# THIRD_PARTY_LICENSES.md — Deback

Every significant third-party component used by this site, with its license
and what the license requires of us. Verified 2026-09-16. If any entry becomes
doubtful, treat it as unverified and re-check before shipping.

## 1. Transformers.js — inference library

- **What:** `@huggingface/transformers` v3.4.0, loaded at runtime from jsDelivr
  (`https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.4.0`). Provides the
  `background-removal` pipeline and WebGPU/WASM execution via ONNX Runtime Web.
- **Source:** https://github.com/huggingface/transformers.js
- **License:** **Apache License 2.0** (LICENSE file in the repo).
- **Commercial use:** permitted. **Redistribution:** permitted with license
  attribution. **Copyleft:** none (permissive).
- **What we must do:** include/retain the Apache-2.0 attribution. This file
  serves as that attribution record; the library's own LICENSE is distributed
  by the CDN copy.
- **Note:** we do not bundle or modify the library — the browser loads the
  published file. No vendored copy in this repo.

## 2. IS-Net model weights — the AI model

- **What:** `onnx/model_quantized.onnx` (~44 MB, int8-quantized) from the
  Hugging Face repo `xrds/isnet-general-onnx-int8` (ONNX conversion of the
  IS-Net "general use" dichotomous segmentation model). Downloaded once per
  visitor into their browser; never hosted by us.
- **Model card license tag:** `license: mit`.
- **LICENSE file:** present in the repo root (MIT License, copyright
  imgly GmbH).
- **License:** **MIT**.
- **Commercial use:** permitted under MIT. **Attribution required:** yes —
  the MIT license requires the copyright notice and permission notice to be
  included. Satisfied by the attribution block at the end of this file.
- **Why this model (recorded honestly):** BiRefNet Lite was evaluated first
  but requires ~4.5 GB RAM at its fixed 1024×1024 inference resolution, which
  exceeds the WASM heap available in browsers — it hard-crashed during
  testing on CPU. IS-Net int8 runs fully in-browser on WASM CPU in every
  modern browser (verified end-to-end), at a 44 MB download.
- **Residual caveats (not blockers, recorded honestly):**
  1. The ONNX export/quantization is a third-party conversion, not published
     by the original IS-Net authors — standard practice on the Hub, but
     worth noting.
  2. As with nearly all public segmentation models, the training data
     (DIS5K) has its own terms; the MIT grant covers the *weights*.
     Flagged in `LEGAL_REVIEW.md` for attorney review.

## 3. IS-Net paper/code (reference only — not shipped)

- **What:** the IS-Net architecture ("Highly Accurate Dichotomous Image
  Segmentation", Qin et al.). Not used at runtime; listed because the
  weights derive from it.
- **License:** per the original authors' release.

## 4. ONNX Runtime Web (transitive)

- **What:** execution backend bundled inside Transformers.js for WebGPU/WASM.
- **License:** Apache License 2.0. Permissive; no action beyond this record.

## 5. Site artwork & sample image (original, no third-party license needed)

- `assets/img/logo.svg`, `favicon.svg`, `og-cover.png`: original artwork
  created for this project.
- `assets/img/sample-balloon.jpg`: synthetic sample image generated
  programmatically for the "try a sample" demo — no stock, no copyrighted
  source material.
- All page copy (guides, FAQ, legal pages, comparison table): written
  originally for this site. Competitor facts are dry factual data (prices,
  plan terms) with sources cited, not copied marketing text.

## 6. Deliberately NOT used (license rejections, for the record)

- `briaai/RMBG-1.4` — BRIA custom license is **non-commercial**; commercial use
  requires a BRIA agreement. Ruled out.
- `onnx-community/BEN2-ONNX` — card declares MIT and the original repo LICENSE
  is MIT, but vendor README distinguishes an open "base" model from paid
  commercial offerings; BiRefNet's explicit author confirmation was cleaner.
- `Xenova/modnet` — Apache-2.0 declared on the conversion card, but the
  original MODNet repo has no license file; quality is also a generation older.
- RobustVideoMatting — **GPL-3.0** (copyleft); incompatible with a proprietary
  commercial site. Ruled out.
- U²-Net — code Apache-2.0 but **no clear weight license** found. Ruled out.

## MIT attribution (BiRefNet)

Per the MIT license terms, the following notice is reproduced:

> BiRefNet — Copyright (c) Zheng Peng and contributors.
> Licensed under the MIT License.
>
> Permission is hereby granted, free of charge, to any person obtaining a copy
> of this software and associated documentation files (the "Software"), to deal
> in the Software without restriction, including without limitation the rights
> to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
> copies of the Software, and to permit persons to whom the Software is
> furnished to do so, subject to the following conditions:
>
> The above copyright notice and this permission notice shall be included in all
> copies or substantial portions of the Software.

*(The precise copyright-holder line should be conformed to the LICENSE text
in the upstream repo when it is added; the grant quoted above is the author's
statement in issue #316.)*