title: Browser Remove Background
emoji: 🖼️
colorFrom: blue
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
short_description: In-browser background removal (WebGPU/WASM) by is.tools
models:
- chenjindu/isnet-web
Browser Remove Background
Removes image backgrounds entirely in your browser — the image is decoded, segmented and composited on your own device. Nothing is uploaded, no server, no API key, no account.
Built with ONNX Runtime Web running the ISNet segmentation model (Apache-2.0), with both WASM and WebGPU execution providers so you can measure the difference yourself.
Live demo
chenjindu.github.io/browser-remove-background
A production version with a proper UI runs at https://is.tools/remove-background — free, no signup, no watermark, no size cap.
Run it locally
npx serve .
Then open http://localhost:3000, pick a photo, and run inference with WASM and with WebGPU to compare.
Benchmark
Measured on a desktop with a discrete GPU, Chrome 113+, 1024×1024 input:
| Execution provider | 2nd inference |
|---|---|
| WebGPU | ~0.6 s |
| WASM | ~17 s |
The first run is always slower — model download, session creation and shader compilation. Read the second number.
WASM — runs everywhere, roughly 25× slower.
WebGPU — Chrome/Edge 113+, sub-second after warm-up.
Requirements
WebGPU needs Chrome or Edge 113+. Without it the page falls back to WASM, which works in any modern browser but is roughly 25× slower on this model.
The model ships as three parts under model/ and is reassembled in the browser,
which keeps every file under GitHub's size limit.
Part of is.tools
This probe came out of building https://is.tools — a set of browser-side tools for PDF, image, text and encoding work, built under one rule: the file never leaves your device.
Same approach throughout: image codecs are MozJPEG / libwebp / libavif / OxiPNG compiled to WASM, PDF work uses pdf-lib and qpdf, OCR uses tesseract.js, and background removal uses the ISNet model in this repo. No accounts, no quotas, no upload step, no watermark on the output.
Available in 8 languages.
License
Apache-2.0. The ISNet model is Apache-2.0 as well.