Sakana AI Reports Vision Model Detects AI Images

Sakana AI posted that its latest results show a strong general-purpose vision model can tell real images from AI-generated ones with only a simple decision rule on frozen representations. The work will be presented at ECCV2026. It introduces the Percept-Lens out-of-distribution evaluation framework and a Mah-NCM-based detector built on those frozen visual features. The post notes that diffusion-based generators have made synthetic images common and that detectors often fail when generator, prompt or source domain all shift at once.