OmniVBench: Omni Reference-to-Video 생성을 위한 벤치마크 및 대규모 데이터셋
본 논문은 더욱 범용적이고 다재다능한 제어를 지향하는 omni Reference-to-Video (R2V) 생성 패러다임에 대응하기 위해 OmniVBench와 Omni-R2V Dataset을 제안합니다. 기존 벤치마크가 가진 제한적인 reference type과 holistic한 평가 방식의 한계를 극복하기 위해 7개의 task families와 18개의 fine-grained tasks를 포함하는 광범위한 평가 체계를 구축하였습니다. 또한, factor-grounded evaluation 방식을 도입하여 reference factor가 적절히 보존, disentanglement, routing 되었는지를 정밀하게 측정합니다.
Reference-to-video (R2V) generation is evolving toward increasingly general and versatile reference control, giving rise to the emerging paradigm of omni R2V generation. However, existing benchmarks fall short of these emerging capabilities: their test cases cover limited reference types and compositions, and their evaluation protocols largely assess holistic reference consistency, overlooking whether reference factors are properly preserved, disentangled, and routed. Meanwhile, the high cost of constructing omni R2V training data makes suitable training resources scarce. To address these gaps, we introduce OmniVBench and the Omni-R2V Dataset for evaluating and training omni R2V models. OmniVBench expands R2V evaluation across broader reference types, fine-grained control tasks, and richer reference compositions, covering 7 task families and 18 fine-grained tasks spanning content, motion, style, structure, narrative, and multi-reference settings. We introduce factor-grounded evaluation