MechVerse is a large-scale video dataset of mechanical assembly animations for fine-tuning and evaluating generative video models on mechanical motion understanding.

Project Page

Visit the project page for interactive demos, dataset visualizations, and model comparisons:

mechverse.pages.dev

Overview

21,156 video clips from 1,357 assemblies across 141 categories
Each clip paired with a structured natural language prompt
Two sources: PartNet-Mobility (904 assemblies) + curated CAD assemblies (453)

Complexity Tiers

Tier	Clips	Description
Easy	15,720	Independent part motion (1–2 parts)
Medium	2,412	Kinematically coupled motion (3–8 parts)
Hard	3,024	Dense coupling, 10–50 parts

Clip Variations

Each assembly rendered across 3 speeds (Slow/Mid/Fast), 2–3 camera viewpoints, and 2 motion directions (Forward/Reversed). All clips are 32 frames at 16 FPS (2 seconds).

Data Splits

File	Clips
`train.json`	18,672
`test.json`	1,004
`full_dataset.json`	21,156

Citation

If you use MechVerse in your research, please cite:

bibtex
@article{jain2026mechverse,
  title={MechVerse: Evaluating Physical Motion Consistency in Video Generation Models},
  author={Jain, Rahul and Patel, Mayank and Unmesh, Asim and Ramani, Karthik},
  journal={arXiv preprint arXiv:2605.14843},
  year={2026}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You are free to use, share, and adapt this dataset for any purpose, including commercial use, provided that you:

Cite our paper when using this dataset in any publication or product
Link back to this dataset: https://huggingface.co/datasets/Mayank-p3108/MechVerse
Indicate any changes made to the original dataset

Downloads last month: 28

Total file size:

6.27 GB

Paper for Mayank-p3108/MechVerse

MechVerse: Evaluating Physical Motion Consistency in Video Generation Models

Paper • 2605.14843 • Published May 14 • 1