AutoSplice: A Text-prompt Manipulated Image Dataset for Media Forensics

This repository provides the processed AutoSplice benchmark dataset for image forgery detection and manipulation localization, formatted consistently with other benchmarks such as KhangTruong/CocoGlide, KhangTruong/OpenSDID, KhangTruong/Casia-v2.0, and KhangTruong/Diffseg30k.

Dataset Overview

• Dataset Name: AutoSplice (Forged_JPEG100 canonical baseline)
• Split: cross_test
• Total Samples: 3621
• Generator: DALL-E 2 language-image model guided by contextual text prompts
• Columns:
  • image (Image): The tampered image (lossless JPEG quality 100).
  • mask (Image): Single-channel ground-truth binary mask (0 = pristine / background, 255 = tampered / spliced region). Samples with only black masks (no tampered regions) have been removed.

Usage

from datasets import load_dataset

dataset = load_dataset("KhangTruong/AutoSplice", split="cross_test")

sample = dataset[0]
image = sample["image"]
mask = sample["mask"]

print(image.size, mask.size)

Citation

@inproceedings{jia2023autosplice,
  title={AutoSplice: A Text-prompt Manipulated Image Dataset for Media Forensics},
  author={Jia, Shan and Huang, Mingzhen and Zhou, Zhou and Ju, Yan and Cai, Jialing and Lyu, Siwei},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  pages={893--903},
  year={2023}
}

License & Usage Notes

Released for non-commercial research and educational purposes in accordance with the original AutoSplice dataset terms.