metadata
license: apache-2.0
OSCBench Dataset
Dataset Description
OSCBench is a benchmark dataset designed to evaluate object state change (OSC) reasoning in text-to-video (T2V) generation models.
OSCBench organizes prompts into three scenario types:
- Regular scenarios: common action–object combinations frequently seen in training data.
- Novel scenarios: uncommon but physically plausible action–object pairs that test generalization.
- Compositional scenarios: prompts that combine multiple actions or conditions to test compositional reasoning.
Dataset Statistics
The OSCBench dataset contains 1,120 prompts organized into three scenario categories:
| Scenario Type | Number of Scenarios | Prompts per Scenario | Total Prompts |
|---|---|---|---|
| Regular | 108 | 8 | 864 |
| Novel | 20 | 8 | 160 |
| Compositional | 12 | 8 | 96 |
| Total | 140 | — | 1,120 |
Dataset Sources
- Dataset: https://huggingface.co/datasets/XianjingHan/OSCBench_Dataset
- Paper: https://arxiv.org/abs/2603.11698
- Project Page: https://hanxjing.github.io/OSCBench
Acknowledgements and Citation
If you find this dataset helpful, please consider citing the original work:
@article{han2026oscbench,
title={OSCBench: Benchmarking Object State Change in Text-to-Video Generation},
author={Han, Xianjing and Zhu, Bin and Hu, Shiqi and Li, Franklin Mingzhe and Carrington, Patrick and Zimmermann, Roger and Chen, Jingjing},
journal={arXiv preprint arXiv:2603.11698},
year={2026}
}