OpenSDI: Spotting Diffusion-Generated Images in the Open World

News

• [2025/3/10] We release dataset and code!

Introduce OpenSDI challenge

As text-to-image (T2I) diffusion models rapidly advance, distinguishing between real and AI-generated images becomes increasingly challenging. This repository introduces OpenSDID (Open-world Spotting of Diffusion Images Dataset), a novel benchmark designed to address the OpenSDI challenge: spotting diffusion-generated images in realistic, open-world scenarios.

OpenSDID Dataset Highlights:

• User Diversity: Simulates the wide range of user intentions and creative styles. We use diverse text prompts generated by VLMs to capture varied real-world editing scenarios.
• Model Innovation: Accounts for the rapid evolution of diffusion models. OpenSDID includes images from multiple state-of-the-art diffusion models (SD1.5, SD2.1, SDXL, SD3, Flux.1) released after 2023.
• Manipulation Scope: Covers the full spectrum of diffusion-based manipulation, from global image synthesis to precise local edits, object insertions, and background changes.

Model	Training Set		Test Set		Total
	Real	Fake	Real	Fake	Images
SD1.5	100K	100K	10K	10K	220K
SD2.1	-	-	10K	10K	20K
SDXL	-	-	10K	10K	20K
SD3	-	-	10K	10K	20K
Flux.1	-	-	10K	10K	20K
Total	100K	100K	50K	50K	300K

Dataset Overview.

Samples.

Download:

We have packaged the dataset into Hugging Face datasets for convenient distribution in the following link: OpenSDI Training Dataset OpenSDI Testing Dataset

We distribute this dataset under the CC BY-SA 4.0 license. Real images come from megalith-10m. This dataset is for academic use. While it has undergone ethical review by the University of Southampton, it is provided 'as is' without warranties. Users assume all risks and liabilities associated with its use; the providers accept no responsibility for any consequences arising from its use.

The original data (indiviual image files) will be uploaded to cloud storage later.

For dataset utilization, we recommend using IMDLBenCo, which offers many methods. And you can use our hf dataset to load the data.

Benchmark and Leaderboard: Join the OpenSDI Evaluation!

We maintain a public leaderboard at OpenSDI Leaderboard. We warmly welcome researchers to evaluate their detection and localization models using our OpenSDID dataset and submit their results to contribute to this growing benchmark.

Quick Start with MaskCLIP

To address the OpenSDI challenge, we introduce MaskCLIP, a novel method based on our Synergizing Pretrained Models (SPM) framework. MaskCLIP effectively enhances the generalization strengths of two powerful pre-trained models at image and pixel levels:

• CLIP (Contrastive Language-Image Pre-training): Provides robust image-level semantic understanding, crucial for generalizable detection.
• MAE (Masked Autoencoder): Offers fine-grained spatial representation learning, essential for accurate forgery localization at pixel level.

MaskCLIP overview.

Installation

conda create -n opensdi python=3.12 -y
conda activate opensdi
pip install -r requirements.txt
wget -P weights https://dl.fbaipublicfiles.com/mae/pretrain/mae_pretrain_vit_base.pth

Train

train.sh

Test

test.sh

Citation

If you find OpenSDI useful for your research and applications, please cite using this BibTeX:

@article{wang2025opensdi,
  title         = {OpenSDI: Spotting Diffusion-Generated Images in the Open World},
  author        = {Wang, Yabin and Huang, Zhiwu and Hong, Xiaopeng},
  year          = {2025},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2503.19653},
}

References & Acknowledgements

We sincerely thank IML-ViT and IMDLBenCo for their exploration and support.