Skip to content

Dataset · OpenEnvisionLab

WorldAtlas benchmark for evaluating generative world models

OpenEnvisionLab/WorldAtlas

A large collection of labeled images and reference videos in English to check consistency and physical plausibility of scene generators.

What's inside

The benchmark includes 5,000 labeled images, 960 starting frames for closed-loop memory verification, and 4,961 reference videos, with a total size estimated at 10-100k entries. English data includes images, text, and videos. Created by the OpenEnvision project for comprehensive evaluation of generative world models.

What you can do with it

  • Evaluating physical plausibility and controllability of video and 3D scene generators
  • Checking long-term visual consistency (memory) in world models
  • Comparing multimodal scene generation systems by a unified standard
Tasks
Other
Languages
English
Size
10K–100K rows
Downloads this month
13K
Likes
8
Access
on request on Hugging Face

How to load it

from datasets import load_dataset

ds = load_dataset("OpenEnvisionLab/WorldAtlas")