LAION-SG is a large-scale, high-quality image understanding dataset built by Zhejiang University, Jiangnan University, Peking University, Alibaba Group, and Ant Group in 2024.LAION-SG: An Enhanced Large-Scale Dataset for Training Complex Image-Text Models with Structural Annotations". LAION-SG contains 540,005 scene graph-image pairs with object, attribute and relationship annotations, which are divided into training, validation and test sets. The images in the dataset are from the LAION-Aesthetics V2 (6.5+) dataset, and the annotation process uses GPT-4o for automatic annotation.

Compared to the original LAION-Aesthetics dataset, LAION-SG has improved both the average annotation length and accuracy. Each sample in this dataset contains an average of 6.39 objects, and the object information has increased by 20%. If abstract proper nouns are excluded, this advantage increases to 216%.

The LAION-SG dataset is suitable for a variety of cross-modal research fields of images and text, including image description generation, visual question answering systems, and image retrieval tasks, all of which rely on a deep understanding and semantic parsing of image content.