alpha channel rather than a downscaled copy.
Two models, and how to choose between them
| Model | Size | Good at | Bad at |
|---|---|---|---|
| People and portraits | 13 MB | Headshots, full-length shots of a person, hair against a clean background | Anything without a person in it, because it was trained only on people |
| Anything else | 84 MB | Products, animals, food, packaging, illustrations, objects on a busy backdrop | Speed. It takes about eight seconds a picture against the portrait model's one |
The split is a difference in what each was taught, not a fast setting and a slow one. The portrait model is a small network trained on a human matting dataset. Given a bicycle it does not struggle, it looks for a person, and there is not one. The general model was trained on a broad segmentation set, which is why it can find a bicycle, and it works at 1024 pixels square rather than 512, which is where its extra seconds go.
Both are convolutional networks, and that is a constraint rather than a preference. The obvious pick for the general slot is one of the newer segmentation transformers, and they do not run here at all: at 1024 pixels square the attention layers exhaust the memory a browser gives WebAssembly, and the runtime gives up before producing anything. That was measured rather than assumed, at two precisions, at three input sizes and on WebGPU. A model card will not tell you this, and neither will the file size.
The colour mode is not a lesser version of the AI
It is the right answer to a different question. A model looks at a picture and decides what the subject is. A scanned logo, a signature, a line drawing or a product on seamless white has no subject to find, so the model produces a confident guess and the guess has soft, wrong edges. Erasing one colour has no opinion about the contents at all, so on a flat backdrop it is both exact and instant, with no model to download.
The part that makes it worth using rather than a plain "make white transparent" filter is that it floods inward from the border of the image. Only white that is connected to the edge of the frame is erased. The white of a shirt, a tooth, a highlight or the inside of a letter O stays, because none of it touches the border. A filter that simply matched every white pixel would punch holes through all of those, which is what most one-click transparency tools actually do.
Edge and Feather, and the halo they exist for
A cutout that looks slightly wrong almost always has a pale fringe around it, and the fringe is the old background. Along any real boundary there are pixels that are genuinely part subject and part background, and the model marks them partly foreground, which is correct and also means their colour comes along. Against a bright backdrop that reads as a glowing outline the moment you put the cutout on something dark.
Edge moves the boundary. Pulled left it stops counting that outermost ring as subject and the fringe goes with it. Two or three notches usually does it, and going much further starts removing real edge, which shows up in hair first. Feather goes the other way and softens the boundary, which is what you want when the cutout is going over a photograph rather than onto flat colour. Both work off the mask the model already produced, so they respond instantly and nothing is re-run.
What it cannot do
Glass, smoke and motion blur. These need a real matte with partial transparency through the middle of the object, not just at its edge, and neither model produces one. A wine glass comes back as a glass-shaped silhouette with the background trapped inside it.
Fine hair against a busy background. Individual strands survive against a clean backdrop and turn to mush against a detailed one. That is a limit of every automatic tool at this size, paid ones included, and no slider here will recover it.
Video, and batches. One image at a time. Video would mean holding hundreds of frames of raw pixels in a tab, and a batch of forty photos through the general model is a request to run your laptop's fan for ten minutes. If you need a still out of a video first, the frame extractor will get you one at full resolution.
Which format to save it as
PNG is the right default and the one to use if you are unsure. It is lossless and every piece of software made in the last twenty years reads its transparency correctly. The file is large, because a photograph compressed losslessly always is.
WebP keeps transparency too and is usually a third of the size, which makes it the better choice for anything going on a web page. Safari cannot write one, so on an iPhone or a Mac in Safari the option is not offered rather than silently handing you a PNG with the wrong extension.
JPG has no alpha channel at all. Saving a transparent cutout as a JPG flattens it onto white, so it is only useful once you have put a background colour behind the subject. The tool says so before you save rather than after.
Frequently asked questions
Which of the two models should I pick?
Pick by what is in the picture, not by which sounds better. The portrait model is a small network trained on photographs of people, so on a headshot or a full-length shot of a person it is excellent, and it is 13 MB. Hand it a shoe and it does not do a worse job, it does the wrong job: it was never taught what a shoe is, so it hunts for a person and finds nothing. The general model handles products, animals, food, packaging and illustrations, and it costs 84 MB and about eight seconds a picture to get there. The tool opens on the portrait model because most photos people bring here have a person in them. Switching models re-runs the cutout on the picture you already have open.
What is downloading the first time I use this?
The model weights, and nothing else. They come from a public CDN, they are the same file for everyone, and they say nothing about you or your picture. 13 MB for the portrait model, 84 MB for the general one. Your browser stores them, so the second cutout starts immediately. To be exact about where: the weights and the runtime that executes them go into Cache Storage, which is the persistent kind, and the small library that drives them is held in the ordinary browser cache. In practice that means the tool keeps working with the network off, and if you have cleared your browser data since, it fetches again. The photo itself is not part of that traffic in either direction: it is decoded to pixels in this tab and the pixels never leave it. If you want to check rather than take our word for it, open your browser's network panel and watch. You will see requests going out for the weights and no request carrying your image.
Do I get the cutout at the full size of my photo?
Yes, always, and this is the thing worth checking against anything else you have used. The most popular tool on this search gives you 500 by 500 pixels for free and puts your real resolution behind a paid unlock. Here, a 4000 by 3000 photo saves as a 4000 by 3000 PNG, because there is no server bill to recover and therefore nothing to meter. The model itself works at a fixed smaller size, 1024 pixels square for the general model and 512 for the portrait one, then the mask it produces is scaled back up and applied to your original pixels. So the file you save is your own picture with an alpha channel, not a downscaled copy of it.
The edge has a faint halo of the old background. Can I fix that?
Yes, that is what the Edge slider is for. A halo happens when the subject was shot against something bright: the outermost ring of pixels is genuinely a mix of subject and background, and the model quite reasonably calls a partly-covered pixel partly foreground. Pulling Edge to the left stops counting that outer ring as subject, and the fringe goes. Two or three notches is usually enough, and pulling too far starts eating the real edge, so watch the hair. Feather does the opposite job: it softens a hard boundary so the cutout does not look pasted on when you put it over something else.
It made a mess of my product photo. What now?
Two things to try, in order. If you were on the portrait model, switch to the general one, because a product is exactly the case the portrait model cannot see. If the product is on plain white or plain grey seamless, switch to the colour mode instead and click the background. That mode does not try to understand the picture at all, it just erases one colour, which on a flat backdrop is not a worse answer than the model. It is a better one, and it is instant.
Why is there a mode that just erases one colour?
Because "remove the white background" is a different job from "cut out the subject", and every tool that offers only the second one does the first one badly. A scanned logo, a signature, a diagram or a product on seamless white has no subject for a model to find, so the model guesses, and a guess on flat white looks like a bad cutout. Erasing the colour has no opinion about what the picture contains. The mode here floods inward from the border rather than matching every pixel, which is the part that matters: the white behind your subject goes, and the white of a shirt, a tooth or a highlight inside the subject stays, because those are not connected to the edge of the frame.