Spotting AI-made images by hunting for hidden traces

AI image generators keep getting better, making fake images harder to spot because obvious tells get buried under realistic details. This work introduces Peak-Guided Calibration, which identifies and amplifies the subtle, discriminative patterns that distinguish AI images from real ones—like tiny statistical anomalies detectors usually miss. On a new benchmark mixing outputs from 15 different generators, it achieves 12.3% higher accuracy; it also improves detection on existing benchmarks by 2–3.5%. Code released.