Face Image Quality Assessment

Tip

The models and functionality in this repository are integrated into UniFace — an all-in-one face analysis toolkit.

PyTorch inference, ONNX export, and ONNX Runtime inference for eDifFIQA — a face image quality estimator that predicts a single quality score from an aligned 112×112 face. Uses UniFace for face detection and alignment.

Models

Variant	Backbone	Params	Input	PyTorch (.pth)	ONNX
eDifFIQA-T	MobileFaceNet	1.7M	112x112	Link	Link
eDifFIQA-S	IResNet-18	24.6M	112x112	Link	Link
eDifFIQA-M	IResNet-50	44.1M	112x112	Link	Link
eDifFIQA-L	IResNet-100	65.7M	112x112	Link	Link

Higher score = better quality. Authors evaluate against 10 FIQA competitors on 7 static-image datasets + 1 video dataset using EDC curves; LFW and XQLFW are the commonly cited references for high-quality vs. low-quality regimes. eDifFIQA-L ranks first on the NIST FATE-Quality Kiosk-to-Entry track. See the paper for per-dataset numbers.

.pth weights are ported from the upstream authors' OneDrive folder; .onnx files are exported from those weights and mirrored to this repo's release.

Installation

pip install -r requirements.txt
bash download.sh

Demo

eDifFIQA-T quality scores on assets/test_images/. The same face, progressively degraded, gets a sharply lower score:

Original	Down-up sampled	Heavy blur
0.762	0.435	0.144
0.737	0.537	0.273

Inference

Inputs are aligned 112×112 face crops. Pass a full image and alignment is handled for you; use --aligned or score_aligned() to skip it for already-aligned crops.

PyTorch

python main.py assets/test_images/image2.jpg --variant s

Pre-aligned 112×112 crops:

python main.py --aligned aligned_face.jpg --variant s

ONNX

python onnx_inference.py --model weights/ediffiqa_s.onnx assets/test_images/image2.jpg

import cv2
from models import eDifFIQAOnnx
from uniface.detection import SCRFD

detector = SCRFD()
quality = eDifFIQAOnnx("weights/ediffiqa_s.onnx")

image = cv2.imread("image.jpg")
for face in detector.detect(image):
    score = quality.get_quality(image, face.landmarks)
    print(f"Quality: {score:.4f}")

ONNX Export

python onnx_export.py -v s -w weights/ediffiqa_s.pth -o weights/ediffiqa_s.onnx --dynamic

Reference

• eDifFIQA — Original PyTorch implementation and paper
• UniFace — Face detection and alignment

License

MIT License