T2I-CoReBench is a comprehensive evaluation benchmark for text-driven image generation models proposed by the University of Science and Technology of China, Kuaishou Technology Kling Team, and the University of Hong Kong in 2025. The relevant paper results are "Easier Painting Than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play?", which aims to simultaneously measure the combination ability and reasoning ability of image generation models.

The dataset contains 1,080 highly challenging prompts and is equipped with approximately 13,500 inspection items covering 12 dimensions, which are used to evaluate whether each expected element in the generated image is correctly presented.

Data composition

This dataset designs prompts and evaluation systems from two dimensions:

• Composition dimension: Construct various composition structures around three types of scene graph elements: instance, attribute, and relation.
• Reasoning dimension: Based on three types of reasoning: deductive, inductive, and abductive.

To facilitate fine-grained evaluation, each prompt is accompanied by a yes/no checklist that notes whether each element implicitly or explicitly required by the prompt is correctly presented.