SPARK: Sim-ready Part-level Articulated Reconstruction with VLM Knowledge

CVPR 2026 · Oral

Website  ·  arXiv  ·  Code

Model checkpoints for SPARK, a framework for articulated 3D reconstruction from a single image. See the paper for the method and results, and the code for the full pipeline.

What's included

Folder	Component
PartCrafter/	3D part-generation checkpoint loaded by inference
TripoSG/	Base-model weights used by the training configuration
RMBG-1.4/	Optional image background removal

The full checkpoint bundle is approximately 15.6 GB.

Get started

Follow the installation instructions first. From the SPARK code directory, download the checkpoints with:

python -m src.utils.download_weights

Weights are saved under pretrained_weights/. Follow the Quick Start for reconstruction, requirements, and API configuration.

License

The SPARK code is MIT-licensed. Bundled third-party checkpoints retain their respective terms; SPARK's code license does not replace them. See the component cards for PartCrafter, TripoSG, and RMBG-1.4. BRIA describes RMBG-1.4 as available for non-commercial use, with a separate commercial agreement required.

Citation

@inproceedings{he2026spark,
  title={Spark: Sim-ready part-level articulated reconstruction with vlm knowledge},
  author={He, Yumeng and Jiang, Ying and Lu, Jiayin and Yang, Yin and Jiang, Chenfanfu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={7233--7243},
  year={2026}
}