WorldGenBench: A World-Knowledge-Integrated Benchmark for Reasoning-Driven Text-to-Image Generation
Daoan Zhang, Che Jiang, Ruoshi Xu, Biaoxiang Chen, Zijian Jin, Yutian Lu, Jianguo Zhang, Liang Yong, Jiebo Luo, Shengda Luo
2025-05-05
Summary
This paper talks about WorldGenBench, a new way to test how well AI models can turn text into images by checking if they really understand facts about the world and can reason about them.
What's the problem?
Many text-to-image AIs can make pictures from words, but they often miss important details or make mistakes because they don't fully understand real-world knowledge or how things connect logically.
What's the solution?
The researchers created a special test, called WorldGenBench, that scores these AIs on how well they use world knowledge and reasoning when making images, helping to spot where the models are lacking.
Why it matters?
This matters because it helps developers improve AI so it can create more realistic and accurate images from text, which is important for education, creativity, and making sure technology understands the world like people do.
Abstract
WorldGenBench evaluates text-to-image models using the Knowledge Checklist Score to identify gaps in world knowledge grounding and implicit reasoning.