RuiSumida/AuraFace-v1-CoreML

AuraFace-v1 — CoreML

CoreML conversion of fal/AuraFace-v1's recognition model (glintr100.onnx, ResNet100 with additive angular margin loss), for on-device face embedding on iOS.

AuraFace-v1 is released by fal.ai under Apache-2.0 as a commercially-usable alternative to the non-commercial InsightFace recognition packs. This repo redistributes a converted copy under the same license.

Files

• FaceEmbedding.mlmodel — neuralnetwork format, float16 weights (~124 MB)

Interface

• Input faceImage: 112×112 RGB image, 5-point aligned (InsightFace norm_crop / ArcFace template). Preprocessing (pixel − 127.5) / 127.5 is baked into the model.
• Output embedding: 512-dim face embedding (L2-normalize before cosine comparison).

Conversion

ONNX → PyTorch (onnx2torch) → traced → coremltools (neuralnetwork) → fp16 weight quantization. Validated against ONNX Runtime: cosine similarity 0.999956 on random input (fp16 quantization is effectively lossless).

Accuracy (from the upstream model card)

LFW	CFP-FP	AgeDB	CALFW	CPLFW
99.65	95.19	96.10	94.70	90.93