Open-source licenses

Background removal runs in your browser. The picture is not uploaded for this step.

Flat background

Flat or nearly flat backgrounds are removed by project code. It reads the border color and flood-fills inward. Edges stay hard, which suits cartoons, logos, and pixel art.

ISNet, general use

Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool, “Highly Accurate Dichotomous Image Segmentation,” ECCV 2022.

• Code and the evaluation metric are Apache License 2.0: xuebinqin/DIS.
• The DIS5K dataset has its own terms. This site does not ship or use that dataset.
• The repository README licenses the code and metric, and publishes the isnet-general-use weights for general use. It does not add a separate non-commercial license for the weight file. A request for an explicit sentence about the weights is still open (DIS issue 150). The file used here is distributed on that basis.
• Source ONNX: rembg v0.0.0 isnet-general-use.onnx (rembg is MIT). MD5 fc16ebd8b0c10d971d3513d564d01e29. SHA-256 60920e99c45464f2ba57bee2ad08c919a52bbf852739e96947fbb4358c0d964a. Release file.
• The browser file is a weight-only per-channel int8 quantization of that ONNX (symmetric DequantizeLinear, axis 0), 46,714,903 bytes. SHA-256 d1216111717a063a165f325aa6c70aeecd4e76fc0f3180af6adf76471d389b9f.
• Not used: onnx-community/ISNet-ONNX (its model card says AGPL-3.0), BRIA RMBG, and @imgly/background-removal (AGPL).

U²-Netp

Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar Zaiane, and Martin Jagersand, “U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection,” Pattern Recognition, 2020. Code: Apache License 2.0, xuebinqin/U-2-Net.

ONNX: rembg v0.0.0 u2netp.onnx. MD5 8e83ca70e441ab06c318d82300c84806. SHA-256 309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8.

Runtime

onnxruntime-web 1.22.0, MIT License, Microsoft. WebGPU is used when the browser provides it; otherwise the model runs on WebAssembly.