README.md

metadata

language:
  - en
license: mit
pretty_name: VisPhyBench
task_categories:
  - video-text-to-text

VisPhyBench

Project Page | Paper | GitHub

To evaluate how well models reconstruct appearance and reproduce physically plausible motion, we introduce VisPhyBench, a unified evaluation protocol comprising 209 scenes derived from 108 physical templates that assesses physical understanding through the lens of code-driven resimulation in both 2D and 3D scenes, integrating metrics from different aspects. Each scene is also annotated with a coarse difficulty label (easy/medium/hard).

Dataset Details

• Created by: Jiarong Liang, Max Ku, Ka-Hei Hui, Ping Nie, Wenhu Chen
• Language(s) (NLP): English
• License: MIT
• Repository: https://github.com/TIGER-AI-Lab/VisPhyWorld

Uses

The dataset is used to evaluate how well models reconstruct appearance and reproduce physically plausible motion.

Dataset Structure

The difficulty of the two sets of VisPhyBench splits, sub (209) and test (49), are as follows: sub is 114/67/28 (54.5%/32.1%/13.4%), and test is 29/17/3 (59.2%/34.7%/6.1%) (Easy/Medium/Hard).

What each sample contains

VisPhyBench is provided as two splits:

• sub: a larger split intended for evaluation and analysis.
• test: a smaller split subsampled from sub for quick sanity checks.

For each sample, we provide:

1. A short video of a synthetic physical scene.
2. A detection JSON (per sample) that describes the scene in the first frame.
3. A difficulty label (easy/medium/hard) derived from the mean of eight annotators’ 1–5 ratings.

Detection JSON format

Each detection JSON includes:

• image_size: the image width/height.
• coordinate_system: conventions for coordinates (e.g., origin and axis directions).
• objects: a list of detected objects. Each object includes:
  • id: unique identifier.
  • category: coarse geometry category.
  • color_rgb: RGB color triplet.
  • position: object position (e.g., center coordinates).
  • bbox: bounding box coordinates and size.
  • size: coarse size fields (e.g., radius/length/thickness).

These fields specify object locations and attributes precisely, which helps an LLM initialize objects correctly when generating executable simulation code.

BibTeX

@misc{liang2026visphyworld,
  title={VisPhyWorld: Probing Physical Reasoning via Code-Driven Video Reconstruction}, 
  author={Jiarong Liang and Max Ku and Ka-Hei Hui and Ping Nie and Wenhu Chen},
  year={2026},
  eprint={2602.13294},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2602.13294}, 
}