WorldScore: A Unified Evaluation Benchmark for World Generation

Stanford's unified world-generation benchmark that lets 3D and 4D scene-generation models and video models be compared on the same scale.

WorldScore is a world-generation evaluation benchmark released in April 2025 by Fei-Fei Li and Jiajun Wu's group at Stanford, included in ICCV 2025. World generation means producing, from an image or text, a scene that can be explored along a camera path; up to that point, 3D scene generation, 4D scene generation, and video generation models were each evaluated separately, with no way to compare across them directly. WorldScore unifies the task into a chain of “generate the next scene” steps, with each step's layout specified by a camera trajectory, over a dataset of 3,000 examples. Metrics fall into three groups: controllability (camera, object, and content alignment), quality (3D consistency, photometric and style consistency, subjective quality), and dynamics (motion accuracy, magnitude, and smoothness), combined into two overall scores, WorldScore-Static and WorldScore-Dynamic.

ExampleThe paper used the same set of examples to evaluate 19 open- and closed-source models across four categories — 3D scene generation, 4D scene generation, image-to-video, and text-to-video — and maintains a public leaderboard.

Also called

WorldScore

Related

World Model · Video Generation Model · 4D World Model · Marble (World Labs) · WorldArena: A Unified Benchmark for Evaluating Perception and Functional Utility of Embodied World Models · VBench: Comprehensive Benchmark Suite for Video Generative Models

Sources

WorldScore (arXiv 2504.00983, ICCV 2025)
WorldScore 项目主页 (Chinese)

As of

2025-11