AI Culler incorporates open-source software and pre-trained AI models. No GPL or AGPL components
are bundled. The complete attribution and license texts are installed with the application in the
licenses folder and in THIRD-PARTY-NOTICES.txt.
1. Runtime components
Components are dynamically linked where their licenses require it, so that you may replace them with modified versions.
| Component | License | Notes |
|---|---|---|
| PySide6 / Qt 6 | LGPL-3.0 | Dynamically linked; the shared libraries ship as separate files and can be replaced. Full license text included on installation. |
| shiboken6 | LGPL-3.0 | Same terms as PySide6. |
| libheif / libde265 | LGPL-3.0 | Decoding only; no HEVC encoder is distributed. |
| LibRaw (via rawpy) | LGPL-2.1 (LibRaw commercial exception) | Uses the LibRaw exception that permits use in closed-source products. |
| certifi | MPL-2.0 | File-level copyleft; unmodified. |
| OpenCV | Apache-2.0 | Headless build; video I/O and FFmpeg are excluded from the distribution. |
| onnxruntime | MIT | Local inference only. |
| NumPy | BSD-3-Clause and others | See bundled license file. |
| Pillow | MIT-CMU | Image decoding. |
| psutil | BSD-3-Clause | Memory and CPU detection. |
| cryptography | Apache-2.0 OR BSD-3-Clause | Signature verification. |
| protobuf, flatbuffers | BSD-3-Clause / Apache-2.0 | Model runtime dependencies. |
| Python (CPython) | PSF-2.0 | Runtime. |
2. Pre-trained models
The following models are distributed with the installer and run locally. They are used unmodified.
| Model | Purpose | License |
|---|---|---|
| RetinaFace (MobileNetV2 backbone) | Face detection and facial landmarks | MIT |
| MediaPipe FaceMesh | 468-point facial mesh for eyelid openness | Apache-2.0 |
| OCEC-S | Closed-eye classification | MIT |
| FER+ | Facial expression (Microsoft / ONNX Model Zoo) | MIT |
| CLIP ViT-B/32 (INT8) | Visual similarity features (OpenAI) | MIT |
| CLIP-IQA | Image quality assessment | MIT |
| HiRQA-S | Imaging quality assessment | Apache-2.0 (ResNet-18 backbone) |
| U²-Net-p | Portrait segmentation | Apache-2.0 |
| EDiffIQA-tiny | Portrait quality assessment | Apache-2.0 |
Model authors retain all rights. AI Culler only performs inference; no model is fine-tuned or redistributed in modified form.
3. Components deliberately excluded
- x265 / libx265 (GPL-2.0) — the HEVC encoder is not needed for reading photos and is stripped from every build; the build aborts if it reappears.
- Ultralytics YOLO (AGPL-3.0) — not used and not bundled.
- FFmpeg (LGPL) — video I/O is excluded; the application handles still images only.
- Qt Quick / QML / Qt PDF / WebEngine — removed to reduce size and licensing surface.
4. Your rights under these licenses
Where a component is licensed under LGPL, you are entitled to obtain, modify and replace the library with a
compatible version. Because the libraries ship as separate DLLs in the installation folder, you can replace
them directly. The full license texts are provided in the licenses folder; source code for
LGPL-licensed libraries is available from the upstream projects named above.
This page is a summary for convenience. The authoritative texts are the files installed with the software. This notice supplements and does not override the AI Culler End User License Agreement.
Last updated: 28 October 2026.