Xargaar Tools Background Remover: third-party licenses and provenance
Reviewed 2026-10-15. This file is not legal advice.

ONNX Runtime Web 1.20.1
Copyright (c) Microsoft Corporation. MIT License.
Full license: vendor/LICENSE-onnxruntime.txt
Source: https://github.com/microsoft/onnxruntime/tree/v1.20.1
Runtime JS, Emscripten module and WASM are bundled locally without runtime CDN dependencies.

U2NETP / U²-Net lightweight salient-object segmentation
Xuebin Qin et al. Apache License 2.0.
Full license: models/LICENSE-u2net.txt
Source: https://github.com/xuebinqin/U-2-Net
Weight license confirmation from upstream authors:
https://github.com/xuebinqin/U-2-Net/issues/208
The authors state that released code and models follow the same license, and
that users should check dataset licenses, which the authors do not control.
They confirm u2net and u2netp were trained with DUTS.
ONNX conversion distributed by rembg, not a rembg relicensing of the weights:
https://github.com/danielgatis/rembg/releases/tag/v0.0.0
Artifact: https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2netp.onnx
The mirror does not provide a complete conversion provenance record.

Training data note:
DUTS project: https://saliencydetection.net/duts/
DUTS-TR images originate from ImageNet DET training/validation sets.
The DUTS page reserves ground-truth annotation rights and does not give a clear
blanket commercial dataset license. This app does not distribute DUTS images
or annotations, and uses the upstream Apache-2.0 released weights. Dataset
rights are not certified by this product. No assertion of complete training-data
clearance is made. Assess that distinction before commercial distribution.
No Bria RMBG or @imgly/background-removal code or weights are included.

Model citation:
Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R. Zaiane,
Martin Jagersand. U²-Net: Going Deeper with Nested U-Structure for Salient
Object Detection. Pattern Recognition, 2020.

Offline bundling modification:
For single-file use, the Emscripten module's import.meta.url references are
replaced by a harmless absolute URL to let its wasmBinary option initialize
from memory under file://. No network request is made to this URL. Original
module is preserved in vendor/. Tools use single-thread WASM (no worker).

Photo note:
The built-in product sample is a simple original drawing, not a benchmark photo.
Quality-test images, reports and licenses are outside the public website.