TGIF2
收藏资源简介:
TGIF2是由根特大学和希腊研究与技术中心联合构建的文本引导修复伪造数据集,作为TGIF的扩展版本,新增了FLUX.1模型生成的编辑图像及随机非语义掩码。该数据集包含约7.5万张高分辨率图像(最高1024×1024像素),源自MS-COCO的3000张基础图像,涵盖拼接和完全再生两类篡改方式。通过集成Stable Diffusion、Adobe Firefly等先进模型,数据集重点捕捉生成式AI在局部图像编辑中产生的法医痕迹,用于评估伪造定位与合成检测方法的鲁棒性,尤其针对超分辨率攻击等新兴挑战。
TGIF2 is a text-guided inpainting forgery dataset jointly developed by Ghent University and the Centre for Research and Technology Hellas. As an extended version of the original TGIF dataset, it newly includes edited images generated by the FLUX.1 model and random non-semantic masks. This dataset contains approximately 75,000 high-resolution images (up to 1024×1024 pixels), which are derived from 3,000 base images from the MS-COCO dataset, and covers two types of tampering methods: splicing and full regeneration. By integrating advanced models such as Stable Diffusion and Adobe Firefly, this dataset focuses on capturing forensic traces generated by generative AI during local image editing, and is designed to evaluate the robustness of forgery localization and synthetic detection methods, particularly against emerging challenges such as super-resolution attacks.
TGIF: Text-Guided Inpainting Forgery Dataset
数据集概述
- 数据量:约75,000张伪造图像。
- 图像来源:原始图像来自MS-COCO,采用CC BY 4.0 许可,分辨率最高达1024x1024像素。
- 伪造方法:使用文本引导的图像修复方法(SD2、SDXL和Adobe Firefly)进行图像篡改。
- 图像类型:包括篡改区域拼接的原图像(SD2-sp, PS-sp)和完全重新生成的图像(SD2-fr, SDXL-fr)。
数据集许可
- 该数据集遵循CC BY-SA 4.0 许可。
数据集内容
- 伪造图像:
- SD2:46 GB
- SDXL:41 GB
- Adobe Firefly:17.8 GB
- 真实图像:
- SD2:4 GB
- SDXL crops:3 GB
- 掩码:
- SD2
- SDXL
- Photoshop masks
- 元数据:
- SD2
- SDXL




