REFERENCE-TO-VIDEO GENERATION

Omni-R2V&OmniVBench

A benchmark and large-scale dataset
for omni reference-to-video generation.

HETEROGENEOUS REFERENCES
FACTOR-GROUNDED EVALUATIONSCROLL TO EXPLORE ↓

01 / DATASET OVERVIEW

Omni-R2V Dataset

Comprehensive Coverage of R2V Tasks

The first large-scale public R2V training dataset covering heterogeneous reference types across seven task families, spanning single-reference tasks and multi-reference compositions.

Industry-Grade Training Data

340K processed training samples built primarily from professional video footage, spanning live-action, 2D animation, and 3D animation.

Reusable R2V Construction Pipelines

Task-specific pipelines construct reference–target pairs and corresponding training instructions, providing a reusable recipe for omni-R2V data construction.

How Omni-R2V compares with existing R2V datasets

Dataset	# Size	Reference Modality		Reference Task Coverage							Processed Data Provided
Dataset	# Size	Image	Video	Content	Motion	Style	Structure	Narrative	Multi-Content	Cross-Aspect	Processed Data Provided
OpenS2V-5M	5.4M	✓		✓					✓		✓
Phantom-Data	1M	✓		✓					✓
MuSS	30K	✓		✓
Omni-R2V (Ours)	340K	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓

Omni-R2V is the only dataset providing processed image and video references across all seven task families.

02 / SINGLE-REFERENCE TASKS

Single-reference examples

Target video

Training instruction

03 / MULTI-REFERENCE COMPOSITION

Multi-reference examples

Target video

Training instruction

Dataset statistics

View full size ↗

04 / CITATION

@article{omnivbench2026,
  title   = {OmniVBench: A Benchmark and Large-Scale Dataset for Omni Reference-to-Video Generation},
  author  = {Li, Wenxue and Guan, Peiyan and Jiang, Haoyang and Cai, Junxian and
             Liu, Hualuo and Zhang, Chunjie and Guan, Chong and Huang, Kai and
             Li, Songlian and
             Wu, Taiyi and Yu, Yongjian and Zhao, Xiaotong and
             Zhao, Alan and Liu, Eric and Chen, Xi and Liu, Yu and Zhu, Lei},
  journal = {arXiv preprint arXiv:2609.22069},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.22069}
}