Relative Patch Response Learning for Generalizable AI-Generated Image Detection
The Relative Patch Response Learning (PRL) method improves AI-generated image detection by learning from how image patches change between two mixed views, instead of using fixed source labels. PRL uses three objectives—relative response, reference coherence, and area ranking—to supervise these patch responses. This approach allows the detector to generalize better to unseen generators and is robust across standard and in-the-wild benchmarks.