Omni-R2V Dataset

Large-scale training data for omni reference-to-video generation

Overview · Task coverage · Quick start · Data format · Citation

_{From individual reference factors to multi-content and cross-aspect combinations.}

Overview

For evaluation, download the companion OmniVBench benchmark, which provides generation instructions and reference media for 813 evaluation cases.

Omni-R2V provides 339,570 processed training samples across seven reference families and 13 subtasks. Given one or more image or video references, a model follows an instruction to generate a target video while preserving the specified reference factors.

	What the dataset provides
🧩 Broad reference coverage	Content, motion, style, structure, narrative, multi-content, and cross-aspect settings in one dataset.
🖼️ Reusable reference collections	Multiple candidate references are retained where available, supporting single-reference selection, multi-reference conditioning, and alternative training pairs.
💬 Bilingual instructions	English and Chinese prompts describe the requested video and identify the role of each reference.
📦 Self-contained samples	References and the target video are stored together, with relative paths in the annotations.

Pairs are built through task-specific cross-pair matching, inverse construction, and complementary composition. Instructions identify the supplied references and the factors they should constrain.

Multiple references, flexible training pairs

We retain multiple candidate references where available, preserving the reference collection associated with each target video. Organizing these candidates alongside the target and reference-aware instructions makes the collected data reusable across different conditioning setups. The released sample count describes the stored annotation records; additional training pairs can be constructed from suitable reference candidates.

Training setup	How to use the retained references
Single-reference training	Select one suitable candidate per sample, either once during preprocessing or dynamically during training.
Multiple reference–target pairs	For interchangeable candidates, pair each reference with the same target video to construct separate training examples.
Multi-reference training	Use the available references together, or sample compatible subsets to vary the number of conditioning inputs.

When selecting or splitting references, update the instruction and its <Figure N> / <Video N> tokens to match the retained inputs. Candidates describing the same subject may be used as alternatives; references that supply complementary subjects or factors in multi-content and cross-aspect tasks should be retained together when required by the instruction. Derived pairs share a target video and should stay in the same data split.

Task coverage

Family	Sub-task	Samples	What the reference constrains
Content	content	18,615	Subject identity (person)
Multi-content	multi_content	19,497	Several subjects (person / object / environment)
Motion	action	29,929	Subject motion
Motion	camera_motion	27,471	Camera movement
Style	style	8,842	Visual style
Structure	lineart	19,243	Single-shot geometry (line art / edges)
Structure	greybox	9,880	Single-shot geometry (greybox layout)
Structure	rough_storyboard	9,714	Single-shot geometry (rough storyboard)
Narrative	preceding_shot	67,882	Cross-shot continuity (continue a previous shot)
Narrative	multi_panel_storyboard	99,999	Cross-shot sequence (multi-panel storyboard)
Cross-aspect	content_lineart	9,682	Subject + line art
Cross-aspect	content_storyboard	9,647	Subject + rough storyboard
Cross-aspect	content_style	9,169	Subject + style

Quick start

Annotations are UTF-8 JSONL files: one record per sample. All media paths are relative to the dataset root. The loader below uses only the Python standard library.

import json
from pathlib import Path

root = Path("/path/to/OmniR2V-Data")


def iter_samples(annotation):
    with (root / "meta" / annotation).open(encoding="utf-8") as handle:
        for line in handle:
            if line.strip():
                yield json.loads(line)


sample = next(iter_samples("content/content.jsonl"))
references = [
    {"role": group["role"], "modality": asset["modality"],
     "path": root / asset["rel_path"]}
    for group in sample["inputs"]
    for asset in group["assets"]
]
target = root / sample["target"]

print(sample["prompt_en"])
print(references)
print(target)

Data format

Directory structure

OmniR2V-Data/
├── README.md
├── figures/
├── meta/                              # 13 annotation files
│   ├── content/content.jsonl
│   ├── motion/
│   │   ├── action.jsonl
│   │   └── camera_motion.jsonl
│   ├── style/style.jsonl
│   ├── structure/
│   │   ├── lineart.jsonl
│   │   ├── greybox.jsonl
│   │   └── rough_storyboard.jsonl
│   ├── narrative/
│   │   ├── preceding_shot.jsonl
│   │   └── multi_panel_storyboard.jsonl
│   ├── multi_content/multi_content.jsonl
│   └── cross_aspect/
│       ├── content_lineart.jsonl
│       ├── content_storyboard.jsonl
│       └── content_style.jsonl
└── assets_tar/                        # Media, packed as tar volumes
    ├── action_0.tar … action_3.tar
    ├── camera_motion_0.tar … camera_motion_11.tar
    └── …                              # 130 volumes in total

Each volume restores the original layout when extracted:

assets/{subtask}/{shard}/{seq}/        # One directory per sample
├── subject_00.jpg                     # Example reference filenames
├── subject_00_1.jpg
└── target.mp4

Media volumes

Media files are packed into 130 tar volumes of roughly 18 GB each, one set per sub-task. Packing keeps the repository at a few hundred files instead of ~848K individual media files, which makes downloads and resuming far more reliable. The archives are uncompressed: extracting a volume costs exactly its archive size in disk space.

Download only what you need:

# all media
hf download wxli318/Omni-R2V --repo-type dataset --include "assets_tar/*"

# a single sub-task
hf download wxli318/Omni-R2V --repo-type dataset --include "assets_tar/action_*"

Extract from the dataset root so that the relative paths in meta/ resolve:

cd /path/to/Omni-R2V
for t in assets_tar/*.tar; do tar -xf "$t"; done

Volumes can be extracted in any order and deleted afterwards to reclaim disk space.

Annotation fields

Field	Description
sample_id	Unique identifier in the form r2v_{subtask}_{seq:06d}.
aspect / subtask	Reference family and subtask.
prompt_en / prompt_cn	English and Chinese instructions.
inputs	Ordered reference groups, each with a role and an assets list.
inputs[].assets[].modality	Reference media type: image or video.
inputs[].assets[].rel_path	Reference path relative to the dataset root.
target	Target video path relative to the dataset root.

View an example annotation

The prompt text below is abbreviated; the two reference images correspond to the first content sample.

{
  "sample_id": "r2v_content_000000",
  "aspect": "content",
  "subtask": "content",
  "prompt_en": "<Subject 1> is the boy in <Figure 1> and <Figure 2> ...",
  "prompt_cn": "<Subject 1> 是 <Figure 1> 和 <Figure 2> 中的少年……",
  "inputs": [
    {
      "role": "subject",
      "assets": [
        {
          "modality": "image",
          "rel_path": "assets/content/00/000000/subject_00.jpg"
        },
        {
          "modality": "image",
          "rel_path": "assets/content/00/000000/subject_00_1.jpg"
        }
      ]
    }
  ],
  "target": "assets/content/00/000000/target.mp4"
}

Reading reference tokens. <Figure N> identifies the N-th image and <Video N> the N-th video, counted separately as assets are traversed in inputs order. <Subject N> is a semantic identifier in the prompt; it does not map one-to-one to an input group.

Use & license

The dataset supports research on reference-conditioned video generation, disentangling factors across heterogeneous references, and multi-reference composition.

Research use only. Commercial use and redistribution of the media are not covered. Annotations and instructions are provided by the authors under the same research-only terms. The detailed TERMS_OF_USE.md document has not yet been included in this directory.

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}
}