Background Remover

about this tool

On your device, first-party

Drop a photo and the tool decodes it, runs a salient-object model over it and returns a PNG with the background made transparent. Everything happens in your browser. The first use fetches two files from this site, once: the model, U²-Net's small variant (4.6 MB, Apache 2.0), and ONNX Runtime Web (about 14 MB, MIT), which executes it in WebAssembly. Both are cached by the browser afterwards. The photo is never sent anywhere — this site has no server that could receive it — and the privacy page's promise that a page here makes no request to anyone else still holds, because the model and the runtime are served from the same origin as the page.

That choice is why the well-known @imgly/background-removal library is not used: it is AGPL-licensed and loads its models from a third-party CDN.

What the model does

U²-Net predicts, for each pixel, how likely it is to belong to the main subject. The image is stretched to 320 × 320 for the model, the resulting map is normalised and resized bilinearly back to the picture's size, and that becomes the alpha channel. It is a saliency model, not a portrait matting model: it excels at a clear subject against a distinct background — people, products, pets, vehicles — and struggles with busy scenes, several subjects, or a subject that fills the frame.

Tuning without re-running

The model runs once per image; every adjustment re-composes from its map. Cut-off moves the point at which a pixel counts as subject; edge softness sets how wide the blend is, from a hard cut for products to a wide feather for hair. Trim to the subject crops to the subject's bounding box with a small margin. The backdrop replaces transparency with white, black or a colour of your choice for a flat PNG, and the panel reports how much of the picture was kept.

Limits

Processing is capped at 2048 px on the longest side to fit in a browser tab's memory; the result is that size. The model's native resolution is 320 px, so edges are as fine as the resized mask allows and hairline detail is approximate. A single-threaded WebAssembly pass takes a second or two on a laptop and longer on a phone. The runtime needs WebAssembly SIMD, which every browser since 2021 has. Nothing is stored between visits beyond the browser's ordinary cache of the two files.

Once the background is gone, image format converter will save it as a PNG or WebP with transparency intact, and image compressor will shrink the result.

questions

Is my photo uploaded anywhere?

No. The model (U²-Net, 4.6 MB) and the runtime (ONNX Runtime Web, about 14 MB) are fetched from this site the first time you use the tool and cached by your browser; the photo is decoded, analysed and re-encoded entirely on your device. No request carries your image, and this site keeps no server for it.

How does the model decide what is the background?

U²-Net is a salient-object detector: it was trained to predict which pixels belong to the main subject of a photograph, and produces a soft map from 0 (background) to 1 (subject). The tool stretches that map to the image size and uses it as the alpha channel. It works best on a clear subject — a person, a product, a pet — against a distinct background.

The edge is wrong; can I fix it?

Two sliders act on the model’s map without re-running it. Cut-off moves the point where a pixel counts as subject — lower it if parts of the subject were lost, raise it if background crept in. Edge softness widens or hardens the transition; soft edges suit hair and fur, hard edges suit products and logos.

Why is the output smaller than my photo?

Images are processed at up to 2048 px on the longest side to keep memory within what a browser tab allows; larger photos are scaled down first and the result is that size. The model itself sees a 320 × 320 version, so fine detail comes from the resized mask, not from the model.

What is the model licence?

U²-Net is released under the Apache License 2.0 and ONNX Runtime under the MIT licence, both permissive. The widely used @imgly/background-removal library was not used because it is AGPL-licensed and fetches its models from a third-party CDN, which this site’s privacy policy rules out.