❯ Background Remover

It uses transformers.js (Apache-2.0) to run a segmentation neural network inside the browser via ONNX Runtime Web. Two models are available: Portrait (MODNet, Apache-2.0, 6.6 MB) for people, and General (BiRefNet lite, MIT, 114 MB) for any subject. Weights download once from Hugging Face and stay cached by your browser; the ONNX Runtime WASM is served from bytefork.tools itself. After that first download, every removal happens entirely in your browser tab — the image bytes never leave your device.

For clear subjects on contrasting backgrounds (product photos, portraits with separable hair, animals against sky), accuracy approaches commercial cloud services like remove.bg. For complex hair detail, transparent objects (glass, water), motion blur, or low-contrast subject/background pairs, results may show halos or missing edges. Portrait uses MODNet, which is tuned specifically for people; General uses BiRefNet lite, which handles arbitrary subjects. Both are strong for browser-side segmentation, neither is magic.

The first run downloads the model. Portrait is 6.6 MB — a second or two on broadband. General is 114 MB, roughly 20–30 seconds on a fast connection and longer on a slow one. After that the model lives in your browser cache and the second run jumps straight to inference (2–10 seconds for a typical photo). If you are cutting out a person, stay on Portrait and you never pay the large download.

Not in this version — one image at a time. Batch is on the roadmap; for now, drop each image into the tool one by one.

Input: PNG, JPG, WebP. Output: PNG with an alpha channel (transparent background) or PNG composited on a solid color you pick.

PNG is the only widely-supported format that handles alpha transparency without artifacts. JPG has no alpha and would force you to flatten against a chosen color (which you can do explicitly via the Color mode). WebP supports alpha but support is uneven in older toolchains.

Yes, and the whole stack is permissively licensed with no commercial tier to buy. bytefork.tools is MIT. transformers.js is Apache-2.0 and ONNX Runtime Web is MIT. The model weights — the layer that usually carries the restriction — are permissive too: MODNet is Apache-2.0, with its author's README explicitly extending the grant to the models, and BiRefNet is MIT. This tool deliberately avoids the popular RMBG-1.4 and RMBG-2.0 weights, which are non-commercial (CC BY-NC 4.0 / Bria's own licence), and IS-Net, whose Apache-2.0 tag covers only "code and evaluation metric" while its DIS5K training data is non-commercial.

Use this when privacy matters (your image never leaves your device), when you need to process many images without paying per credit, and when you're OK with browser-side accuracy. Use remove.bg when you need their tuned pro model for edge cases (very fine hair, complex translucent objects). For most everyday use, the difference is small enough that the privacy + zero-cost trade is the right one.

ONNX Runtime Web — the inference engine that runs the background-removal neural network — JIT-compiles its tensor kernels using `new Function()`. That requires `unsafe-eval` in the page's Content Security Policy. Every other tool on bytefork.tools runs under a strict CSP that forbids `eval()` and `new Function()`. We scoped the relaxation to this one URL using a per-path `Content-Security-Policy` header in public/_headers — you can verify in DevTools → Network → response headers that the CSP only differs here. Practical risk is low because the site never accepts user-submitted scripts; the additional eval surface has no inputs to exploit. If you prefer not to run a page with `unsafe-eval`, skip this tool.

Yes. No signup, no usage limit, no watermark, no ads. The model download is the only "cost" — and it's a one-time, no-data-leaked transfer.