MultiEdit
收藏资源简介:
MultiEdit是由中国科学院计算技术研究所和国科大学者构建的大规模多工具人脸编辑数据集,旨在模拟真实场景中多步组合编辑操作。该数据集包含517,540张编辑后的面部图像,覆盖六种代表性编辑工具(如Deepfake、美化工具)及三种编辑类型,编辑步骤从1至5步不等。数据创建过程基于200张真实人脸图像,系统枚举所有合理编辑序列并结合质量过滤(SSIM与视觉语言模型评估)生成。该数据集旨在推动多工具图像编辑溯源任务,解决现有方法仅能识别单一工具的局限,提升深度伪造检测的透明度和准确性。
MultiEdit is a large-scale multi-tool facial editing dataset constructed by the Institute of Computing Technology, Chinese Academy of Sciences and scholars from the University of Chinese Academy of Sciences, aiming to simulate multi-step combined editing operations in real-world scenarios. This dataset contains 517,540 edited facial images, covering six representative editing tools such as Deepfake and beauty tools and three editing types, with the number of editing steps ranging from 1 to 5. The data creation process is based on 200 real facial images, where the system enumerates all reasonable editing sequences and generates the dataset with quality filtering using SSIM and vision-language model evaluation. This dataset aims to promote the task of multi-tool image editing traceability, address the limitation that existing methods can only identify a single tool, and improve the transparency and accuracy of deepfake detection.
MIEA 数据集概述
MIEA(Multi-tool Image Attribution)是一个面向面部伪造场景的多工具图像溯源数据集。该数据集聚焦于识别和归因面部伪造图像所使用的具体生成工具,支持多工具场景下的图像来源追踪研究。
该数据集目前处于即将发布状态,官方代码与数据即将在项目主页公开。

- 1Multi-Tool Image Editing Attribution in Facial Forgery中国科学院计算技术研究所; 中国科学院大学 · 2026年



