SpatialVID: 1-10 Million Videos with Spatial Annotations for 3D Generation

SpatialVID provides between 1 million and 10 million video records paired with spatial annotations, developed by researchers at Nanjing University and the Chinese Academy of Science for CVPR 2026. The data supports multi-modal generative tasks by linking video sequences with 3D spatial metadata and English text descriptions.

Use Cases

• Training text-to-3d models using spatial annotations to define geometric structure
• Developing image-to-video generation pipelines that require spatial consistency
• Analyzing video-text alignment using the provided English language descriptions

Strengths

• Scale of 1 to 10 million records
• Expert-authored for CVPR 2026
• Includes spatial annotations for 3D-aware generation

Limitations

• Restricted to non-commercial use under CC BY-NC-SA 4.0 license
• Potential for high storage and compute requirements due to 1M+ video records

Provenance

Source

Nanjing University and Institute of Automation, Chinese Academy of Science (Arxiv 2509.09676)

Collection Method

Annotated

Freshness

Last updated March 2026; associated with CVPR 2026 publication.