PhyBench
收藏资源简介:
PhyBench是由上海人工智能实验室创建的综合性T2I评估数据集,包含700个提示,涵盖力学、光学、热力学和材料属性四大类物理知识,共涉及31种不同的物理场景。该数据集旨在评估文本到图像模型在遵循物理常识方面的能力,特别是在世界模拟和日常任务中的应用。PhyBench通过严格的提示收集流程,确保评估集中在图像背后的物理知识,而非图像与提示的简单对齐。数据集的应用领域包括提升T2I模型在物理常识推理方面的性能,推动更接近真实世界模拟的技术发展。
PhyBench is a comprehensive text-to-image (T2I) evaluation dataset developed by the Shanghai AI Laboratory. It contains 700 prompts covering four categories of physical knowledge: mechanics, optics, thermodynamics, and material properties, involving a total of 31 distinct physical scenarios. This dataset aims to evaluate the capability of text-to-image models to comply with physical common sense, especially for applications in world simulation and daily tasks. PhyBench employs a rigorous prompt collection process to ensure that the evaluation focuses on the physical knowledge underlying the generated images, rather than the simple alignment between images and prompts. The application fields of this dataset include enhancing the performance of T2I models in physical common sense reasoning, and promoting the development of technologies closer to real-world simulation.

- 1PhyBench: A Physical Commonsense Benchmark for Evaluating Text-to-Image Models上海人工智能实验室 · 2024年



