LAYOUTBENCH

LAYOUTBENCH是由北卡罗来纳大学教堂山分校和微软研究院合作创建的数据集，专注于评估布局引导的图像生成模型。该数据集包含8000张图像，每张图像都针对四个关键的空间控制技能（数量、位置、大小、形状）进行了设计。数据集的创建过程涉及从CLEVR模拟器中抽样场景，并使用Blender渲染图像以获取边界框布局。LAYOUTBENCH的应用领域主要集中在解决图像生成模型在面对任意、未见布局时的表现问题，特别是在超出分布（OOD）布局上的泛化能力。

LAYOUTBENCH is a dataset jointly developed by the University of North Carolina at Chapel Hill and Microsoft Research, focusing on evaluating layout-guided image generation models. This dataset consists of 8000 images, each designed to target four core spatial control capabilities: count, position, scale, and shape. The dataset was created by sampling scenes from the CLEVR simulator and rendering images via Blender to obtain bounding box layouts. The primary application scenario of LAYOUTBENCH is to evaluate the performance of image generation models when facing arbitrary, unseen layouts, particularly their generalization ability on out-of-distribution (OOD) layouts.