STANCE_dataset

# STANCE Scenes (Composite & Synthetic)

**Snapshot:** 95k scenes total — **75,000 composite** + **20,000 synthetic/simple**.
Built with **Kubric** for short clips of rigid-body interactions.

- **Synthetic/simple:** minimal scenes with one or more rigid objects; randomized shape (ball / a few primitives), mass, initial linear velocity, and pose. Lighting uses three area lights plus a sun with randomized intensity/temperature.
- **Composite:** replaces primitives with **GSO** assets; backgrounds sampled from ~5,000 environment maps; randomized object selection/placement/pose to induce diverse contacts and occlusions.
Camera intrinsics/extrinsics and renderer settings are consistent within each scene; material, friction, restitution, and object counts are sampled within bounded ranges.

## Files in this release
- `composite_scene.tar.part-00`
- `composite_scene.tar.part-01`
- `composite_scene.tar.part-02`
- `composite_scene.tar.part-03`
- `synthetic_scene.tar`

> `composite_scene.tar` is split into 4 parts; `synthetic_scene.tar` is a single archive.

## Reassemble & Extract
Linux / macOS:
```bash
cat composite_scene.tar.part-0{0..3} > composite_scene.tar
tar -xf composite_scene.tar
tar -xf synthetic_scene.tar
````

(Optional) verify:

```bash
sha256sum composite_scene.tar synthetic_scene.tar
```

## Project & Code

For detailed information about our work, please visit our **project page**:
[https://envision-research.github.io/STANCE/](https://envision-research.github.io/STANCE/)

If you find this work useful, please consider **citing our paper** and leaving a **⭐ on GitHub**:
[https://github.com/EnVision-Research/STANCE](https://github.com/EnVision-Research/STANCE)

## Citation

```bibtex
@article{chen2025stance,
title = {STANCE: Motion Coherent Video Generation Via Sparse-to-Dense Anchored Encoding},
author = {Zhifei Chen and Tianshuo Xu and Leyi Wu and Luozhou Wang and Dongyu Yan and Zihan You and Wenting Luo and Guo Zhang and Yingcong Chen},
journal = {arXiv preprint arXiv:2510.14588},
year = {2025}
}
```

## License

CC BY 4.0 (you may switch to CC BY-NC 4.0 if preferred).

```
::contentReference[oaicite:0]{index=0}
```

# STANCE 场景集（复合场景与合成场景）

**数据集概览**：总场景数共计95,000个——**75,000个复合场景** + **20,000个合成/简易场景**。本数据集基于**Kubric**构建，用于生成刚体交互的短片段。

- **合成/简易场景**：仅包含单个或多个刚体对象的极简场景；对象形状（球体/若干基础图元）、质量、初始线速度及位姿均为随机生成。光照配置采用三盏区域光加一盏太阳光，强度与色温均随机设置。
- **复合场景**：将基础图元替换为**GSO（Google Scanned Objects，谷歌扫描对象）**资产；背景采样自约5000张环境贴图；通过随机化对象选择、放置方式与位姿，以生成多样化的接触与遮挡效果。每个场景内的相机内参、外参及渲染器设置保持一致；材质、摩擦系数、恢复系数及对象数量均在限定范围内随机采样。

**本版本发布文件清单**：
- `composite_scene.tar.part-00`
- `composite_scene.tar.part-01`
- `composite_scene.tar.part-02`
- `composite_scene.tar.part-03`
- `synthetic_scene.tar`

> `composite_scene.tar` 被拆分为4个分卷；`synthetic_scene.tar` 为单归档文件。

**重组与解压**：
Linux / macOS 系统下可执行以下命令：
bash
cat composite_scene.tar.part-0{0..3} > composite_scene.tar
tar -xf composite_scene.tar
tar -xf synthetic_scene.tar

（可选）校验文件完整性：
bash
sha256sum composite_scene.tar synthetic_scene.tar

**项目与代码**：
如需了解本研究的详细信息，请访问我们的**项目页面**：[https://envision-research.github.io/STANCE/](https://envision-research.github.io/STANCE/)
若您认为本工作对您有所帮助，请引用我们的论文，并在 GitHub 上点亮 **⭐**：[https://github.com/EnVision-Research/STANCE](https://github.com/EnVision-Research/STANCE)

**引用格式**：
bibtex
@article{chen2025stance,
title = {STANCE: Motion Coherent Video Generation Via Sparse-to-Dense Anchored Encoding},
author = {Zhifei Chen and Tianshuo Xu and Leyi Wu and Luozhou Wang and Dongyu Yan and Zihan You and Wenting Luo and Guo Zhang and Yingcong Chen},
journal = {arXiv preprint arXiv:2510.14588},
year = {2025}
}

**许可证**：
采用 CC BY 4.0 协议（您也可选择改用 CC BY-NC 4.0 协议）。