FaceBench

<div align="center"> <h2>FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs [CVPR 2025]</h2> Xiaoqin Wang, Xusen Ma, Xianxu Hou, Meidan Ding, Yudong Li, Junliang Chen, Wenting Chen, Xiaoyang Peng, Linlin Shen*

</div>

Overview

In this work, we introduce FaceBench, a dataset featuring hierarchical multi-view and multi-level attributes specifically designed to assess the comprehensive face perception abilities of MLLMs. We construct a hierarchical facial attribute structure, which encompasses five views with up to three levels of attributes, totaling over 210 attributes and 700 attribute values. Based on the structure, the proposed FaceBench consists of 49,919 visual question-answering (VQA) pairs for evaluation and 23,841 pairs for fine-tuning. Moreover, we further develop a robust face perception MLLM baseline, Face-LLaVA, by training with our proposed face VQA data.

<div align="center"><img src="./assets/overview.png" width="100%" height="100%"></div>

Distribution of visual question-answer pairs

<div align="center"><img src="./assets/VQAs.jpg" width="100%" height="100%"></div>

Some samples from our dataset

<div align="center"><img src="./assets/example.png" width="100%" height="100%"></div>

News

• [2024-08-20] The Face-LLaVA model is released on HuggingFace🤗.
• [2024-03-27] The paper is released on ArXiv🔥.

TODO

• [X] Release the Face-LLaVA model.
• [X] Release the evaluation code.
• [X] Release the dataset.

Evaluation

Model inference

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=0 python evaluation/inference.py \
    --data-dir ./datasets/example/test.jsonl \
    --images-dir ./datasets/example/images/ \
    --model-name face_llava_1_5_13b \
    --question-type "TFQ, SCQ, MCQ, OEQ" \
    --save-dir "./responses-and-results/"

Calculate metrics

OMP_NUM_THREADS=8 CUDA_VISIBLE_DEVICES=5 python evaluation/evaluation.py \
    --data-path ./responses-and-results/face_llava_1_5_13b_test_responses.jsonl"

Results

Experimental results of various MLLMs and our Face-LLaVA across five facial attribute views.

<div align="center"><img src="./assets/five-view-results.jpg" width="100%" height="100%"></div>

Experimental results of various MLLMs and our Face-LLaVA across Level 1 facial attributes.

<div align="center"><img src="./assets/level-1-results.jpg" width="100%" height="100%"></div>

Citation

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

@inproceedings{wang2025facebench,
  title={FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs},
  author={Wang, Xiaoqin and Ma, Xusen and Hou, Xianxu and Ding, Meidan and Li, Yudong and Chen, Junliang and Chen, Wenting and Peng, Xiaoyang and Shen, Linlin},
  booktitle={Proceedings-2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025},
  year={2025}
}

@article{wang2025facebench,
  title={FaceBench: A Multi-View Multi-Level Facial Attribute VQA Dataset for Benchmarking Face Perception MLLMs},
  author={Wang, Xiaoqin and Ma, Xusen and Hou, Xianxu and Ding, Meidan and Li, Yudong and Chen, Junliang and Chen, Wenting and Peng, Xiaoyang and Shen, Linlin},
  journal={arXiv preprint arXiv:2503.21457},
  year={2025}
}

If you have any questions, you can either create issues or contact me by email wangxiaoqin2022@email.szu.edu.cn.

Acknowledgments

This work is heavily based on LLaVA. Thanks to the authors for their great work.