Skip to content

Dataset · NU-World-Model-Embodied-AI

Physics benchmark in generated video

NU-World-Model-Embodied-AI/phyground

250 prompts with initial frames and labels for 13 laws of physics to test the physical validity of video generation models.

What's inside

PhyGround is an English-language criterion benchmark (fewer than 1,000 records) for diagnosing physical errors in AI-generated videos. The set includes 250 prompts covering 13 observable laws of physics from solid mechanics, fluid dynamics, and optics. Each example is supplemented with a first-frame image, 10 generation configurations, and applicable law labels.

What you can do with it

  • Evaluation and diagnosis of physical hallucinations in text-to-video and image-to-video models
  • Comparing generative world models for adherence to mechanical and optical laws
Tasks
Text → video, Image → video
Languages
English
Size
up to 1K rows
Downloads this month
4K
Likes
48

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("NU-World-Model-Embodied-AI/phyground")