Background Remover

The model doing the cutting out

This page loads an open source segmentation model that runs on your own device. The library is @imgly/background-removal, pulled in as a module the first time you press the button, and its neural network weights plus the runtime that executes them are fetched as separate asset files, roughly 40 MB in total. Your browser caches those files, so the wait only happens once. The network looks at the picture and produces an alpha value for every pixel, an estimate of how much that pixel belongs to the subject rather than the background. Those values become the transparency channel of a new PNG. Nothing about the picture is sent anywhere: the model comes to the image, not the other way round. Everything happens in your browser and nothing is uploaded to a server.

Before any of that, the page checks that your browser can compile a WebAssembly module at all. If it cannot, the button is disabled and a message explains why rather than letting you wait for a failure. Firefox and Safari users see a note about tighter memory limits, because those browsers are the ones most likely to run out of room on a very large photo.

What it handles well, and what it struggles with

A person or a product photographed against a reasonably distinct background is the easy case, and hair against a plain wall usually comes out better than a hand traced selection would. It also copes with pets, plants, shoes, furniture and bottles. The harder cases are predictable once you know what the model is estimating: transparent or reflective objects such as glass and jewellery confuse it, because the background is genuinely visible through the subject; a subject wearing a colour very close to the wall behind it gives the network little to work with; fine flyaway hair against a busy background keeps some of that background; and motion blur at the edge of an arm blends the two regions together. Group shots where people overlap the frame edge can lose a limb. If the result is close but not perfect, retaking the photo against a plainer backdrop usually fixes more than any amount of fiddling.

Transparent, or a colour behind the subject

The preview shows the original and the result side by side, both on a grey checkerboard, which is how you tell real transparency from white. Tick the fill option and pick a colour, and the page redraws the cut out onto a canvas filled with that colour first, so the subject sits on a solid background. That is what you want for marketplace listings that reject transparency, for a plain backdrop behind a staff headshot, or for a thumbnail that needs to sit on a brand colour. Untick it and you are back to transparency. The download is always a PNG, named after your original file with -no-bg added, so portrait.jpg becomes portrait-no-bg.png.

A typical run

Take a 2048 x 1365 JPEG of a jacket on a hanger, about 900 KB. The first time you press the button the progress bar tracks the model download, then flips to processing and finishes in a few seconds on a desktop, longer on a phone. The output is a PNG at the same 2048 x 1365 pixels with the wall and hanger hook gone. Expect the PNG to be larger in bytes than the JPEG you started with, often several megabytes, because PNG is lossless and has to store the alpha channel. Load a second image afterwards and the model is already cached, so it starts almost immediately.

Limits and gotchas

• Files over 25 MB trigger a warning. There is no hard cap, but memory is the real limit and a failed run on a phone usually means the image was too big.
• If a large photo fails, resize it to about 2000 pixels on the long edge and try again.
• The first run needs a working connection to fetch the model. After that, processing itself does not.
• One image at a time. There is no batch queue and no folder upload.
• There is no manual brush, no edge feather control and no way to correct a mistake the model makes.
• Output is PNG only, always at the input resolution, with no cropping or straightening applied.
• Clearing site data or browser storage removes the cached model, so the next run downloads it again.

There is no fee to use it. Once the credit runs out, a rewarded ad of a few seconds plays before the next cut out.

When another tool fits the job

For an ID photo on a regulation background with the right dimensions, the Passport Photo Maker is built for exactly that. To trim the frame or square it up before cutting out, use Crop Image, and to bring a huge camera file down to a manageable size first try Resize Image. If the finished PNG is too heavy for a listing or an email, Compress Image to KB will aim at a target size, and the Image Converter can turn it into a WebP or JPEG when transparency is no longer needed. To hide a face or a number plate rather than cut a subject out, use Pixelate & Blur Image. To add a caption or a price to the finished cut out, see Add Text to Image, and for sorting a shoot down to the keepers before you edit anything, Photo Cull.

FAQ

Does my photo get sent to a server for processing?

No. The segmentation model is downloaded to your device and runs there. The picture itself is read from disk, processed locally and saved back out, so it never travels anywhere.

Why is the first run so much slower?

The model weights and runtime are about 40 MB and have to be fetched once. Your browser stores them, so the second and every later image skip that download and go straight to processing.

Can I replace the background with a colour instead of transparency?

Yes. Tick the fill option and choose a colour, and the cut out is drawn over a solid background before you download. Leave it unticked to keep a genuinely transparent PNG.

Which formats can I feed it, and what comes out?

Drop in a JPG, PNG or WEBP. The result is always a PNG, since that is the common format that can carry an alpha channel, and it is named after your original with -no-bg on the end.

The result kept part of the wall. What can I do?

That usually means the subject and the background are too similar in colour, or the edge is blurred. Reshoot against a plainer backdrop, improve the lighting so the subject separates, or crop tighter so there is less background to judge.

It failed or the tab reloaded. Why?

Almost always memory. Very large images need a lot of it, and Safari and Firefox are stricter than Chrome or Edge. Reduce the photo to around 2000 pixels on the long edge, close other tabs, and run it again.