DRAGON: 2.5 Million Realistic Images Generated by 25 Diffusion Models

A large-scale dataset of 2.5 million training and 100,000 test images generated using 25 diffusion models, created by lesc-unifi and released in May 2025. The dataset is designed to include images from both recent and older, well-established generative architectures.

Use Cases

• Benchmarking the performance and artifacts of different diffusion models based on the included 25 architectures.
• Training classifiers or detectors for synthetic image identification based on the large-scale labeled collection.
• Analyzing the evolution of image quality and characteristics across model generations based on the span of architectures.
• Developing data augmentation or generation techniques using a diverse corpus of synthetic images.

Strengths

• Contains 2.5 million training images and 100,000 test images, providing substantial scale.
• Includes images generated by 25 different diffusion models, offering architectural diversity.
• Spans both recent advancements and older, well-established models, allowing for temporal comparison.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is unspecified beyond the provided totals, which may limit suitability assessment for specific splits.
• The dataset is synthetic, which may limit applicability to tasks requiring real-world data distributions.

Provenance

Source

lesc-unifi

Collection Method

Images generated programmatically using a collection of 25 diffusion models.

Time Range

Models span both recent and older architectures, but specific generation dates are not provided.

Freshness

Last updated 2025-05-19 12:27:29.

Geography

null