<p>The rapid advancement of generative models has increasingly blurred the boundary between synthetic and real imagery, raising urgent challenges for authenticity verification and copyright protection. To address these issues, this paper presents a multi-module detection framework that integrates lightweight visual features, auxiliary descriptors, and an dual-scale model collaborative reasoning paradigm. The system first extracts complementary micro- and macro-level visual features and fuses them into a compact representation. A logical reasoning and dual-scale collaborative loop module, implemented with three fine-tuned Llama models, performs structured reasoning and iterative calibration, effectively mitigating misclassification arising from semantic ambiguity and subtle artifacts. To further enhance robustness in challenging scenarios such as post-edited or low-resolution images, a plug-in auxiliary module supplements metadata consistency cues and frequency-domain descriptors, producing an enriched feature set for final decision making. Extensive comparative and ablation experiments demonstrate that the proposed framework consistently achieves superior accuracy, stability, and generalization compared with existing VLM-based baselines. This work provides both methodological insights and practical solutions for trustworthy image authenticity verification.</p>