OpenCausaLab/LINA-PAP

该数据集名为物理对齐探针（PAP），是LINA论文中引入的核心诊断数据集。它基于因果场景图（CSG）构建，旨在诊断扩散模型中的物理推理和分布外（OOD）生成能力。数据集包含三个核心部分：1）结构化提示库，包含287个结构化提示，分为光学、密度和OOD三个诊断子集；2）大规模图像语料库，包含28,700张图像，每个提示生成50张图像；3）诊断分割掩码，用于关键元素的细粒度分割。数据集还提供了评估指标，用于量化纹理对齐和物理对齐的成功率。

The dataset, named Physical Alignment Probe (PAP), is the core diagnostic dataset introduced in the LINA paper. It is a multi-modal corpus based on Causal Scene Graphs (CSG) designed to diagnose physical reasoning and out-of-distribution (OOD) generation capabilities in diffusion models. The dataset comprises three core components: 1) A Structured Prompt Library with 287 structured prompts divided into three diagnostic subsets (Optics, Density, and OOD); 2) A Large-Scale Image Corpus of 28,700 images, with 50 images generated per prompt; and 3) Diagnostic Segmentation Masks for fine-grained segmentation of critical elements. The dataset also includes evaluation metrics to quantify the success rates of Texture Alignment and Physical Alignment.