README.md

---
license: mit
language:
- en
size_categories:
- n<1K
tags:
- benchmark
- world-model
- video-generation
- programmable-world-models
arxiv: "2610.02205"
pretty_name: PROWBench — Episodes
configs:
- config_name: default
  default: true
  data_files:
  - split: full
    path: splits/full.parquet
  - split: verified_first_frame
    path: splits/verified_first_frame.parquet
  - split: unverified_first_frame
    path: splits/unverified_first_frame.parquet
  - split: long_video
    path: splits/long_video.parquet
  - split: multi_view
    path: splits/multi_view.parquet
  - split: gunfight
    path: splits/gunfight.parquet
---

# PROWBench Dataset

PROWBench evaluates whether video models follow program-specified motion and interactions. It contains **170 episodes, 200 camera views, and 600 proxy videos**, with prompts, recorded world states, and RGB first frames where available.

[Paper](https://arxiv.org/abs/2610.02205) · [Project](https://alaya-lab.github.io/PROWBench/) · [Gallery](https://alaya-lab.github.io/PROWBench/gallery.html) · [Code](https://github.com/AlayaLab/PROWBench) · [Evaluation guide](https://github.com/AlayaLab/PROWBench/blob/main/eval/USAGE.md)

## Download

```bash
pip install -U huggingface_hub
hf download AlayaLab/PROWBench --repo-type dataset --local-dir PROWBench
```

Load sample metadata and preview images in Python (`pip install -U datasets pillow`):

```python
from datasets import load_dataset

dataset = load_dataset("AlayaLab/PROWBench")
sample = dataset["full"][0]
```

The `*_file` fields contain paths relative to the dataset root. Use the download command above for the full videos, prompts and engine records.

## Data

| Split | Episodes | Duration | RGB first frames |
|---|---:|---:|---:|
| `verified_first_frame` | 40 | 5 s | 40 |
| `unverified_first_frame` | 90 | 5 s | 0 |
| `long_video` | 10 | 30 s | 0 |
| `multi_view` | 10 | 5 s | 0 |
| `gunfight` | 20 | 5 s | 20 |

Each view has three proxy representations: `white` (coarse 3D), `mixed` (semantic proxy), and `colored-obb` (3D boxes). Multi-view episodes have four synchronized cameras.

The `full` Viewer split contains **600 rows**, one per camera–representation pair; the other splits are subsets. Full prompts are available through `prompt_file`. `preview_kind` distinguishes RGB first-frame references from proxy previews.

A typical episode:

```text
episodes/<group>/<episode>/
├── manifest.json
├── scene/                              # Recorded world states
├── cameras/                            # Camera records
└── views/<camera>/
    ├── <representation>/video.mp4
    ├── <representation>/prompt.txt
    ├── prompt-context/first_frame.png   # When available
    └── visibility.json                 # Multi-view
```

Long-video prompts use JSON files or a `prompts/` directory of windowed prompts. Use `prompt_file` and the adjacent `sample.json` for exact input paths, including versioned episodes.

This release contains conditioning inputs and annotations; raw depth maps and generated model outputs are not included. [Release metadata](metadata/release.json) · [Sample metadata](metadata/samples.jsonl.gz).

## License

This dataset is released under the [MIT License](LICENSE).

## Citation

If you find PROWBench helpful for your research, please consider citing our paper:

```bibtex
@misc{huang2026prowbench,
  title={PROWBench: Do Video Models Render What the Program Specifies?},
  author={Zheng-Hui Huang and Guixu Lin and Yu-Ju Tsai and Jian-Kai Zhu and Fengbo Lan and Yu-Lun Liu and Yung-Yu Chuang and Kaipeng Zhang and Zhixiang Wang},
  year={2026},
  eprint={2610.02205},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.02205}
}
```