README.md

metadata

pretty_name: VWG-Bench
language:
  - en
task_categories:
  - image-to-video
tags:
  - video-generation
  - reasoning
  - benchmark
  - image-to-video
  - eccv-2026
license: cc-by-nc-4.0
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test/**

VWG-Bench

VWG-Bench (Video World Generalist Benchmark) is an image-to-video reasoning benchmark for evaluating whether video generation models can follow symbolic rules, respect physical and commonsense constraints, and realize intended goals over time.

The test split contains 380 initial images and structured annotations organized into 9 reasoning dimensions and 38 ten-sample task groups.

Data layout

data/test/
├── metadata.jsonl
└── images/
    ├── 000000.png
    └── ...
taxonomy.json
evaluation_rules.json
schema.json
checksums.sha256
LICENSE.md
evaluation_code/

The layout follows the Hugging Face ImageFolder convention. Each metadata.jsonl row has a file_name field relative to data/test/. The evaluation_code/ directory contains the cleaned VWG-Bench evaluation toolkit and helper scripts.

Annotation fields

Field	Meaning
id	Stable sample identifier, 0–379
file_name	Relative initial-image path
dimension_id, dimension_name	One of the 9 benchmark dimensions
task_group_id	Stable identifier for one of the 38 ten-sample groups
task_group_name	Current human-readable group name
task_name	Task label present in the source data
user_prompt	Image-to-video generation instruction
last_frame_goal	Expected final state
progress_goal	Optional expected intermediate process
foreground_rule	Optional foreground consistency constraint
background_rule	Optional background consistency constraint
implicit_rule	Optional implicit physical or logical constraint
has_progress_goal	Whether progress_goal is applicable
image_*	Image integrity and shape metadata

Evaluation

The bundled evaluation_code/ package evaluates five 1–5 metrics when applicable:

1. video quality;
2. progress consistency;
3. implicit-rule following;
4. progress-goal realization;
5. last-frame-goal realization.

Optional metrics are omitted when their annotation is null; they are not treated as zero.

Minimal local validation after cloning the dataset repository:

cd evaluation_code
python -m pip install -e .
PYTHONPATH=src python -m vwg_bench.cli validate-data --dataset-root ..

To evaluate generated videos:

cd evaluation_code
export GEMINI_API_KEY="..."
bash scripts/eval_vwg.sh .. /path/to/generated_videos outputs/results.jsonl 0,1,2

Generated videos are expected by default as {id}_seed{seed}.mp4, for example 0_seed0.mp4. See evaluation_code/README.md for details and external benchmark adapters.

Dataset statistics

• Samples: 380
• Images: 380
• Reasoning dimensions: 9
• Stable task groups: 38
• Samples per task group: 10
• Split: test only
• Evaluation score range: integer 1–5

License

VWG-Bench is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). See LICENSE.md.

Citation

Citation metadata should be added after the camera-ready bibliographic record is finalized.