MedVideoCap-55K

# Introduction

MedVideoCap-55K is a high-quality text-to-video dataset designed for research institutions to support medical video generation. It contains 55,803 medical videos covering areas like medical education, clinical practice, medical imaging, medical teaching and medical popular science videos. The videos are carefully selected for clear visuals, smooth motion, and good consistency. Unlike other datasets that use category labels, MedVideoCap-55K provides detailed and natural captions generated with the help of multimodal large language models (MLLMs), making it more useful for medical video generation model training. For more details, refer to our paper and github.

- **📄 Paper**: [MedGen: Unlocking Medical Video Generation by Scaling Granularly-annotated Medical Videos](https://huggingface.co/papers/2507.05675)
- **🗃️Github**: [https://github.com/FreedomIntelligence/MedGen](https://github.com/FreedomIntelligence/MedGen)

# Usage

We provide all the medical videos (`video_*.zip`) and their corresponding caption file (`MedVideoCap-55K.json`) in this repository. For easier management, the videos have been split and packaged into separate archives.
```
DATA_PATH
└─ MedVideoCap-55K.json
└─ videos_1.zip
└─ videos_2.zip
└─ videos_3.zip
└─ videos_4.zip
└─ videos_5.zip
└─ videos_6.zip
```

You can download this dataset, and unzip the all videos using the following code:
```linux
# Extract all zip files into the "videos" directory
mkdir -p videos && for f in videos_*.zip; do unzip -q "$f" -d videos/; done

# Move all .mp4 files to the "videos" directory and remove empty directories
find videos/ -type f -name "*.mp4" -exec mv -t videos/ {} + && find videos/ -type d -empty -delete
```

# Warning

The dataset is intended solely and strictly for research purposes and should not be used for nonresearch settings, especially in clinical practice.

# Citation

```bibtex
@misc{wang2025medgenunlockingmedicalvideo,
title={MedGen: Unlocking Medical Video Generation by Scaling Granularly-annotated Medical Videos},
author={Rongsheng Wang and Junying Chen and Ke Ji and Zhenyang Cai and Shunian Chen and Yunjin Yang and Benyou Wang},
year={2025},
eprint={2507.05675},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2507.05675},
}
```

# 简介
MedVideoCap-55K是一款高品质的文本到视频数据集，专为科研机构设计，用于支撑医学视频生成相关研究。该数据集包含55803条医学视频，涵盖医学教育、临床实践、医学影像、医学教学以及医学科普等多个领域。所有视频均经过严格筛选，具备画面清晰、运动流畅、一致性佳的特点。与其他仅使用类别标签的数据集不同，MedVideoCap-55K借助多模态大语言模型（Multimodal Large Language Models，MLLMs）生成了详细且自然的字幕标注，可更好地服务于医学视频生成模型的训练。如需了解更多细节，请参阅我们的论文及GitHub仓库。
- 📄 论文：[MedGen：通过规模化细粒度标注医学视频解锁医学视频生成](https://huggingface.co/papers/2507.05675)
- 🗃️ GitHub：[https://github.com/FreedomIntelligence/MedGen](https://github.com/FreedomIntelligence/MedGen)

# 使用方式
本仓库提供了全部医学视频文件（`video_*.zip`）及其对应的标注字幕文件（`MedVideoCap-55K.json`）。为便于管理，视频文件已拆分并打包为多个独立压缩包。

数据集路径
└─ MedVideoCap-55K.json
└─ videos_1.zip
└─ videos_2.zip
└─ videos_3.zip
└─ videos_4.zip
└─ videos_5.zip
└─ videos_6.zip

你可通过以下代码下载本数据集并解压所有视频文件：
linux
# 将所有压缩包解压至“videos”目录
mkdir -p videos && for f in videos_*.zip; do unzip -q "$f" -d videos/; done

# 将所有.mp4文件移动至“videos”目录并删除空文件夹
find videos/ -type f -name "*.mp4" -exec mv -t videos/ {} + && find videos/ -type d -empty -delete

# 警告
本数据集仅可用于科研用途，严禁用于非科研场景，尤其是临床实践活动。

# 引用格式
bibtex
@misc{wang2025medgenunlockingmedicalvideo,
title={MedGen: Unlocking Medical Video Generation by Scaling Granularly-annotated Medical Videos},
author={Rongsheng Wang and Junying Chen and Ke Ji and Zhenyang Cai and Shunian Chen and Yunjin Yang and Benyou Wang},
year={2025},
eprint={2507.05675},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2507.05675},
}