Defactify: 10K-100K AI-Generated Images for Veracity Benchmarking

Between 10,000 and 100,000 images and metadata comprise this benchmark for AI-generated content detection, created by Rajarshi-Roy-research in 2025. The collection pairs real-world photographs with synthetic images generated by SD21, SDXL, SD3, DALLE3, and Midjourney.

Use Cases

• Binary classification of real vs. synthetic images using the veracity labels
• Source attribution to identify which specific model (SD21, SDXL, SD3, DALLE3, or Midjourney) produced an image
• Benchmarking detection algorithms against the specific artifacts of the Stable Diffusion 3 (SD3) architecture

Strengths

• Includes 10,000 to 100,000 image records
• Covers five distinct AI generation models including SD3 and Midjourney
• Distributed in optimized Parquet format for high-performance data loading

Limitations

• Geographic bias toward the US region as noted in the metadata tags
• Undocumented schema for metadata columns in the initial release
• Unknown license restrictions which may limit commercial application

Provenance

Source

Rajarshi-Roy-research

Collection Method

Synthetic generation using AI models and collection of real-world images.

Freshness

Last updated November 2025.

Geography

United States