Why the benchmark matters
The CIFAR Synthetic Evidence Corpus is designed for evidentiary image authentication rather than general synthetic-image detection. It covers surveillance, dashcam, and consumer photographs across three manipulation tiers, with metadata for evidentiary family, manipulation tier, subtype, source dataset, generator, prompt, and scene attributes.
The benchmark shows weak zero-shot performance
Across 1,505 items, the best tested detector, d3, achieved a ROC-AUC of 0.674. The other tested detectors performed near chance. The evaluation used released weights and default thresholds without fine-tuning, so these are benchmark measurements rather than evidence of courtroom reliability or admissibility.