Dataset Card for DRAGON

🧾 ArXiv Preprint

DRAGON is a large-scale Dataset of Realistic imAges Generated by diffusiON models.

The dataset includes a total of 2.5 million training images and 100,000 test images generated using 25 diffusion models, spanning both recent advancements and older, well-established architectures.

Dataset Details

Dataset Description

The remarkable ease of use of diffusion models for image generation has led to a proliferation of synthetic content online. While these models are often employed for legitimate purposes, they are also used to generate fake images that support misinformation and hate speech. Consequently, it is crucial to develop robust tools capable of detecting whether an image has been generated by such models. To support the development of such tools, we introduce DRAGON: a large-scale dataset of images generated by 25 diffusion models, featuring a diverse range of subjects and is available in multiple sizes (from extra-small to extra-large) to suit various research needs. The dataset also includes a dedicated test set designed as a standardized benchmark for evaluating detection methods and model attribution systems.

• Curated by: Signal Processing and Communications Lab (LESC)
• License: Creative Commons Attribution Share Alike 4.0 International

Uses

DRAGON is designed to support the development of new multimedia forensics tools for diffusion models, with a specific focus on synthetic image detection and model attribution tasks.

Dataset Structure

Each image in the DRAGON dataset is annotated with both the generative model used to create it and the input prompt. For each of the 1,000 ImageNet classes, a corresponding prompt was generated. Using these prompts, 100 training images and 4 test images were produced per model, resulting in a total of 2.5 million training images and 100,000 test images.

To support a range of research needs, DRAGON includes predefined subsets of varying sizes:

• ExtraSmall (XS): 10 prompts, 250 training images, 1,000 test images
• Small (S): 100 prompts, 2,500 training images, 10,000 test images
• Regular (R): 100 prompts, 25,000 training images, 10,000 test images
• Large (L): 1,000 prompts, 250,000 training images, 100,000 test images
• ExtraLarge (XL): 1,000 prompts, 2,500,000 training images, 100,000 test images

Citation

@misc{bertazzini2025dragon,
  title={DRAGON: A Large-Scale Dataset of Realistic Images Generated by Diffusion Models}, 
  author={Giulia Bertazzini and Daniele Baracchi and Dasara Shullani and Isao Echizen and Alessandro Piva},
  year={2025},
  eprint={2505.11257},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2505.11257}, 
}

Dataset Card Contact

• Giulia Bertazzini: [email protected]
• Daniele Baracchi: [email protected]