Key Info
Tencent released OmniVBench, a benchmark for reference-to-video generation, alongside Omni-R2V, a 340K-sample ready-to-train dataset built primarily from licensed footage. Both are open for research.
Highlights
- Omni-R2V contains 340K training samples covering 7 task families and 18 fine-grained tasks, spanning content, motion, style, structure, narrative and multi-reference compositions.
- OmniVBench evaluates reference fidelity, instruction realization and video quality separately rather than as a single holistic similarity score.
- Across 11 models, average scores were 72 for content, 58 for motion and 57 for style, showing uneven performance across reference types.
- Its factor-grounded evaluation is designed to pinpoint specific failures in reference preservation.
- Dataset, benchmark and paper links were provided in the thread.
