PhyAV-Sound-11K

This is the official release of the PhyAV-Sound-11K dataset introduced in the paper: PhyAVBench: A Challenging Audio Physics-Sensitivity Benchmark for Physically Grounded Text-to-Audio-Video Generation.

All videos in this dataset are released without audio. To prevent potential data leakage, the ground-truth audio files are not publicly released. Instead, we provide the corresponding ground-truth audio embeddings (CLAP and ImageBind) to support reproducible evaluation.

Data Organization

• prompt_all.jsonl: 337 variable-controlled prompt groups.
• videos_gt_silent_part_0-1: 11,605 silent videos.
• ref.lst: a list of reference videos, mainly used for video-to-audio generation.
• ref_first_frames: first-frame images extracted from the reference videos, mainly used for image-to-video generation.
• embeddings: 377 embedding direction vectors (CLAP and ImageBind), computed from the audio of all 11,605 ground-truth videos and averaged within each group.

Evaluation Code

Please visit https://github.com/imxtx/PhyAVBench, which provides:

• Evaluation code.
• Evaluation prompts.
• Ground-truth audio embedding direction vectors.

Citation

If you find this dataset useful, please consider citing our paper:

@article{xie2025phyavbench,
  title={Phyavbench: A challenging audio physics-sensitivity benchmark for physically grounded text-to-audio-video generation},
  author={Xie, Tianxin and Lei, Wentao and Jiang, Kai and Huang, Guanjie and Zhang, Pengfei and Zhang, Chunhui and Ma, Fengji and He, Haoyu and Zhang, Han and He, Jiangshan and Wang, Jinting and Fang, Linghan and Gao, Lufei and Ablet, Orkesh and Zhang, Peihua and Hu, Ruolin and Li, Shengyu and Lin, Weilin and Feng, Xiaoyang and Yang, Xinyue and Rong, Yan and Wang, Yanyun and Shao, Zihang and Zhao, Zelin and Li, Chenxing and Yang, Shan and Wang, Wenfu and Yu, Meng and Yu, Dong and Liu, Li},
  journal={arXiv preprint arXiv:2512.23994},
  year={2025}
}