README.md

---
license: apache-2.0
task_categories:
- image-to-3d
---

# WorldSculpt Dataset

This dataset accompanies the paper [WorldSculpt: Generating Compositional Worlds from Grounded Videos](https://huggingface.co/papers/2609.05416). It provides the **UE-MeshyScene** benchmark of densely cluttered scenes with hundreds of objects, per-object annotations, and ground-truth meshes, as well as additional Marble demo data for converting 3DGS worlds into compositional mesh scenes.

- **Project page:** [https://alaya-lab.github.io/WorldSculpt/](https://alaya-lab.github.io/WorldSculpt/)
- **Code:** [https://github.com/AlayaLab/WorldSculpt](https://github.com/AlayaLab/WorldSculpt)

## Download

The dataset can be downloaded with the Hugging Face CLI as described in the [GitHub README](https://github.com/AlayaLab/WorldSculpt):

```bash
# UE-MeshyScene
hf download AlayaLab/WorldSculpt_data --include "UE-MeshyScene/*" --local-dir ./input --repo-type=dataset
cd ./input/UE-MeshyScene
tar xvf scene*.tar
cd ../..

# Marble DEMO data
hf download AlayaLab/WorldSculpt_data --include "Marble/*" --local-dir ./input --repo-type=dataset
cd ./input/Marble
tar xvf Marble.tar.gz
cd ../..
```

## License

Apache 2.0