UniVA-Bench Leaderboard

UniVA-Bench is an agent-oriented benchmark for unified video intelligence, covering Understanding, Generation, Editing, Segmentation, and agentic probing. We report CLIP / DINO / MLLM preference, segmentation J/F/J&F, and long-video QA accuracy, following the evaluation protocol described in our paper.

Table 1: Comparison across LongText2Video, Entities2Video and Video2Video


LTX-Video	0.281	0.903	3.333	0.287	0.88	1.789	0.226	0.894	4.068


UniVA	0.2814	0.9026	3.333	0.2868	0.8796	1.789	0.262	0.8939	4.068
LTX-Video	0.2161	0.9392	1.125	0.221	0.8452	1.281	0.2263	0.9943	2.123
Wan	0.2028	0.6779	3.183	0.3106	0.7043	1.65	0.2632	0.9188	2.034
Seedance	0.2157	0.8836	2.65	0.3039	0.88	2.7	0.2684	0.9518	2.621