FVQ: A Large-Scale Dataset and an LMM-based Method for Face Video Quality Assessment (ACM MM 2025)

February 5, 2026 · View on GitHub

Paper | Video | Dataset

We explore the in-the-wild face video quality assessment problem for the first time. Concretely, we present FVQ-20K, the first large-scale in-the-wild face video quality assessment dataset, which contains 20,000 face videos with (a) diverse source video content, (b) various distortions in both spatial and temporal domains, (c) a variety of facial attributes, and (d) high-quality MOS annotation for each video. Along with the FVQ-20K dataset, we propose FVQ-Rater, the first LMM-based method elaborately designed for the face video quality assessment task as illustrated in (e).

📌 TODO

  • Release the training code of FVQ-Rater.
  • Release the evaluation code of FVQ-Rater.
  • Release the FVQ-20K dataset.

✨ FVQ-20K Dataset

Download

The dataset can be downloaded from Hugging Face.

Overview

  • FVQ-20K is an in-the-wild face video quality assessment (FVQA) dataset, which contains 20,000 face videos with MOS annotations.
  • The FVQ-20K dataset is divided into training, validation, and test sets with a ratio of 80% : 5% : 15%.

Data Structure

FVQ-20K
│
├── train
│   ├── labels.txt
│   └── videos
│        ├── *.mp4
│        └── ...
├── val
│   ├── labels.txt
│   └── videos
│        ├── *.mp4
│        └── ...
└── test
    ├── labels.txt
    └── videos
         ├── *.mp4
         └── ...

• labels.txt contains video names and their corresponding MOS scores (ranging from 0 to 100).

🎯 FVQ-Rater Method

🛠 Installation

⚡ Training and Evaluation on FVQ-20K

🚀 Training and Evaluation on Custom Datasets

📄 Citation

If you find our dataset or code helpful, please consider citing

@inproceedings{wu2025fvq,
  title={FVQ: A Large-Scale Dataset and an LMM-based Method for Face Video Quality Assessment},
  author={Wu, Sijing and Li, Yunhao and Xu, Ziwen and Gao, Yixuan and Duan, Huiyu and Sun, Wei and Zhai, Guangtao},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={6928--6937},
  year={2025}
}

🤗 Contact