CLIP-BCP: Generalizable AI-Generated Image Detection via CLIP Mid-Layer Representations and Bright Channel Prior

The rapid progress of AI-generated image (AIGI) synthesis poses substantial challenges to AIGI detection, especially under cross-generator and cross-domain distribution shifts. We propose CLIP-BCP, a dual-branch framework that combines the mid-layer CLS token of a frozen CLIP encoder with a bright channel prior (BCP). On two benchmarks, we show that mid-layer CLIP features provide stronger real/fake separability while balancing semantic context and fine-grained artifacts. CLIP-BCP freezes the pretrained CLIP encoder and uses the 13th-layer CLS token as its semantic backbone. In parallel, the BCP branch encodes multiscale bright channel response distributions through morphological dilation and convolutional encoding. Feature-level fusion combines these distributions with the mid-layer CLIP representation. Experiments on AIGCDetect-Benchmark and GenImage show that CLIP-BCP achieves mean accuracy (mAcc) values of 94.28% and 98.29%, outperforming the previous best methods by 1.51 and 2.49 percentage points, respectively. Strict cross-dataset evaluation, postprocessing robustness tests, and ablations of the prior and fusion strategy further support the complementarity of mid-layer CLIP features and the BCP under the evaluated settings. The source code is publicly available at https://github.com/LeiChen0407/CLIP-BCP