Background Remover

Remove the background from a photo in your browser. The model runs on your device and nothing is uploaded.

Free, no limit Nothing uploaded No watermark Runs on your device

Drop a photo here

JPG, PNG and WebP. Nothing is uploaded, ever.

Choose a file

The only one of these that does not upload your photograph

Every background remover you have used sends your picture to a server. It has to: deciding which pixels are the subject and which are behind it needs a neural network, and until recently a neural network needed a machine in a data centre.

It does not any more. This page downloads the model, about sixteen megabytes the first time and nothing afterwards, and runs it in your browser. Your photograph is read from your own disk and never leaves it. There is no upload, no queue, no account and nothing on our side to store or delete.

The model is U^2-Net, in its small form, trained for exactly this job. It produces a mask saying how much of every pixel belongs to the subject, and that mask becomes the transparency of the result.

The background being removed from a photograph, leaving the subject
How it works

Three steps, still no upload

The model is downloaded once and then runs on your own processor.

1

Add a photo

Drop it on the tool or pick it from your device. Nothing is sent anywhere.

2

Wait for the model

The first run fetches about 16 MB and takes a few seconds. Every run after that is quick.

3

Choose and download

Leave it transparent or put a colour behind the subject, then download a PNG.

Two different pictures, two different jobs

Not everything dropped here is a photograph. A logo, a banner or a product shot sits on one flat colour, and for those a neural network is the wrong instrument: it is trained to find the salient object in a scene, and given white lettering on a solid red field it will happily hand back the outlines of the letters.

A flat background needs no guessing at all. The background is that colour, and it is connected to the edge of the picture. So the border is checked first, and if it is one colour the background is keyed out exactly: every pixel that matches and can be reached from the edge goes, which is why the enclosed white inside a letter O disappears while the white of an eye stays. It takes no time and there is no model to download.

Connectivity is not the whole answer on its own. The middle of a letter P is background too, exactly the same colour as the rest of it, and it cannot be reached from the edge because the letter is in the way. Those enclosed pockets go as well, or lettering comes back with its counters filled in. What keeps that from punching a hole through a white shirt against a white wall is size: a letter counter is a fraction of a per cent of the picture and a shirt is not.

The edge of lettering is not one colour or the other, it is a blend of both, so those pixels get part of the transparency rather than all or none of it. How much depends on how far the subject actually sits from the background, measured from your own picture: a fixed threshold leaves a coloured rim on high contrast artwork, and a coloured rim is what a bad cut-out looks like.

Anything else is a real scene, and that goes to the model.

How the model decides

Your picture is squashed to 320 by 320, which is the size U^2-Net was trained at, and each colour channel is shifted and scaled the way the training data was. Squashed rather than fitted inside a square: the network saw stretched images during training, and the mask is stretched back to your picture's real shape afterwards, so nothing is cropped and nothing is lost.

The network returns a value for every one of those 102,400 positions saying how much it believes that pixel belongs to the subject. Those values are stretched to fill the range 0 to 1, scaled back up to your image, and used as its transparency.

Values in the middle become partial transparency rather than a hard yes or no. A hard cut at exactly one half gives a jagged outline on hair; too wide a band leaves a faint halo of the old background. The band used here sits between the two, which is what anybody would settle on after trying all three, so you are not asked to choose it.

The mask is 320 squares wide and your photograph is not, so each pixel reads the four mask values around its position and mixes them rather than snapping to the nearest one. That is the difference between an outline that follows the subject and a staircase along every edge.

There is nothing to set either way, and you are told which of the two happened. Drop a picture, get a transparent PNG.

All of this runs through ONNX Runtime compiled to WebAssembly, on one thread. More threads would need SharedArrayBuffer, which needs headers that would change how the rest of this site behaves, for a second or two on a photograph.

What is downloaded, and when

FileSizeWhat it is
ort.min.js0.4 MBONNX Runtime, the part that runs a model.
ort-wasm-simd-threaded.wasm10.5 MBThe same runtime compiled to WebAssembly, which is what actually does the arithmetic.
u2netp.onnx4.4 MBThe trained weights. This is the model.

All three are served from this domain rather than from a content network, so opening this page tells nobody else that you did. Your browser caches them, so the download happens on your first visit and not again.

Nothing is fetched until you drop a picture, so arriving at this page and reading it costs nothing at all.

What people use this for

Product photographs for a marketplace

Amazon, Etsy and most marketplaces want a plain white background. Photograph the item on any surface, remove the background, choose white, download. That is the single most common use and the one the model is best at.

A profile picture that has to be plain

A great many forms want a photograph against a plain background and most photographs are not. Take the transparent PNG from here and put it through our PNG to JPG converter, which fills the transparency with white: a form expecting a photograph will not know what to do with transparency, but plain white is exactly what it asks for.

A cutout for a design

Transparent PNG, dropped into a slide, a poster or a page. Check the edges at full size before you commit: an edge that looks fine at thumbnail size can look cut out at full size.

Removing something distracting

Sometimes the point is not the subject but what is behind it. A photograph taken at a desk, in a shop, or with somebody walking past. Replacing the background with a flat colour removes all of it at once.

What runs through all four is that the picture is usually one you would rather not hand to a stranger, which is the reason this page exists in the form it does.

What it is good at, and what it is not

Worth reading before you judge the result, because these limits are the model's rather than the page's, and they are the same limits the paid services have in a milder form.

  • Good: a person or a product with a reasonably clear outline, against a background that is a different colour or brightness. This is the ordinary case and it works well.
  • Good: product photographs on any plain surface, which is most of what people need this for.
  • Weak: fine hair, fur and anything wispy. This is where every one of these models struggles, and this is the small model rather than the large one.
  • Weak: a subject the same colour as what is behind it. Hard for a network for the same reason it is hard for you.
  • Weak: several subjects at different depths, where it will often keep the nearest and drop the rest.
  • Handled without the model: a logo, a banner or a product shot on one flat colour. That is not a guessing problem, so the page recognises it and cuts it out by colour instead, which is both instant and exact.

Why it is a large download

Sixteen megabytes is the runtime and the weights: about eleven for the WebAssembly build of ONNX Runtime and four and a half for the model itself. Both are cached by your browser after the first visit, so the second photograph is fast.

The alternative is a two hundred kilobyte page that uploads your picture instead. That is the trade, and it is worth being explicit about which side of it this page is on.

Why this matters more here than anywhere else

Think about what people remove backgrounds from. Photographs of themselves, for a profile or a passport application. Photographs of their children, for a school form. Product shots of something not launched yet. A picture taken inside a workplace with things in the background that should not be in it.

Every other tool that does this asks you to upload all of that. The good ones delete it afterwards and say so in a policy. You have no way of checking, and the picture has already left your machine by the time you think about it.

This page never receives your photograph, because there is nowhere for it to go. The model runs on your own processor. You can watch it work with the network disconnected once the page and the model have loaded, which is the only proof of this kind that means anything.

The cost is a large first download and a few seconds of your own machine's time. That seems a reasonable price for a photograph of your child not being on somebody else's server.

How this compares with the alternatives

remove.bg and the services like it

Better results, particularly on hair, because they run much larger models on much larger machines. They also want your photograph, and most of them want an account and a payment after a few uses. If the picture is not sensitive and the quality has to be perfect, use one.

Photoshop or an editor with a select subject tool

The best result available, because you can correct what the automatic selection gets wrong. Also a subscription and a learning curve for something people need occasionally.

A phone's built in cutout

Recent iPhones and Android phones will lift a subject out of a photograph by long pressing it, locally and very well. If you are on a phone and it offers that, it is a genuinely good option and this page is not trying to beat it.

Doing it by hand

Still the only way to get a perfect edge on hair, and still what a professional does when it matters. Everything automatic is an approximation, this one included.

If the result is not right

The hair looks chewed

That is this model's main weakness and it is real. It is the small version of U^2-Net running on your own machine in a second or two, against services running the large one on a server. If the edge on hair is not good enough here, that is the point where a paid service or a manual selection in an image editor is genuinely better, and no setting on this page would have changed it.

Part of the subject was removed

Usually a subject the same colour and brightness as the background, or an arm or object at a different depth from the rest. That is hard for a network for the same reason it is hard for a person looking at the photograph. A picture with clearer separation between the subject and what is behind it will come out far better.

There is a faint outline of the old background

A thin rim of the old colour can survive around the outline, because the pixels there were genuinely a mixture of the subject and the background before anything was removed. Placing the cut-out on a background of a similar colour hides it completely. On a very different colour, an image editor's defringe or a one pixel contract of the selection is the fix.

Nothing happens, or it fails

The first run fetches about sixteen megabytes and prepares the model, which takes a few seconds and needs a current browser with WebAssembly. On a slow connection it can take a while. If it fails outright, reload the page and try once more.

It is slow

It is running a neural network on your own processor, on one thread. A few seconds a picture on a laptop, longer on an older phone. Every other tool is fast because it does the work somewhere else, on your uploaded photograph.

My iPhone photo will not open

iPhones save as HEIC and only Safari decodes it. Convert first with our HEIC to JPG page, or set your phone to save as JPEG under Settings, Camera, Formats.

Free, with no account and no watermark

No sign-up, no credits, no daily limit and no watermark. There is no paid tier holding back the resolution, because the work happens on your computer and costs us nothing.

The result is always a PNG, because PNG is the only common format that can hold transparency. If you need it on a solid white background instead, run the PNG through our PNG to JPG converter, which fills transparency with white as JPEG has no way to store it.

If you picked a white background and now need the file smaller, our image compressor hits an exact size in kilobytes. If you need different pixel dimensions, the resizer does that. Neither uploads anything either.

FAQ

Background Remover

The questions people ask most about removing a background.

Drop the photograph on the tool above. That is the whole thing: there is nothing to choose and no button to press. It runs in your browser, takes the background off, and a download button appears with a transparent PNG.

No, and that is the point of this page. The model is downloaded to your device and runs there. Your photograph never leaves your machine, which no other background remover can say.

Because it fetches about sixteen megabytes: the runtime that executes the model and the model's own weights. Your browser caches both, so every use after the first is much faster.

No. They run a much larger model on much larger machines and get better edges, particularly on hair. They also need your photograph. This is the trade, and which side of it you want depends on the picture.

Fine detail is where every one of these models is weakest, and this is the small version running on your own device. A photograph where the hair stands against a plainer, more contrasting background comes out noticeably better.

Not on this page, and it does not need one. Download the transparent PNG and run it through our PNG to JPG converter, which fills transparency with white. That is what most upload forms want, since they cannot handle transparency.

A PNG, always, because it is the only common format that can carry transparency. If you need a white background instead, put that PNG through our PNG to JPG converter, which fills transparency with white.

Yes, and better than the model would. A picture on one flat colour is recognised as such and cut out by colour and connectivity instead, exactly rather than approximately, and with no download at all.

Yes, on a current one. It is doing real work on your processor, so an older phone will be slow and a very old one may run out of memory.

After the first run, yes. Once the model is cached, the page works with the network disconnected, which is the simplest proof that nothing is being uploaded.

Completely. No account, no credits, no daily limit and no watermark, because the work happens on your device rather than on ours.

Yes. A bg remover, an image background remover and an online background remover all describe this page. The names differ because people describe the same job differently; the job itself does not change.
More tools

What to do with it next

Same engine, different job. Every one runs in your browser.

All 26 image tools

Image Resizer

Change the pixel size of the cutout, keeping its shape.

Image Compressor

Bring the result under an exact size in kilobytes.

PNG to JPG

Flatten a transparent PNG onto white for a form that wants JPG.

Color Picker

Take a colour out of the photo to use behind the subject.

Need a different tool?

Over a hundred free tools, all running in your browser. Nothing is uploaded and nothing is capped.

Browse all tools Contact us