VBVR-Pro-Bench / README.md
junxiangjoe's picture
Align badges with scorer README
a056972 verified
|
Raw
History Blame Contribute Delete
7.9 kB
metadata
license: cc-by-nc-4.0
language:
  - en
tags:
  - video-generation
  - image-to-video
  - visual-reasoning
  - benchmark
pretty_name: VBVR-Pro-Bench
size_categories:
  - n<1K
configs:
  - config_name: video
    default: true
    data_files:
      - split: test
        path: preview/video.parquet
  - config_name: image
    data_files:
      - split: test
        path: preview/image.parquet

VBVR-Pro-Bench

The frozen evaluation split of VBVR-Pro. 100 parameterized reasoning tasks, 5 held-out instances each, in both a video (I2V) and an interleaved-image setting.

At a glance

Property Value
Tasks 100 (In-Domain_50 50 + Out-of-Domain_50 50)
Instances per task 5
Total instances 500 per setting (1,000 across both)
Settings Video and Image
Resolution 1024 × 1024
Video 16 fps, MPEG-4 Part 2, as generated (no re-encode)
Size 86.0 MB (video) + 11.7 MB (image)

In-Domain_50 covers task families that also appear in the training splits; Out-of-Domain_50 covers task families that do not, and is the split to report generalization on.

Layout

.
├── VBVR-Pro-Bench-Video.tar.gz
└── VBVR-Pro-Bench-Image.tar.gz

VBVR-Pro-Bench-Video.tar.gz extracts to:

VBVR-Pro-Bench-Video/
├── In-Domain_50/
│   ├── G-131_select_next_figure_increasing_size_sequence_data-generator/
│   │   ├── 00000/
│   │   │   ├── first_frame.png        # conditioning image  (1024 × 1024)
│   │   │   ├── prompt.txt             # instruction
│   │   │   ├── ground_truth.mp4       # reference video     (16 fps)
│   │   │   ├── final_frame.png        # last frame of the reference video
│   │   │   └── metadata.json          # task parameters, ground truth, scoring contract
│   │   ├── 00001/ … 00004/            # same five files
│   │   └── …
│   └── …                              # 50 tasks
└── Out-of-Domain_50/                  # 50 tasks, same shape

VBVR-Pro-Bench-Image.tar.gz extracts to the same tree, with the reference output as an image sequence instead:

VBVR-Pro-Bench-Image/
├── In-Domain_50/
│   ├── G-131_select_next_figure_increasing_size_sequence_data-generator/
│   │   ├── 00000/
│   │   │   ├── first_frame.png        # conditioning image  (1024 × 1024)
│   │   │   ├── prompt.txt             # instruction
│   │   │   ├── frame_1.png            # reference output, step 1
│   │   │   ├── frame_2.png            # … step 2, when the task has more steps
│   │   │   ├── …                      # up to frame_N.png
│   │   │   └── metadata.json          # task parameters, ground truth, scoring contract
│   │   ├── 00001/ … 00004/
│   │   └── …
│   └── …                              # 50 tasks
└── Out-of-Domain_50/                  # 50 tasks, same shape

N is task-dependent (1–21): it is the number of steps the instruction asks the model to render, and it is stated in prompt.txt (e.g. "Output 4 images: …"). Most tasks are single-step.

Usage

huggingface-cli download Video-Reason/VBVR-Pro-Bench --repo-type dataset --local-dir ./VBVR-Pro-Bench
cd VBVR-Pro-Bench && tar xzf VBVR-Pro-Bench-Video.tar.gz

Generate one video (or image sequence) per instance from first_frame.png + prompt.txt, then score with VBVR-Pro-Bench.

License

VBVR-Pro source code, scripts, configuration files and task-specific scoring software — including everything in this repository — are licensed under the Apache License 2.0. VBVR-Pro data and benchmark materials are separately licensed under CC BY-NC 4.0. Model weights and third-party materials remain subject to their applicable model-card and upstream terms. See LICENSE.md for details.

Citation

@misc{xu2026vbvrproscalableverifiablesuite,
      title={VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning},
      author={Junxiang Xu and Ruisi Wang and Fanyi Pu and Maijunxian Wang and Ran Ji and Tongxi Zhou and Chenyang Gu and Jing Zuo and Hongcan Xiao and Yimeng Geng and Wanqi Yin and Wei Chen and Oscar Qian and Zhengan Yan and Ziqi Huang and Haiwen Diao and Liang Pan and Bo Li and Xiangyu Fan and Dezhi Luo and Fengyuan Yu and Zehong Zhao and Qingying Gao and Tinghui Zhu and Yilan Zhang and Jingqi Tong and Pinyuan Feng and Zhengze Jiang and Letian Wang and Ziyu Guo and Renrui Zhang and Jieneng Chen and Sonia Joseph and Constantin Venhoff and Saman Motamed and Mengyue Yang and Chandra Sripada and Alan Yuille and Philip Torr and Lvmin Zhang and Vikash Kumar and Daniel Khashabi and Nikolaus Kriegeskorte and Rapha\"{e}l Milli\`{e}re and Vincent C. M\"{u}ller and Anyi Rao and Quan Wang and Ziwei Liu and Dahua Lin and Lei Yang and Hokin Deng and Zhongang Cai},
      year={2026},
      eprint={2608.26105},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.26105},
}