# Product background remover licence gate — 2026-09-30

Selected: **U²-Net portable (u2netp), float32 ONNX, opset 11**, by Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R. Zaiane and Martin Jagersand (Pattern Recognition 106, 107404, 2020).

Weights URL (immutable host revision): https://huggingface.co/edgetools/u2netp/resolve/25dee37ab19c5b6ad64ba6578eba63f1ae07720c/u2netp.onnx

SHA-256: `309c8469258dda742793dce0ebea8e6dd393174f89934733ecc8b14c76f4ddd8`
Size: **4,574,861 bytes** (4.575 MB / 4.363 MiB). Vendored unchanged as `u2netp.onnx`; below the 25 MB storage limit. No local export, quantisation or modifications.
The host states this is a byte-for-byte mirror of [rembg's v0.0.0 ONNX release](https://github.com/danielgatis/rembg/releases/download/v0.0.0/u2netp.onnx), retaining the upstream weights licence.

## Separate code and weights findings

- Authors' code: **Apache-2.0**. [Authors' LICENSE](https://github.com/xuebinqin/U-2-Net/blob/master/LICENSE) states “Apache License” and “Version 2.0, January 2004”. The complete licence is included in `LICENSE-U-2-Net.txt`; it grants reproduction, derivative works and distribution, and use and sale, without a non-commercial restriction.
- Weights: **Apache-2.0**. [Weights host model card](https://huggingface.co/edgetools/u2netp/blob/25dee37ab19c5b6ad64ba6578eba63f1ae07720c/README.md) states `license: apache-2.0` and “The U²-Net weights are Apache-2.0.” [Authors' README](https://github.com/xuebinqin/U-2-Net#usage-for-salient-object-detection) publishes `u2netp.pth` alongside the code under the repository's Apache licence. This is the salient-object model, not the portrait or human-segmentation variants.
- ONNX distributor rembg: **MIT**, [LICENSE.txt](https://github.com/danielgatis/rembg/blob/main/LICENSE.txt): “MIT License”. This does not replace the upstream Apache weights licence.
- Browser inference runtime: **onnxruntime-web 1.17.3, MIT**, [Microsoft LICENSE](https://github.com/microsoft/onnxruntime/blob/v1.17.3/LICENSE): “MIT License”. Its permission includes “use, copy, modify, merge, publish, distribute, sublicense, and/or sell”. Vendored from the pinned npm tarball `https://registry.npmjs.org/onnxruntime-web/-/onnxruntime-web-1.17.3.tgz`. Only `ort.wasm.min.js`, plain `ort-wasm.wasm`, and plain SIMD `ort-wasm-simd.wasm`; no threaded/JSEP files. Full MIT licence in `../onnxruntime/LICENSE`.

## Candidate comparison

MB below means decimal megabytes; sizes refer to published files, not speculative exports.

| Candidate | Code licence | Weights licence on host | Size | Verdict |
|---|---|---|---|---|
| U²-Net u2netp | Apache-2.0 | Apache-2.0 (edgetools host above) | 4.575 MB ONNX | PASS, selected; smallest checked product-capable model |
| U²-Net full | Apache-2.0 | Apache-2.0 (https://huggingface.co/Heliosoph/u2net-onnx) | ~176 MB | Over 50 MB |
| IS-Net general-use / DIS | Apache-2.0 (https://github.com/xuebinqin/DIS/blob/main/LICENSE.md) | Apache-2.0 (https://huggingface.co/jellybox/isnet-general-use) | 178.648 MB ONNX | Over 50 MB; authors' README expressly licences code/evaluation, host supplies weights field |
| BiRefNet | MIT (https://github.com/ZhengPeng7/BiRefNet/blob/main/LICENSE) | MIT (https://huggingface.co/ZhengPeng7/BiRefNet) | 444.474 MB safetensors | Over 50 MB as published |
| BiRefNet_lite | MIT (same author repository) | MIT (https://huggingface.co/ZhengPeng7/BiRefNet_lite) | 177.634 MB safetensors; ~115 MB fp16 ONNX (https://huggingface.co/onnx-community/BiRefNet_lite-ONNX/tree/main/onnx) | Over 50 MB as published |
| rembg silueta | MIT distributor; Apache-2.0 U²-Net origin | No explicit model licence field on checked weights host (https://huggingface.co/fofr/comfyui); MIT distributor alone does not establish weights terms | 44.2 MB ONNX | Reject under explicit weights-host gate; also larger |
| MODNet | Apache-2.0 (https://github.com/ZHKKKe/MODNet#license) | Author README expressly covers models under Apache-2.0 | ~26 MB | Portrait only, unsuitable for products |
| BRIA RMBG-1.4 / 2.0 | Restricted BRIA terms | Non-commercial / commercial agreement (https://huggingface.co/briaai/RMBG-1.4 and https://huggingface.co/briaai/RMBG-2.0) | Not downloaded | Reject |
| @imgly/background-removal | AGPL | Irrelevant after runtime rejection | Not downloaded | Reject |

## Training data and practical limits

The [authors' paper](https://arxiv.org/abs/2005.09007) describes training on DUTS-TR (10,553 training images); the repository also mentions DUTS-TR in its comparison with a separate human segmentation model. This selection does not use the APDrawing portrait checkpoint or Supervisely human checkpoint. The authors do not provide a separate training-data commercial-use warranty in the cited sources; permissive model licensing is not a licence to redistribute training images. No training data is bundled here.

Input: RGB float32 NCHW `[1,3,320,320]`, resize without letterboxing, divide RGB by per-image maximum then ImageNet mean/std (`[.485,.456,.406]`, `[.229,.224,.225]`). First fused output is min/max normalised, bilinearly upscaled, thresholded and feathered. This small saliency model is intended for a clear product on a simple background; hair, glass and busy scenes can need touch-up.