All-around forgery clues for generalizable AI-generated image detection

The rapid advancement of generative models poses significant challenges to reliable forgery detection, as existing methods often lack generalization by overfitting to a single type of artifact. We observe that generative processes introduce clues at two distinct levels: low-level forensic traces from up-sampling and high-level semantic features from the generative model. To exploit this observation, we propose a novel forensic–semantic framework that analyzes forgery evidence across these different levels. At the forensic level, the framework introduces a residual analysis method that isolates up-sampling artifacts. By computing the difference between an image and its re-sampled version, this approach effectively removes content-related information to focus on critical forgery traces. Simultaneously, at the semantic level, it employs a vision-language model fine-tuned with low-rank adaptation (LoRA) to capture the inherent semantic features. Evaluated on 30 sub-datasets, our method outperforms representative approaches by an average of 2.1% in accuracy (ACC) and 1.8% in average precision (AP), demonstrating strong generalization.