How RemoveBkgd removes a background without uploading your image

Background removal is a computer-vision task: a model looks at every pixel and predicts how likely it is to belong to the foreground subject. For years those models were too large and too slow to run on a normal laptop, so a background-removal website had to send your image to a server, run the model on a GPU there and return the result. That is why most services meter usage — server GPUs cost money, and the cost scales with every image you process.

The browser changed that. WebAssembly lets compiled C++ and Rust run at near-native speed inside a web page, and WebGPU exposes a device’s graphics chip to JavaScript. Together they can run optimised neural networks locally. RemoveBkgd ships a compact segmentation model that fits these constraints, which is why the compute happens on your device and why the exports can stay free.

RemoveBkgd runs a single open, MIT-licensed segmentation model (ISNet general-use, about 42 MB) that handles people, products and complex subjects alike. It preserves fine structures — hair, fur, leaves, glass and thin straps — so there is no need to choose between a "portrait" and an "object" engine. It runs fastest with WebGPU and also works on CPU-only browsers.

The model is downloaded once and then cached by your browser. On later visits — and even when you go offline — there is no model download at all, so processing is essentially instant.

Automatic masking is very good, but a single threshold is rarely perfect for every image. A hard mask edge can look pasted-on; a slightly soft edge composites more convincingly. The feather control blurs the alpha channel around the boundary so the subject blends rather than cuts. The expand/contract control pushes the mask outward or inward by a few pixels, which recovers a shoulder, a jersey number or a product rim that the model trimmed. The drop-shadow control adds a soft, offset shadow beneath the subject so it sits in the new scene instead of floating in it.

Because everything is drawn to a canvas in real time, you see each adjustment immediately at the same resolution you will export. There is no render queue and no round trip.

Model quality is only half the result — the source photo matters too. A subject that is clearly separated from the background, lit evenly and not blurred by motion gives the model an easier job. When an image has a busy background with similar colours to the subject, use expand/contract to correct the boundary. For hair and fur, add a small amount of feather rather than none, and export PNG so transparency is preserved.