OmniVBench

A Benchmark for Omni Reference-to-Video Generation

813 Evaluation Cases  ·  7 Task Families  ·  18 Subtasks

OmniVBench evaluates reference-to-video generation with bilingual instructions, 1,041 reference images, and 421 reference videos (approximately 3.96 GB).

The task families are Content, Motion, Style, Structure, Narrative, Multi-content, and Cross-aspect. This release provides generation inputs and reference media. Evaluation materials and model outputs will be released separately.

Data layout

task_json/<family>/<subtask>.json   # 18 annotation files
task_assets/<family>/...           # Reference images and videos

Every JSON contains task_family, subtask, path_base, sample_count, and a flat samples list.

Sample field	Description
case_id	Globally unique ID: task_id/sample_id. Use it to identify model predictions.
task_id, sample_id	Task identifier and sample identifier within that task.
category, modalities	Sample category (null if unspecified) and input modalities.
inputs.text_en, inputs.text_cn	English and Chinese generation instructions.
inputs.image, inputs.video	Ordered reference paths; either list may be absent.
inputs.reference_mapping	Optional reference-role annotations in English.

The 18 subtask files group 20 task IDs. Content + Structure and Content + Narrative each contain two task IDs; case_id uniquely identifies every sample.

Quick start

Download the repository, then read any annotation file using Python:

import json
from pathlib import Path

root = Path("/path/to/OmniVBench")
doc = json.loads((root / "task_json/motion/action.json").read_text())
sample = doc["samples"][0]

references = {
    kind: [root / doc["path_base"] / p
           for p in sample["inputs"].get(kind, [])]
    for kind in ("image", "video")
}
print(sample["case_id"], sample["inputs"]["text_en"])
print(references)

Media paths resolve as repository_root / path_base / reference_path, where path_base is task_assets. Use the provided English or Chinese instruction and preserve the reference-list order, including repeated entries.

Reference roles

reference_mapping, available for 473 cases, describes the subjects or factors supplied by each reference. subject_en describes the subject; role_en describes its intended role. entity_id, when present, links views of the same entity within a sample.

ref_index uses one-based numbering within the image or video list. ref_type takes precedence; otherwise, type containing video or ref_index: "video" identifies a video reference ("video" means the first video). Other indices refer to images. These annotations help interpret the references; generation uses the provided prompt and ordered media.

License

The annotations and reference media are available for non-commercial academic research only. Please acknowledge OmniVBench and cite the paper. See Terms of Use for permitted uses and attribution requirements.