ArcFace ResNet-50 face-embedding model, fine-tuned from InsightFace's buffalo_l for the Zekra Alzheimer's care companion app. 512-dimensional embeddings, packaged as a single ONNX file in fp16 (≈83 MB) for on-device inference via onnxruntime.

⚠ Non-commercial license

This model is licensed for non-commercial research and educational use only. It is a derivative of InsightFace buffalo_l, whose pretrained weights are released under their non-commercial research license. Any commercial use of this model (or the Zekra app shipping it) requires a separate license from InsightFace.

This restriction does not apply to the companion language model published at amaniee/zekra-care-companion-model, which carries the Gemma Terms of Use.

What it does

Produces a 512-dimensional face embedding from a 112×112 aligned face crop. Used by Zekra to recognize family members from photos the caregiver uploads, then surface the story behind their face when the patient encounters them via the in-app camera.

Search: cosine similarity against an on-device HNSW index of family-member embeddings.
Threshold: cos ≥ 0.55 for confident match, 0.40–0.55 for "might be" / show-candidate.
Validation: 95.5% top-1 accuracy on the Zekra family-graph validation set, up from 75.1% on the unfine-tuned buffalo_l baseline.

Files

arcface_zekra_r50_fp16.onnx — fine-tuned ResNet-50 ArcFace, fp16 weights, ONNX format.

Intended use

On-device face recognition inside the Zekra Flutter app, running through onnxruntime.
Research / educational use under the non-commercial license.

Out of scope

No commercial use without a separate license from InsightFace.
Not designed for general-purpose face recognition outside the dementia-care companion context.
Not validated for use in surveillance, identification of strangers, law-enforcement applications, or any setting where the consent of the photographed person is not explicit.

Training

Base model: InsightFace buffalo_l (ResNet-50 ArcFace, 512-d embedding).
Method: CosFace classification head fine-tune (s=15, m=0.35) with BatchNorm frozen and a head-only warm-up phase. Memory: pair-contrastive collapses on this task — wrong loss family.
Data: Zekra family-graph pairs, on-device-captured + caregiver-uploaded family photos. No data left the device during training-data collection.
Format: ONNX export at fp16 for ~83 MB on-device footprint (INT8 variant validated 0.5pp behind on 13.8K val pairs — fp16 is the current ship).

Privacy

This model produces embeddings, not pixels. The training photos never left the device during data collection. The shipped ONNX file contains weights only — no identifiable face data is recoverable from the model parameters.

License (full text)

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

You are free to:

Share — copy and redistribute the material in any medium or format
Adapt — remix, transform, and build upon the material

Under the following terms:

Attribution — credit InsightFace + Zekra; provide a link to the license and indicate changes.
NonCommercial — you may not use the material for commercial purposes.

Full license text: https://creativecommons.org/licenses/by-nc/4.0/

Citation

@misc{zekra-face-2026,
  title  = {Zekra Face Embedder: ArcFace fine-tune for dementia-care companion},
  author = {El-Hajj, Amanie},
  year   = {2026},
  url    = {https://huggingface.co/amaniee/zekra-face-embedder},
  note   = {Derived from InsightFace buffalo_l; non-commercial research license}
}

@inproceedings{deng2019arcface,
  title     = {ArcFace: Additive Angular Margin Loss for Deep Face Recognition},
  author    = {Deng, Jiankang and Guo, Jia and Xue, Niannan and Zafeiriou, Stefanos},
  booktitle = {CVPR},
  year      = {2019}
}

Companion model

The language model used alongside this face embedder is published separately at amaniee/zekra-care-companion-model under the Gemma Terms of Use. The two models are independent and licensed under different terms.