Why AI-generated images have a telltale frequency signature

Real and generated images have fundamentally different frequency decay patterns: while real images follow power-law physics, AI-generated ones show an unusual energy spike in the ultra-high-frequency tail. The researchers traced this to nonlinear harmonic accumulation in generative models and built STAL, which uses a frequency-aware teacher to guide spatial detectors during training, then discards all frequency machinery at inference. Tested on 9 datasets spanning different generators and real-world conditions, it generalizes better than prior frequency-domain methods without slowing down predictions.