Researchers have introduced VA-Judger, a reward model for post-training video-audio joint generation systems, along with the VAPref-10K dataset of 10.3 thousand pairwise comparisons and the VA-Judger-Bench benchmark. The work was published on August 19, 2026 on
arXiv. The authors note that existing metrics evaluate audio quality, visual accuracy, and synchronization separately and fail to capture the overall semantic and temporal coherence between the text prompt, video, and audio. VA-Judger is trained in stages: first on pairs with an explicit quality gap, then on more complex comparisons via rejection sampling with verification against human annotations, followed by dimension-wise reinforcement learning. The code is released as open source.
What it means
The news relates to video generation in general — our
AI Video section collects text-to-video and image-to-video tools. VA-Judger is not a generator but a reward model for evaluating and post-training generative systems, so it does not directly overlap with our tool pages.
Among the reference tools, the closest in task to text-to-video are
CogVideoX and
Pyramid Flow. Both are open-source video generators, but their fact sheets contain no information about audio support or the use of reward models for post-training.
LTX-Video handles the image-to-video task and also does not mention audio processing.
Joint video and audio generation is a separate direction not yet covered in our fact sheets. If VA-Judger gains traction as a post-training component, it could affect future versions of generators, but the news does not impact the current tool pages in the reference.
