tencent-hunyuan/omnicapbench is a caption evaluation framework that handles audio and video together. The repository introduces tools and evaluation material for validating video descriptions split into object references, visual scenes, and audio events. The repository description mentions NeurIPS 2026, and the latest release version is not given in the provided information. It is written in Python and uses the Apache-2.0 license.
Instead of assigning a single score to an entire caption, the framework is built to split descriptions into verifiable units so that you can see where the differences arise. In the validation stage it checks the structure and time range of each unit along with the links between tracks, and then evaluates the meaning of each unit aligned in time. That puts the focus on locating errors that a total score alone would not reveal. The material presented in the README consists of 786 videos, 5,818 references, 11,419 scenes, 39,160 sub-scenes, and 6,537 audio events. The README also explains the annotation procedure along with installation and usage examples for the scoring tool.
That said, the provided FACTS present no comparison with other evaluation frameworks and no real adoption cases. The GitHub star count is 7. The material is therefore useful as a reference on which units and procedures a structured audio and video caption evaluation covers, but this information alone makes it difficult to judge performance or field applicability.