About
Who builds Backgrounds Remover, and how we test it
Backgrounds Remover cuts the background out of a photo inside your browser, so the picture never reaches a server.
Why we built it
The usual way to remove a background online is to send your photo to a server, run the model there, and send the cutout back. For a product shot that may be fine. For a portrait, a signature, or a picture of your home, it means handing the image to a company you know little about.
Browsers can now run a real segmentation model through WebAssembly and WebGPU, so the upload is no longer necessary. We moved the whole job onto your device: the browser decodes the image, the model finds the subject, and the result is composed and saved in your tab. The details are on how it works.
What we believe
Your photo stays on your device
The AI model runs in your browser tab. The site has no upload endpoint, so there is nowhere for your image to go.
Limits stated up front
Fine hair, glass, and busy scenes can come out imperfect, and the first image downloads a 176 MB model plus the engine (11.3 MB for the CPU build). We say so on the page instead of hiding it.
Free, no account
Removing a background, choosing a color, and downloading the result cost nothing and need no sign-up.
Our own engine build
We compile the inference engine from source and export the model ourselves, so we know exactly what runs on your device.
Who builds it
Backgrounds Remover is built and operated by Alder Street Imaging. The engine, the model export, and the website are developed together in one code base.
The site shows no ads, sets no cookies, and loads no analytics script or provider; the analytics layer in its code has no provider attached, so it sends nothing. The tool never receives your image, so we could not keep it even if we wanted to. The details are in the privacy policy, and the privacy and security page shows how to check them yourself.
How we test it
On every engine build, the engine is checked step by step against the original research code, so a change that moves the cutout beyond the tolerances below fails the check:
- Reference results. The authors' model runs in PyTorch on 14 test images: real photos (a portrait, a cat, a cup of coffee, a rocket) and synthetic edge cases such as white on white, black on black, low contrast, a 1 x 1 image, and a small subject in a large frame.
- Every stage compared. Our ONNX export and the WebAssembly engine, with and without threads, are compared against those reference results, and our C++ resizing and masking is compared against the same steps in the research pipeline. The masks must agree on at least 99.9% of their area, and pixel values may differ by no more than a few levels out of 255.
- WebGPU against CPU, by hand. A browser test that we run by hand, not on every change, puts the same photo through the WebGPU build and the CPU build on a software graphics adapter; the two masks must agree on at least 99% of their area, with an average difference under one level out of 255. It passed on September 17 and September 23, 2026. Real graphics cards have not been measured, for the cutout or for speed.
- The page itself. An automated browser test loads the site, removes a background, and checks that the result is transparent where it should be, that no request leaves the site, and that the image is never uploaded (while it works, every request is a GET or HEAD with no body). It also covers the white and picked-color backgrounds, JPG export, the square size presets, the sample photo, and a return visit that loads the model from the device.
On September 17, 2026 we also ran three test images through the site's engine and looked at the edges ourselves: a flat logo, a signature on textured paper, and a studio portrait. What we saw is written up in how it works, including where the tool fell short.
Contact
Questions, bug reports, or an image the tool handles badly: write to support@backgrounds-remover.online. Please describe the problem rather than attaching a private photo; we never need your image to help.
Related reading
Try it on your own photo
Runs in your browser. Your image never leaves your device.