Synthetic Image Detection with CLIP: Understanding and Assessing Predictive Cues

A new study reveals how CLIP-based models detect synthetic images with up to 96% mAP accuracy, but struggle to generalize across different AI generators, dropping to as low as 37% mAP. The research introduces SynthCLIC, a high-quality dataset to reduce bias, and finds detectors rely on high-level photographic attributes like style and lens flare, not just low-level artifacts. This highlights the need for continual updates to keep detection robust against evolving generative AI.