# AuraFace v1 — CoreML pack for Run It Back

Weights: [fal/AuraFace-v1](https://huggingface.co/fal/AuraFace-v1) (`glintr100.onnx`),
Apache License 2.0 (see `LICENSE.md`). AuraFace is a ResNet100 trained with the
Additive Angular Margin (ArcFace) loss on commercially licensed data by fal.ai.

Conversion: `scripts/convert-auraface.py` in the Run It Back repository
(onnx2torch → coremltools 9, fp16 mlprogram, compiled with coremlcompiler).
The InsightFace input normalization `(x - 127.5) / 127.5` on RGB is folded
into the model's image input; L2 normalization is applied by the consumer.
Numerical parity against onnxruntime on the repository's face fixtures:
cosine ≥ 0.99994.

Changes from the source: format only (ONNX → CoreML), no retraining.