nebula/OpenSDI_test

OpenSDI: Spotting Diffusion-Generated Images in the Open World

This dataset is designed to address the OpenSDI challenge: spotting diffusion-generated images in realistic, open-world scenarios. It is described in the paper:

Project Page: https://iamwangyabin.github.io/OpenSDI/

OpenSDID Dataset Highlights:

• User Diversity: Simulates a wide range of user intentions and creative styles using diverse text prompts generated by VLMs.
• Model Innovation: 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.

Dataset Overview:

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

Download:

• OpenSDI Training Dataset: https://huggingface.co/datasets/nebula/OpenSDI_train
• OpenSDI Testing Dataset: This repository

This dataset is for academic use and is distributed under the CC BY-SA 4.0 license. Real images come from megalith-10m.