Dataset · worldbenchmark
Intuitive physics benchmark for video and VLM models
worldbenchmark/IntuitivePhysics
A large English WorldBench benchmark for evaluating physical understanding in generative and vision models.
What's inside
A dataset ranging from 100,000 to 1,000,000 records featuring two components to test physical understanding. The first consists of 425 simulated scenes with 132 frames each, providing RGB, normal, depth, optical flow, and segmentation maps for Cosmos-tier models. The second adds English text questions to 181 video clips for the VQA task.
What you can do with it
- Evaluating physical realism in video-to-video and text-to-video models
- Testing vision-language models on physical law understanding from video
- Tasks
- Visual question answering, Text → video
- Languages
- English
- Size
- 100K–1M rows
- Downloads this month
- 235.6K
- Likes
- 12
How to load it
from datasets import load_dataset
ds = load_dataset("worldbenchmark/IntuitivePhysics")