| --- |
| 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. |
|
|
