"LAION-Comp"

"LAION-Comp is a large-scale dataset designed to support compositional and controllable image generation through explicit structural annotations. It comprises 540,005 high-quality image\u2013scene graph pairs, built upon the LAION-Aesthetics V2 (6.5+) subset. Each sample includes a detailed scene graph that encodes multiple objects, their abstract attributes, and intricate inter-object relations using concrete verbs, moving beyond simple spatial descriptors. The annotation pipeline leverages GPT-4o with carefully designed prompts, followed by partial human verification, achieving high accuracy (98.8% for objects, 97.5% for attributes, 95.7% for relations). The dataset features an average of 6.39 objects per scene (excluding proper nouns), with diverse relations and attributes\u2014non-spatial relations dominate (77.48%), capturing rich functional and interaction-oriented semantics. Scene graph lengths vary widely, balancing expressiveness and learning efficiency. LAION-Comp is split into training (480k), validation (10k), and test (50k) sets, offering a foundational resource for advancing structured conditioning in image synthesis."

LAION-Comp 是一款旨在通过显式结构化标注支持组合式与可控式图像生成任务的大规模数据集。其包含540,005张高质量图像-场景图（scene graph）对，构建自LAION-Aesthetics V2（6.5+）子集。
每个样本均附带一份详细的场景图，该场景图通过具体动词编码了多类对象、其抽象属性以及复杂的对象间关系，突破了仅依赖简单空间描述符的局限。
该数据集的标注流水线依托GPT-4o并搭配精心设计的提示词，随后辅以部分人工核验，最终实现了较高的标注精度：对象标注准确率达98.8%，属性标注准确率达97.5%，关系标注准确率达95.7%。
该数据集平均每个场景（不含专有名词）包含6.39个对象，关系与属性类型丰富多样——非空间关系占比达77.48%，涵盖了丰富的功能性与交互导向语义。
场景图的长度跨度较大，兼顾了表达能力与学习效率。
LAION-Comp 被划分为训练集（480k）、验证集（10k）与测试集（50k），可为推进图像合成中的结构化条件生成提供基础支撑资源。