OmniDance

# 💃 OmniDance Dataset [](#) [](#) [](#) [](#) **OmniDance** is a large-scale dataset for **multimodal dance video generation**, built from Internet dance videos and curated for research on: - **Text-Image-to-Video (TI2V)** - **Music-Image-to-Video (MI2V)** - **Text-Music-Image-to-Video (MTI2V)** The dataset focuses on single-dancer dance videos with strong choreography content, stable visual appearance, and structured text supervision. --- ## ✨ Highlights - 🎬 Large-scale dance video collection from web sources - 📝 Choreography-informed text annotations - 🕺 Focus on single-dancer performance - 🎵 Suitable for music-conditioned dance generation - 🔎 Filtered for dance validity, reference clarity, and scene stability --- ## 📂 File Structure ```text Opensource_Data/ ├── text/ └── video/ ``` - `Opensource_Data/video/`: dance video clips - `Opensource_Data/text/`: paired text annotations for the videos --- ## 📦 Data Contents Each sample is organized around a dance video and its corresponding text description. ### `video/` Contains the dance video clips used for training or evaluation. ### `text/` Contains choreography-aware text annotations describing key properties of the dance video, including: - body dynamics - choreographic content - expressiveness - camera presentation - overall visual appearance These annotations are designed for dance-specific generation and provide more useful supervision than generic video captions. --- ## 🧠 Supported Tasks OmniDance is intended for research on: - **TI2V**: text + reference image → dance video - **MI2V**: music + reference image → dance video - **MTI2V**: text + music + reference image → dance video It can also support related tasks such as: - dance motion understanding - music-motion alignment - choreography-conditioned generation - identity-consistent human video synthesis --- ## 🛠️ Data Pipeline The dataset is constructed with a progressive filtering and annotation pipeline tailored for dance generation. Main stages include: - reference clarity verification - dance video verification - single-dancer filtering - scene stability filtering - choreography-aware text annotation This pipeline improves semantic precision and overall data quality for multimodal dance video generation. --- ## ⚠️ Limitations As a web-collected dataset, OmniDance may still contain some bias and noise: - female performers are more common than male performers - many samples come from Asian online dance communities - some videos may contain motion blur, compression artifacts, or local visual defects These characteristics mainly reflect source-platform distributions and practical quality-control trade-offs during large-scale curation. --- ## 🤝 Usage Notice This dataset is released for **research purposes only**. Please use it responsibly and pay attention to: - privacy and portrait-related concerns - demographic and regional bias - identity-sensitive misuse risks - compliance with local regulations and source-platform policies --- ## 📚 Citation If you use this dataset in your research, please cite: ```bibtex @article{omnidance2026, title={OmniDance: Multimodal Driven Dance Video Generation with Large-scale Internet Data}, author={Anonymous}, journal={ECCV}, year={2026} } ``` --- ## 📬 Contact For questions, suggestions, or issues, please open an issue in this repository.