README.md

---
license: apache-2.0
base_model: fal/AuraFace-v1
tags:
- coreml
- face-recognition
- arcface
- auraface
library_name: coreml
---

# AuraFace-v1 — Core ML (fp16)

Core ML conversion of the face-recognition model from
[fal/AuraFace-v1](https://huggingface.co/fal/AuraFace-v1), for use in
[Visuals](https://bloosoftware.com) on macOS and iOS.

**This is a modified work.** The original `glintr100.onnx`
(SHA-256 `a7933ea5330113b01c9b60351d8f4c33003f145d8470ac5f0e52ee2effe25c60`) was converted from
ONNX to a Core ML ML Program package with float16 weights. No weights were retrained, pruned or
otherwise altered beyond the precision change inherent to the conversion. Output parity against
the ONNX reference was verified at cosine ≥ 0.99995 over random inputs.

## Files

```
auraface_v1.mlpackage/
├── Manifest.json                             617 B
├── Data/com.apple.CoreML/model.mlmodel       253,683 B
└── Data/com.apple.CoreML/weights/weight.bin  130,364,480 B
```

## Contract

| | |
|---|---|
| Input | `data` — Float32 `[1, 3, 112, 112]`, **RGB**, `(pixel / 127.5) - 1` |
| Output | `embedding` — Float32 `[1, 512]`, not normalised (L2-normalise downstream) |
| Architecture | ArcFace-style ResNet100 |

Preprocessing follows InsightFace's `ArcFaceONNX`: 112×112, mean 127.5, std 127.5, RGB channel
order. Best results come from a 5-point aligned crop; a tight face crop also works, with reduced
separation.

## Licence and attribution

Original model © fal, released under the **Apache License 2.0** — see `LICENSE.md`, reproduced
unmodified from the upstream repository. This conversion is distributed under the same licence.
The upstream model card states the model "has been trained on commercially and publicly available
data sources to enable its usage in commercial setting."

Upstream fairness note, carried forward: the training data "may not extensively cover all
ethnicities", and fal recommends downstream users assess fairness in their own context.