✈️ Introduction

The challenge of spatiotemporal consistency has long existed in the field of video generation. We have released the open-source dataset DropletVideo-10M —the world's largest video generation dataset with spatiotemporal consistency. It contains 10 million high-quality video clips (totaling 20,400 hours) and redefines the training benchmarks for video generation technology from a data perspective.

Traditional datasets (such as OpenVid-1M) primarily focus on object motion in static camera settings. In contrast, DropletVideo-10M is the first to integrate both object motion and camera motion, introducing dual dynamics. Through an innovative four-stage construction process, it extracts spatiotemporally coherent segments from long videos, ensuring that each sample contains both subject actions and camera movements such as zooming and panning. This provides multi-dimensional data support for models to better understand real-world motion patterns. You can check our paper for the data processing details.

The dataset employs a dual filtering mechanism based on aesthetic scores and image quality (see Figure 4). As a result, 95% of the videos meet professional visual aesthetic standards (aesthetic score > 3.5), and 78% of the clips achieve cinematic-grade quality (quality score > 4.0). This combination of artistic value and technical parameters effectively mitigates issues such as frame jitter and distortion caused by low-quality data.

✈️ Download

We released two versions of dataset. DropletVideo-10M and DropletVideo-1M. DropletVideo-10M needs the gated grant requirements but DropletVideo-1M is no needed.

✈️ Model

The DropletVideo-5B video generation model, trained on the DropletVideo-10M dataset, is the first known model in the industry to support motion intensity control. Its video generation performance has reached the level of leading open-source models.

Model weights: DropletVideo-5B

Project : DropletX

Paper : Arxiv

Citation

🌟 If you find our work helpful, please leave us a star and cite our paper.

@article{zhang2025dropletvideo,
        title={DropletVideo: A Dataset and Approach to Explore Integral Spatio-Temporal Consistent Video Generation},
        author={Zhang, Runze and Du, Guoguang and Li, Xiaochuan and Jia, Qi and Jin, Liang and Liu, Lu and Wang, Jingjing and Xu, Cong and Guo, Zhenhua and Zhao, Yaqian and Gong, Xiaoli and Li, Rengang and Fan, Baoyu},
        journal={arXiv preprint arXiv:2503.06053},
        year={2025}
      }

Datasets:

DropletX
/

DropletVideo-1M

You need to agree to share your contact information to access this dataset