Community Forensics Evaluation Set with 21 Generative Models

Community Forensics Eval is an evaluation dataset for fake image detection, containing images from 21 generative models paired with real source datasets. The dataset was created by OwensLab for the CVPR 2025 paper "Community Forensics: Using Thousands of Generators to Train Fake Image Detectors". It was last updated in October 2025.

Use Cases

• Benchmarking fake image detectors using the 21 generative models as distinct classes.
• Analyzing detection performance across different real source datasets like RAISE, COCO, FFHQ, and LAION.
• Training classifiers to distinguish between real images and specific generator outputs.
• Studying the generalization of detectors across a diverse set of image generation techniques.

Strengths

• Includes 21 distinct generative models, providing a diverse testbed.
• Pairs generated images with real images from established datasets like COCO and FFHQ.
• Created for a peer-reviewed CVPR 2025 publication.

Limitations

• Dataset size, row count, and specific image counts are not provided.
• Limited to non-commercial research and educational use only.
• Evaluation set only; does not contain the full training data from the paper.

Provenance

Source

OwensLab, associated with the CVPR 2025 paper.

Collection Method

Curated as an evaluation benchmark from generated and real image datasets.

Freshness

Last updated October 2025.