OmniVBench addresses the lack of adequate evaluation for omni reference-to-video generation, where models must preserve, disentangle, and correctly route many different types of reference input. It introduces 18 fine-grained tasks across 7 task families and a factor-grounded evaluation protocol with over 12,000 case-specific checklist items that check whether each reference factor is faithfully preserved and correctly bound. The accompanying Omni-R2V Dataset provides 340K training samples built via task-specific reference-target pipelines, and evaluation of current open- and closed-source models on the benchmark reveals clear, previously invisible performance gaps.
