---
license: other
license_name: mixed-permissive
license_link: LICENSE.md
tags:
  - face-detection
  - face-recognition
  - source-archive
  - lightcella
---

# Lightcella Face Pipeline — Source Archive

Preservation copies of the **exact upstream artifacts** the Lightcella ncnn face models were converted
from — kept so the conversion stays **reproducible and verifiable** even if an upstream repo moves,
is renamed, or is taken down. All three upstreams are permissively licensed (MIT / MIT / Apache-2.0),
so redistribution here is explicitly allowed; upstream authors retain all credit.

**Converted models:** [`lightcella/face-pipeline-ncnn`](https://huggingface.co/lightcella/face-pipeline-ncnn)
(the shippable ncnn weights). Conversion recipe: see that repo's `CONVERSION.md`.

> These files are **not** needed to run the pipeline — the ncnn weights are self-contained. This
> archive exists only for re-verification / re-conversion / provenance.

## Manifest

| file | provenance | upstream | version | license | SHA-256 |
|---|---|---|---|---|---|
| `face_detection_yunet_2023mar.onnx` | raw upstream | [OpenCV Zoo](https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet) | 2023mar | MIT © 2020 Shiqi Yu | `8f2383e4…52fa4` |
| `face_landmarker_2023-05-03.task` | raw upstream (bundle; only `face_landmarks_detector.tflite` inside is used) | Google MediaPipe | 2023-05-03 build | Apache-2.0 © Google | `64184e22…0bc9ff` |
| `GhostFaceNet_W1.3_S2_ArcFace.h5` | raw upstream | [HamadYA/GhostFaceNets](https://github.com/HamadYA/GhostFaceNets) | W1.3 S2 ArcFace | MIT © 2022 HamadYA | `9147e497…abb333` |
| `GhostFaceNet_W1.3_S2_ArcFace.onnx` | **derived** (fp32 rebuild via tf2onnx + onnxsim from the `.h5`) | ↑ | — | MIT © 2022 HamadYA | `86da1f57…bd3b4` |

Full SHA-256 in `SHA256SUMS`. Verify with `sha256sum -c SHA256SUMS`.

## Notes

- The GhostFaceNet `.h5` ships with per-layer `mixed_float16` dtype policies (Keras returns NaN on
  out-of-fp16-range input). The `.onnx` here is the fp32-rebuilt export (the actual pnnx input),
  included so re-conversion needs no TF/Keras. The `.h5` is the true upstream.
- YuNet `.onnx` and the MediaPipe `.task` are byte-for-byte upstream downloads.

## Licenses

Same licenses as the artifacts they preserve — full texts in `LICENSE.*`; map in `LICENSE.md`.