Dataset · ccmoony
Physically grounded video editing benchmark
ccmoony/PhyCoVE
508 tasks for altering physical object properties in video with reference clips and English/Chinese instructions.
What's inside
The physics-simulation-based benchmark includes 90 scenes, 90 source videos, and 508 editing tasks (totaling fewer than 1,000 entries). Each task consists of a source clip, an English or Chinese text instruction, a reference edited video, and simulator markup. Scenarios cover rigid body behavior, elastic materials, fluids, and their combinations.
What you can do with it
- Evaluating video generation and editing models (video-to-video) on physics knowledge
- Verifying understanding of text instructions for changing dynamic object properties
- Tasks
- Video → video
- Languages
- English, Chinese
- Size
- up to 1K rows
- Downloads this month
- 615
- Likes
- 2
How to load it
from datasets import load_dataset
ds = load_dataset("ccmoony/PhyCoVE")