We released OmniVBench and Omni-R2V, a benchmark and a training set for reference-to-video. Both open for research. Omni-R2V is 340K ready-to-train samples, built primarily from licensed footage. Covering 7 task families and 18 fine-grained tasks, the dataset spans content, motion, style, structure, narrative and multi-reference compositions. OmniVBench scores reference fidelity, instruction realization and video quality separately, instead of one holistic similarity. Across 11 models, content averages 72, compared to 58 for motion and 57 for style, revealing uneven performance across reference types. Its factor-grounded evaluation helps pinpoint failures in reference preservation, factor disentanglement, and instruction following.