README.md

---
license: apache-2.0
datasets:
- Seed3D/Articulation-XL2.0
---
<div align="center">

<h1>MagicArticulate: Make Your 3D Models Articulation-Ready</h1>

<p>
  <a href="https://chaoyuesong.github.io"><strong>Chaoyue Song</strong></a><sup>1,2</sup>,
  <a href="https://jfzhang95.github.io/"><strong>Jianfeng Zhang</strong></a><sup>2*</sup>,
  <a href="https://lixiulive.com/"><strong>Xiu Li</strong></a><sup>2</sup>,
  <a href="https://scholar.google.com/citations?user=afDvaa8AAAAJ&hl"><strong>Fan Yang</strong></a><sup>1,2</sup>,
  <a href="https://buaacyw.github.io/"><strong>Yiwen Chen</strong></a><sup>1</sup>,
  <a href="https://zcxu-eric.github.io/"><strong>Zhongcong Xu</strong></a><sup>2</sup>,
  <a href="https://liewjunhao.github.io/"><strong>Jun Hao Liew</strong></a><sup>2</sup>,
  <strong>Xiaoyang Guo</strong><sup>2</sup>,
  <a href="https://sites.google.com/site/fayaoliu"><strong>Fayao Liu</strong></a><sup>3</sup>,
  <a href="https://scholar.google.com.sg/citations?user=Q8iay0gAAAAJ"><strong>Jiashi Feng</strong></a><sup>2</sup>,
  <a href="https://guosheng.github.io/"><strong>Guosheng Lin</strong></a><sup>1*</sup>
  <br>
  *Corresponding authors
  <br>
    <sup>1 </sup>Nanyang Technological University
  <sup>2 </sup>Bytedance Seed
  <sup>3 </sup>A*STAR
</p>

<h3>CVPR 2025</h3>

<p>
  <a href="https://chaoyuesong.github.io/MagicArticulate/"><strong>Project</strong></a> |
  <a href="https://chaoyuesong.github.io/files/MagicArticulate_paper.pdf"><strong>Paper</strong></a> |
  <a href="https://github.com/Seed3D/MagicArticulate"><strong>Code</strong></a> |
  <a href="https://www.youtube.com/watch?v=eJP_VR4cVnk"><strong>Video</strong></a>
</p>


</div>

<br />

### Update
- 2025.8.07: We updated the [skeleton generation weights](https://huggingface.co/Seed3D/MagicArticulate) to fix a training data loader normalization bug causing mesh-skeleton misalignment.
- 2025.4.16: Release weights for skeleton generation trained on Articulation-XL2.0.

### Overview
This repository includes weights of MagicArticulate trained on Articulation-XL2.0. These are the numbers that you should be able to reproduce using the released weights and the current version of the codebase:
<table>
  <thead>
    <tr>
      <th rowspan="2">Test set</th>
      <th colspan="3">Articulation-XL2.0-test</th>
      <th colspan="3">ModelResource-test</th>
    </tr>
    <tr>
      <th>CD-J2J</th>
      <th>CD-J2B</th>
      <th>CD-B2B</th>
      <th>CD-J2J</th>
      <th>CD-J2B</th>
      <th>CD-B2B</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Paper (train on 1.0, spatial)</td>
      <td>-</td>
      <td>-</td>
      <td>-</td>
      <td>4.103</td>
      <td>3.101</td>
      <td>2.672</td>
    </tr>
    <tr>
      <td>Paper (train on 1.0, hier)</td>
      <td>-</td>
      <td>-</td>
      <td>-</td>
      <td>4.451</td>
      <td>3.454</td>
      <td>2.998</td>
    </tr>
    <tr>
      <td>train on Arti-XL2.0 (spatial)</td>
	  <td><b>3.034</b></td>
      <td><b>2.263</b></td>
      <td><b>1.910</b></td>
      <td><b>3.999</b></td>
      <td><b>3.027</b></td>
      <td><b>2.591</b></td>
    </tr>
    <tr>
      <td>train on Arti-XL2.0 (hier)</td>
      <td>3.264</td>
      <td>2.503</td>
      <td>2.123</td>
      <td>4.114</td>
      <td>3.137</td>
      <td>2.693</td>
    </tr>
  </tbody>
</table>
The performance comparison between models trained on Articulation-XL1.0 versus 2.0 demonstrates the importance of dataset scaling with high quality. If you wish to compare your methods with MagicArticulate trained on Articulation-XL2.0, you may reference these results as a baseline for comparison.


### Citation

```
@article{song2025magicarticulate,
      title={MagicArticulate: Make Your 3D Models Articulation-Ready}, 
      author={Chaoyue Song and Jianfeng Zhang and Xiu Li and Fan Yang and Yiwen Chen and Zhongcong Xu and Jun Hao Liew and Xiaoyang Guo and Fayao Liu and Jiashi Feng and Guosheng Lin},
      journal={arXiv preprint arXiv:2502.12135},
      year={2025},
}
```