Dalision/Omni2Sound_Result

Omni2Sound Evaluation Results

CVPR 2026 (Highlight)

Overview

This repository contains the evaluation results of Omni2Sound on three sub-tasks:

• VT2A (Video + Text → Audio)
• V2A (Video → Audio)
• T2A (Text → Audio)

All results are evaluated on the VGGSound-Omni benchmark and stored as JSON files for reproducibility.

Evaluation Setup

Benchmark: Dalision/Omni2Sound_Benchmark (VGGSound-Omni)

Evaluation Toolkit: AV-Benchmark — the standardized evaluation toolkit from MMAudio, applied on 8-second clips following prior work.

Metrics cover four dimensions:

Dimension	Metrics
Distribution Matching	FAD, FD, FD_PaSST, KL, KL_PaSST
Audio Quality	IS, IS_PaSST, PQ (Production Quality)
Semantic Alignment	CLAP, MS-CLAP (text-audio), IB / ImageBind (video-audio)
Temporal Alignment	DS / Desynchronization Score (Synchformer)

All baseline models are re-evaluated using their official checkpoints with the same standardized toolkit and identical video/text conditions for fair comparison.

Links

• Model: Dalision/Omni2Sound
• Benchmark & Dataset: Dalision/Omni2Sound_Benchmark
• Evaluation Toolkit: hkchengrex/av-benchmark
• Paper: arXiv:2601.02731
• Project Page: omni2sound.github.io
• Code: github.com/omni2sound/Omni2Sound

Citation

@article{dai2026omni2sound,
  title   = {Omni2Sound: Towards Unified Video-Text-to-Audio Generation},
  author  = {Dai, Yusheng and Chen, Zehua and Jiang, Yuxuan and Gao, Baolong and
             Ke, Qiuhong and Cai, Jianfei and Zhu, Jun },
  journal = {arXiv preprint arXiv:2601.02731},
  year    = {2026}
}

License

Released under CC BY-NC 4.0 (non-commercial use only).