Dataset · vLAR
Multimodal dataset for studying visual physics
vLAR/PhysInOne
Millions of videos and 3D assets for training models to understand physical laws and generate video.
What's inside
Large-scale English dataset (1 to 10 million records) from researchers at Hong Kong Polytechnic University, Syai Singapore, and Meta (presented at CVPR 2026). Includes rendered videos, 3D assets, and detailed annotations for depth estimation, segmentation, and object detection. Designed for comprehensive study and evaluation of physical reasoning in visual tasks.
What you can do with it
- Training physically plausible text-to-video and image-to-video generation models
- Training depth estimation, segmentation, and detection models in 3D scenes
- Evaluating physics understanding in video models using the built-in leaderboard
- Tasks
- Text → video, Image → video, Video → video, Depth estimation, Image segmentation, Object detection
- Languages
- English
- Size
- 1M–10M rows
- Downloads this month
- 19.7K
- Likes
- 26
How to load it
from datasets import load_dataset
ds = load_dataset("vLAR/PhysInOne")