AI Generated image manipulation dataset

NCL_IMD.v2 (BMVC Workshop on Media Authenticity 2024)

NCL_IMD.v2 dataset

The NCL_IMD.v2 dataset was proposed in the paper "FUSION++: A Method to Detect Generative AI Manipulated Images" , submitted at the BMVC 2024 Workshop on Media Authenticity in the Age of Artificial Intelligence.
NCL_IMD.v2 dataset contains AI-manipulated images. Generated to support research in AI-manipulated image detection systems. It consists of 12k original and manipulated images along with its ground truth and prompts and spans various manipulation types, including removal, creation, replacement, and combination of manipulations. It provides a comprehensive resource for developing detection methods.

NCL_IMD.v2 dataset

Creation Methodology:

  • The original images in this dataset are taken from the COCO dataset.
  • The dataset was created in an orderly fashion using available generative AI tools such as LaMa, DALL-E, PowerPaint, and Paint by Example.

Folder Structure:

/ NCL_IMD.v2 dataset
      │
      ├── /Original/                 Folder containing the original (unmodified) images.
      │
      ├── /Manipulated/              Folder containing the images after manipulation.
      │
      ├── /Mask/                     Folder containing ground truth mask images.
      │
      └── /Caption/                  Folder containing JSON files with prompts and descriptions for each image.

Format and Size:

  • Number of Images: There are 12,000 images in each of the original, manipulated, and mask folders. Each image in the original folder has a corresponding manipulated version in the manipulated folder and an associated mask (ground truth) in the mask folder.
  • The caption folder contains 10,000 JSON files, as images resulting from removal manipulations do not have captions.
  • File Format: All images are in PNG format.

Download link will be available soon.

  • This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0).
  • Please cite the following paper if you intend to use this dataset:
    @inproceedings{aljuaid2024fusion++, title={FUSION++: A Method to Detect Generative AI Manipulated Images}, author={Aljuaid, L and Bhowmik, D}, year={2024}, publisher={Newcastle University} }