DiffSeg30k
收藏资源简介:
DiffSeg30k是由新加坡国立大学与华南理工大学联合构建的扩散编辑检测基准数据集,包含3万张具备像素级标注的扩散编辑图像。该数据集融合真实场景图像与AI生成内容,涵盖八种前沿扩散模型的多轮序列编辑,每幅图像最多经历三次局部修改。通过视觉语言模型驱动的自动化流程,系统实现语义区域识别与上下文感知的编辑提示生成,支持添加、删除及属性修改三类操作。该数据集专为细粒度AIGC检测而设计,推动研究从二值分类转向语义分割范式,致力于解决扩散模型局部编辑的精准定位与溯源 attribution 等核心挑战。
DiffSeg30k is a benchmark dataset for diffusion edit detection, jointly constructed by the National University of Singapore and South China University of Technology. It contains 30,000 diffusion-edited images with pixel-level annotations. This dataset blends real-world scene images and AI-generated content, covering multi-round sequential edits from eight state-of-the-art diffusion models, with each image undergoing up to three local modifications. Through an automated pipeline driven by vision-language models, the system enables semantic region recognition and context-aware edit prompt generation, supporting three types of operations: addition, deletion, and attribute modification. Designed specifically for fine-grained AIGC detection, this dataset promotes the shift of research from binary classification to semantic segmentation paradigms, and aims to address core challenges such as accurate localization and attribution of local edits by diffusion models.
DiffSeg30k 数据集概述
数据集简介
DiffSeg30k 是一个用于分割基于扩散编辑的多轮编辑数据集,适用于训练和评估能够定位编辑区域并识别底层扩散模型的模型。
数据集内容
- 编辑图像文件:格式为
xxxxxxxx.image.png,每张图像可能经过1次、2次或3次编辑操作 - 对应掩码文件:格式为
xxxxxxxx.mask.png,像素值编码编辑类型和使用的扩散模型
数据加载方式
python from datasets import load_dataset dataset = load_dataset("Chaos2629/Diffseg30k", split="train") image, mask = dataset[0][image], dataset[0][mask]
掩码标注规范
掩码为灰度图像(PNG格式),像素值与特定编辑模型的对应关系:
| 掩码值 | 编辑模型 |
|---|---|
| 0 | background |
| 1 | stabilityai/stable-diffusion-2-inpainting |
| 2 | kolors |
| 3 | stabilityai/stable-diffusion-3.5-medium |
| 4 | flux |
| 5 | diffusers/stable-diffusion-xl-1.0-inpainting-0.1 |
| 6 | glide |
| 7 | Tencent-Hunyuan/HunyuanDiT-Diffusers |
| 8 | kandinsky-community/kandinsky-2-2-decoder-inpaint |
重要说明
- 每张编辑图像可能经过多轮编辑,对应掩码可能包含多个不同的标签值(范围0-8)
许可证
Apache-2.0




