FragFake
收藏资源简介:
FragFake是一个用于编辑图像检测的大规模数据集,由多个高级图像编辑模型生成的编辑图像组成,包括对象添加和对象替换两种类型的编辑操作。数据集包含超过20,000个图像-文本对,用于训练视觉语言模型。FragFake旨在解决现代图像编辑技术产生的局部编辑图像的检测问题,通过自动化的数据生成管道构建,以减少对昂贵的像素级注释的依赖。数据集的创建过程包括从COCO数据集中随机采样图像、使用GPT-4o生成编辑指令、使用四个编辑模型生成编辑图像、将图像转换为图像-文本对格式,并进行人工审核以确保正确性。FragFake适用于多模态内容真实性的研究,并有望推动该领域后续研究的发展。
FragFake is a large-scale dataset for edited image detection, composed of edited images generated by multiple state-of-the-art image editing models, covering two types of editing operations: object addition and object replacement. The dataset contains over 20,000 image-text pairs for training vision-language models. FragFake aims to address the detection problem of locally edited images produced by modern image editing technologies, and is constructed via an automated data generation pipeline to reduce reliance on expensive pixel-level annotations. The dataset creation process includes randomly sampling images from the COCO dataset, generating editing instructions using GPT-4o, generating edited images using four editing models, converting the images into image-text pair format, and conducting manual reviews to ensure correctness. FragFake is applicable to research on multimodal content authenticity, and is expected to promote the development of subsequent research in this field.




