PhysEditWorld

A Large-Scale Dataset Toward Physics-Editable World Models

Bin Hu1,* Yanwen Ma2,* Jiehui Huang3,‡ Ziliang Zhang4 Haoning Wu5 Ruicheng Zhang1 Yaokun Li6 Zijun Wang6 Yuechen Zhang7 Chun-Mei Tseng1 Hanhui Li6 Shengju Qian8 Jun Zhou1 Kaipeng Zhang9 Xiaodan Liang6 Jiaya Jia3,† Xiu Li1,†

1Tsinghua Shenzhen International Graduate School, Tsinghua University 2Beihang University 3The Hong Kong University of Science and Technology 4Independent Researcher 5Shanghai Jiao Tong University 6Sun Yat-sen University 7The Chinese University of Hong Kong 8Alaya Studio 9Tencent

*Equal contribution Project lead Corresponding authors

Physics Editing Matched replay 8 synchronized views 60M+ frames
12
UE5 scenes
100h+
interactions
60M+
rendered frames
8
camera views
7
gravity levels

Overview

Matched replay groups keep scene, action, controller, and cameras fixed while gravity changes.

Each sample includes synchronized videos and engine-side metadata.

Part of Dataset

Representative scenes and gravity settings.

MarsRover

Low, normal, and high gravity.

G_0.25sample 00000002
G_1.0sample 00000005
G_10sample 00000006

CyberCity

Urban motion under edited gravity.

G_0.25sample 00000022
G_1.0sample 00000025
G_10sample 00000026

MoonWalk

Multi-view lunar-style motion.

G_0.25sample 00000042
G_1.0sample 00000045
G_10sample 00000046

Evaluation

Held-out gravity prediction and response analysis.

95.29% VLM class accuracy after PhysEditWorld SFT
90.59% Predictions within 10% of true gravity
6.22% Median absolute percentage error
100.0% Gravity alignment in the VGGT response case

Gravity-aware VLM prediction

Qwen3-VL-8B on 170 held-out rollouts. +SFT denotes PhysEditWorld LoRA tuning.

Model Class Acc. (%) MAE Median APE (%) Within 10%
Qwen3-VL-8B 24.71 5.4573 95.00 9.48
Qwen3-VL-8B + SFT 95.29 0.8305 6.22 90.59

Gravity response analysis

Representative held-out matched replay group. Acceleration proxy uses normalized VGGT units.

Model Accel. @0.25x Accel. @0.75x Accel. @10.0x Mean R2 Gravity Align.
Wan2.2-TI2V-5B 0.0002 0.0046 0.0009 0.066 33.3%
Wan2.2-TI2V-5B + SFT 0.1766 0.3983 0.6214 0.570 100.0%

Data

<dataset_root>/
  Video/<sample_id>/<camera_name>.mp4
  Meta/<sample_id>_meta_frame.csv
  Meta/<sample_id>_meta_time.csv
  PhysicalConfig/<physical_config_name>.json
  Aux/<aux_type>/<sample_id>/...
  Details.json

Citation

@misc{hu2026physeditworldlargescaledatasetphysicseditable,
      title={PhysEditWorld: A Large-Scale Dataset Toward Physics-Editable World Models},
      author={Bin Hu and Yanwen Ma and Jiehui Huang and Ziliang Zhang and Haoning Wu and Ruicheng Zhang and Yaokun Li and Zijun Wang and Yuechen Zhang and Chun-Mei Tseng and Hanhui Li and Shengju Qian and Jun Zhou and Kaipeng Zhang and Xiaodan Liang and Jiaya Jia and Xiu Li},
      year={2026},
      eprint={2606.26694},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.26694},
}