X-AIGD
收藏资源简介:
X-AIGD是一个专为可解释性AI生成图像检测设计的细粒度基准数据集。该数据集提供了AI生成图像中感知伪影的像素级人工标注,涵盖低级失真、高级语义和认知级反事实伪影,旨在推动鲁棒且可解释的AI生成图像检测方法的发展。数据集包含超过18,000个伪影实例,覆盖3,000多个标注样本,并包含大规模未标注数据。数据字段包括图像、生成器名称、唯一标识符、伪影标注(类别和多边形坐标)、生成提示文本、生成参数(如推理步数、引导尺度等)以及图像格式和压缩细节。数据集分为标注训练集、标注测试集、未标注训练集和未标注测试集,采用CC BY 4.0许可证发布。
X-AIGD is a fine-grained benchmark dataset specifically designed for explainable AI-generated image detection. This dataset provides pixel-level manual annotations of perceptual artifacts in AI-generated images, covering low-level distortions, high-level semantic and cognitive-level counterfactual artifacts, which aims to advance the development of robust and explainable AI-generated image detection methods. The dataset contains over 18,000 artifact instances, covering more than 3,000 annotated samples, and includes large-scale unannotated data. The data fields include images, generator names, unique identifiers, artifact annotations (category and polygon coordinates), generation prompt texts, generation parameters (such as inference steps, guidance scale, etc.), as well as image format and compression details. The dataset is divided into annotated training set, annotated test set, unannotated training set and unannotated test set, and is released under the CC BY 4.0 license.



