PGC: Peak-Guided Calibration for Generalizable AI-Generated Image Detection

The rapid evolution of generative AI, from GANs to modern diffusion models, has resulted in increasingly subtle discriminative clues. These fine-grained signals are often overshadowed by dominant, high-fidelity image content (e.g., the main subject), limiting the reliability of existing detectors that predominantly rely on global representations. To address this challenge, we propose the Peak-Guided Calibration (PGC) framework. PGC introduces a novel strategy that aggregates salient features via a peak-focusing mechanism. Specifically, by employing a peak-sensitive aggregation that accentuates the most discriminative local clues, PGC leverages these critical signals to calibrate the global decision. This approach recovers subtle patterns that would otherwise be submerged in the global context. Furthermore, to better simulate real-world threats, we introduce the CommGen15 dataset, a challenging benchmark comprising samples from 15 commercial models. Extensive experiments demonstrate that PGC achieves state-of-the-art performance. Specifically, it improves mean accuracy by +12.3% on our CommGen15 dataset, and sets new records on standard benchmarks, including GenImage (+2.1%), AIGI (+3.5%), and UniversalFakeDetect (+1.7%). Code is available at https://github.com/xiaoyu6868/PGC.

As AI image generators improve, their outputs become increasingly difficult for both people and existing detectors to distinguish from real images. This makes detection challenging because useful clues may be subtle and localized to certain regions. Many existing detectors summarize an image globally, so these local clues can be weakened by the realistic main content of the image. We propose Peak-Guided Calibration, a method that identifies the most informative local regions and uses them to adjust the detector’s overall decision. Rather than treating all parts of an image as equally important, the method gives more weight to regions that provide stronger evidence for whether an image is AI-generated. We also introduce CommGen15, a benchmark containing images and video frames from 15 commercial generation models, to better reflect current real-world generators. Across several benchmarks, our method improves detection accuracy, especially for images produced by commercial models. These results suggest that focusing on localized evidence can make AI-generated image detection more reliable.