MLX-UniFace: Blazing-Fast Face Analysis on Apple Silicon

MLX-UniFace is a high-performance face analysis library optimized for Apple Silicon. It provides face detection, recognition, landmark detection, and attribute analysis with native MLX acceleration.

This is a fork of yakhyo/uniface with added MLX backend support for Apple Silicon Macs.

Why MLX-UniFace?

Performance Benefits on Apple Silicon

• Unified Memory: No CPU-GPU data transfer overhead
• Native Acceleration: Optimized for Apple's Neural Engine and GPU
• Lazy Evaluation: Automatic graph optimization
• Numerical Parity: Identical results to ONNX (correlation = 1.0)

Installation

For Apple Silicon (Recommended)

pip install mlx-uniface

With MLX Backend (Explicit)

pip install mlx-uniface[mlx]

With ONNX Fallback

pip install mlx-uniface[onnx]

Install from Source

git clone https://github.com/CodeWithBehnam/mlx-uniface.git
cd mlx-uniface
pip install -e ".[mlx]"

Quick Start

Face Detection

import cv2
from uniface import RetinaFace

# Automatically uses MLX on Apple Silicon
detector = RetinaFace()

image = cv2.imread("image.jpg")
faces = detector.detect(image)

for face in faces:
    bbox = face['bbox']  # [x1, y1, x2, y2]
    confidence = face['confidence']
    landmarks = face['landmarks']  # 5-point landmarks
    print(f"Face detected with confidence: {confidence:.2f}")

Face Recognition

from uniface import ArcFace, RetinaFace
from uniface import compute_similarity

detector = RetinaFace()
recognizer = ArcFace()

# Detect and extract embeddings
faces1 = detector.detect(image1)
faces2 = detector.detect(image2)

embedding1 = recognizer.get_normalized_embedding(image1, faces1[0]['landmarks'])
embedding2 = recognizer.get_normalized_embedding(image2, faces2[0]['landmarks'])

# Compare faces
similarity = compute_similarity(embedding1, embedding2)
print(f"Similarity: {similarity:.4f}")

Age & Gender Detection

from uniface import RetinaFace, AgeGender

detector = RetinaFace()
age_gender = AgeGender()

faces = detector.detect(image)
gender, age = age_gender.predict(image, faces[0]['bbox'])
print(f"{'Female' if gender == 0 else 'Male'}, {age} years old")

Supported Models

Detection

Recognition

Attributes

Backend Selection

MLX-UniFace automatically selects the best backend:

1. Apple Silicon + MLX installed → Uses MLX (fastest)
2. Otherwise → Uses ONNX Runtime

Force a Specific Backend

import os
os.environ['UNIFACE_BACKEND'] = 'mlx'   # Force MLX
os.environ['UNIFACE_BACKEND'] = 'onnx'  # Force ONNX

from uniface import RetinaFace
detector = RetinaFace()  # Uses the specified backend

Check Current Backend

from uniface.backend import get_backend, Backend

backend = get_backend()
print(f"Using: {backend}")  # Backend.MLX or Backend.ONNX

Benchmarks

Run benchmarks on your hardware:

# Quick benchmark
python scripts/test_mlx_detection.py

# Full benchmark with visualization
jupyter notebook notebooks/benchmark_mlx_vs_onnx.ipynb

Numerical Parity Verification

python scripts/verify_numerical_parity.py

Output:

✓ SUCCESS: MLX and ONNX outputs match within tolerance!
  All outputs have correlation > 0.999

Development

Setup

git clone https://github.com/CodeWithBehnam/mlx-uniface.git
cd mlx-uniface
pip install -e ".[dev]"

Run Tests

pytest

Code Formatting

ruff format .
ruff check . --fix

Project Structure

mlx-uniface/
├── uniface/
│   ├── detection/       # Face detection (RetinaFace, SCRFD, YOLOv5)
│   │   ├── retinaface.py      # ONNX implementation
│   │   ├── retinaface_mlx.py  # MLX implementation
│   │   └── ...
│   ├── recognition/     # Face recognition (ArcFace, MobileFace)
│   ├── landmark/        # 106-point landmarks
│   ├── attribute/       # Age, Gender, Emotion
│   ├── nn/              # MLX neural network modules
│   │   ├── backbone.py  # MobileNetV1/V2
│   │   ├── conv.py      # Conv layers with fused BatchNorm
│   │   ├── fpn.py       # Feature Pyramid Network
│   │   └── head.py      # Detection heads
│   ├── backend.py       # Backend auto-selection
│   ├── mlx_utils.py     # MLX utilities
│   └── onnx_utils.py    # ONNX utilities
├── scripts/
│   ├── convert_onnx_to_mlx.py    # Weight conversion
│   ├── verify_numerical_parity.py # Accuracy validation
│   └── test_mlx_detection.py      # End-to-end tests
├── notebooks/
│   └── benchmark_mlx_vs_onnx.ipynb
└── weights_mlx/         # Pre-converted MLX weights

Credits

• Original UniFace: yakhyo/uniface by Yakhyokhuja Valikhujaev
• MLX Framework: Apple MLX
• Model Architectures:
  • RetinaFace
  • SCRFD
  • ArcFace

License

MIT License - see LICENSE for details.

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

• Issues: GitHub Issues
• Original Project: yakhyo/uniface