Spectral Tail Auxiliary Learning for AI-Generated Image Detection
Spectral Tail Auxiliary Learning (STAL) is a method to detect AI-generated images by using frequency information during training. It found that generated images have an unusual pattern in their high-frequency spectrum called 'spectral tail uplift.' STAL transfers this spectral signal to a spatial detector without adding any extra steps during image prediction, leading to better detection accuracy and stability across different AI generators.