---
license: apache-2.0
tags: [coreml, face-recognition, arcface]
---
# ArcFace LResNet100E-IR — CoreML (fp16)

CoreML conversion of [arcfaceresnet100-8.onnx](https://github.com/onnx/models/tree/main/validated/vision/body_analysis/arcface)
from the ONNX Model Zoo (Apache-2.0), for on-device face embedding.

- Input: `faceImage` — 112x112 RGB image (raw 0-255 pixels; BN at graph input)
- Output: `embedding` — 512-d float vector (L2-normalize before use)
- mlprogram / fp16, iOS 17+
- Conversion: onnx (opset 8->13) -> onnx2torch -> coremltools.
  Numerical parity vs onnxruntime verified (cosine 1.0 on real faces).

Packaged as `.mlpackage.tar.gz` for on-device compilation (mlprogram models
cannot ship as a single .mlmodel file).