DRAGON: A Large-Scale Dataset of Realistic Images Generated by Diffusion Models

TL;DR

Diffusion models enable widespread synthetic image generation, raising misinformation risks and the need for robust detection and attribution tools. The paper introduces DRAGON, a large-scale dataset with 2.6 million synthetic images from 25 diffusion models plus 1.33 million real ImageNet images, enhanced by an LLM-driven prompt-expansion pipeline and organized into scalable subsets with a dedicated test set. Through quality and forensic analyses, the work shows that prompt expansion improves perceptual quality and reveals model-specific spectral fingerprints, and that retraining detectors on DRAGON enhances robustness to post-processing. DRAGON thus provides a valuable, up-to-date benchmark for developing and evaluating diffusion-model detection and attribution techniques in a rapidly evolving landscape.

Abstract

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. Many current detection methods, however, require large volumes of sample images for training. Unfortunately, due to the rapid evolution of the field, existing datasets often cover only a limited range of models and quickly become outdated. In this work, we introduce DRAGON, a comprehensive dataset comprising images from 25 diffusion models, spanning both recent advancements and older, well-established architectures. The dataset contains a broad variety of images representing diverse subjects. To enhance image realism, we propose a simple yet effective pipeline that leverages a large language model to expand input prompts, thereby generating more diverse and higher-quality outputs, as evidenced by improvements in standard quality metrics. The dataset is provided in multiple sizes (ranging from extra-small to extra-large) to accomodate different research scenarios. DRAGON is designed to support the forensic community in developing and evaluating detection and attribution techniques for synthetic content. Additionally, the dataset is accompanied by a dedicated test set, intended to serve as a benchmark for assessing the performance of newly developed methods.

Figures (5)

• Figure 1: DRAGON dataset example for the prompt "woodland scene featuring a cottontail rabbit in its natural habitat, soft focus background of trees and undergrowth, high resolution detailing the fur texture, close-up shot capturing the distinct white tuft on its tail, using Canon EOS R5, wide lens" across all models.
• Figure 2: DRAGON prompt expansion pipeline. The black path shows the baseline approach, in which the label is used directly as the prompt. The green dashed path shows the enhanced approach, where a LLM, guided by multiple seed examples, expands the label into a higher-quality prompt. The improved prompt yields a higher Multi-dimensional Preference Scoring (MPS).
• Figure 3: Comparison of image quality (MPS score) with and without LLM-based prompt expansion. For each diffusion model, the left image is generated using the original ImageNet label as the prompt, while the right image is generated using the corresponding LLM-expanded prompt.
• Figure 4: Fourier transform (amplitude) of the average of 1000 noise residuals for each model.
• Figure 5: Model-wise attribution accuracy of DE-FAKE trained on the DRAGON-R training set. The average accuracy across all models is 0.62.