• 🎯 Introduction
• 🏆 Challenge Results
• 📦 Installation
• 📊 Dataset
• 🔧 Training
• 🧪 Testing
• 📚 Citation

Face Image Quality Assessment (FIQA) is crucial for various face-related applications such as face recognition, face detection, and biometric systems. While significant progress has been made in FIQA research, the computational complexity remains a key bottleneck for real-world deployment.

This repository presents Efficient-FIQA, a novel approach that achieves state-of-the-art performance with extremely low computational overhead through:

• 🔬 Self-training Strategy: Enhances teacher model capacity using pseudo-labeled data
• 🎓 Knowledge Distillation: Transfers knowledge from powerful teacher to lightweight student
• ⚡ Efficient Architecture: Student model achieves comparable performance with minimal computational cost

• 🥇 1st Place in ICCV VQualA 2025 FIQA Challenge

Score = (SRCC + PLCC) / 2

For more results on the ICCV VQualA 2025 FIQA Challenge, please refer to the challenge report.

• Python >= 3.9
• PyTorch >= 1.13
• CUDA >= 11.0 (for GPU training)

# Create and activate conda environment
conda create -n EfficientFIQA python=3.9
conda activate EfficientFIQA

# Install other dependencies
pip install -r requirements.txt

dataset/
├── train/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
├── val/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
└── train.csv  # Format: image_name,score

1. Configure paths in config_SwinB.py:

# Data paths
data_dir: str = '/path/to/your/training/images'
csv_path: str = 'data_file/train.csv'

# Model save path
model_save_dir: str = '/path/to/save/teacher/model'

2. Start training:

python train_teacher_model.py

The teacher model is first used to generate pseudo-labels for unlabeled images to enhance training data. Since we cannot provide the original unlabeled images due to copyright restrictions, we use GFIQA-20K as a representative example dataset.

1. Generate pseudo-labels using teacher model:

# Configure in test_unlabeled_images.py
image_dir = '/path/to/GFIQA/images'
output_csv = 'data_file/gfiqa_results.csv'

python test_unlabeled_images.py

2. Configure student training in config_Edgenet.py:

# Data paths
data_dir: str = '/path/to/original/training/images'
gfiqa_data_dir: str = '/path/to/GFIQA/images'
csv_path: str = 'data_file/train.csv'
gfiqa_csv_path: str = 'data_file/gfiqa_results.csv'

# Model save path
model_save_dir: str = '/path/to/save/student/model'

3. Start student training:

python train_student_model.py

• Student Model (Winner): EdgeNeXt-XXS
• Teacher Model: Swin-B
• Teacher+ Model: Swin-B+

python test.py \
    --model_name FIQA_EdgeNeXt_XXS \
    --model_weights_file ckpts/EdgeNeXt_XXS_checkpoint.pt \
    --image_file your_image.jpg \
    --image_size 352 \
    --gpu_ids 0

# Test with student model (recommended)
python test.py \
    --model_name FIQA_EdgeNeXt_XXS \
    --model_weights_file ckpts/EdgeNeXt_XXS_checkpoint.pt \
    --image_file demo_images/z06399.png \
    --image_size 352 \
    --gpu_ids 0

# Test with teacher model
python test.py \
    --model_name FIQA_Swin_B \
    --model_weights_file ckpts/Swin_B_plus_checkpoint.pt \
    --image_file demo_images/z06399.png \
    --image_size 448 \
    --gpu_ids 0

# CPU inference
python test.py \
    --model_name FIQA_EdgeNeXt_XXS \
    --model_weights_file ckpts/EdgeNeXt_XXS_checkpoint.pt \
    --image_file demo_images/z06399.png \
    --image_size 352 \
    --gpu_ids cpu

Try our online demo on Hugging Face Spaces:

If you find this work useful for your research, please cite our paper:

@inproceedings{sun2025efficient,
  title={Efficient Face Image Quality Assessment via Self-training and Knowledge Distillation},
  author={Sun, Wei and Zhang, Weixia and Cao, Linhan and Jia, Jun and Zhu, Xiangyang and Zhu, Dandan and Min, Xiongkuo and Zhai, Guangtao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision Workshops},
  pages={1-9},
  year={2025}
}