OmniPhys
收藏资源简介:
OmniPhys是由浙江大学等机构联合构建的物理常识基准数据集,旨在系统评估文本到图像生成模型的物理合理性。该数据集包含1551个精心设计的样本,覆盖力学、光学和物体属性三大领域,基于物理知识图谱(PKG)和PhET模拟构建,确保每个样本均对应可视觉验证的物理知识节点。数据集通过隐式查询和双路径验证协议,诊断模型对基础物理定律的内化理解,而非表面模式匹配,主要应用于提升生成式AI的物理对齐能力,解决现有模型在流体静力学、光学折射等场景中的“物理幻觉”问题。
OmniPhys is a physical commonsense benchmark dataset jointly constructed by Zhejiang University and other institutions, aiming to systematically evaluate the physical plausibility of text-to-image generation models. This dataset includes 1551 meticulously designed samples covering three core domains: mechanics, optics, and object properties. It is built based on the Physical Knowledge Graph (PKG) and PhET simulations, ensuring that each sample corresponds to a visually verifiable physical knowledge node. Through implicit queries and a two-path validation protocol, the dataset diagnoses the model's internalized understanding of fundamental physical laws rather than superficial pattern matching. Its primary application is to enhance the physical alignment capability of generative AI and resolve the "physical hallucination" issues of existing models in scenarios such as hydrostatics and optical refraction.
数据集概述:OmniPhys
- 名称:OmniPhys
- 来源地址:https://github.com/zjukg/OmniPhys
- 描述:该数据集当前在GitHub上托管,项目名称为OmniPhys,具体内容尚未在README文件中详细说明。

- 1OmniPhys: Knowledge-Graph-Driven Benchmarking and Collective Optimization for Physical Commonsense in Text-to-Image Generation浙江大学; 曼彻斯特大学; 爱丁堡大学; 蚂蚁集团 · 2026年



