CatAID: Category-Guided AI-Generated Image Detection via Vision-Language Model Adaptation

The growing use of AI-generated content (AIGC) has spurred research on detecting AI-generated images. While current methods prioritize generalization capability across unseen generators, they may overlook the limitation posed by unseen semantic content. This work seeks a semanticagnostic detection method for better generalization and robustness. Leveraging the powerful semantic representation of Vision-Language Models (VLMs), we propose a Category-guided AI-generated Image Detection, termed CatAID. Our approach integrates Category-Contextualized Prompts to adapt VLMs for the detection task, informing the detector with explicit semantic concepts. This reduces reliance on semantic learning and thereby encourages the VLM's responses to align with more semantic-invariant forgery patterns. Extensive evaluation of AI-generated images from 33 generators in 4 datasets and various unknown semantic contents demonstrates the improved performance of CatAID over state-of-the-art methods.