<p>The rapid advancement of digital image processing has made image manipulation increasingly accessible, raising significant concerns about misinformation and visual forgery, particularly in critical domains such as journalism, forensics, and public security. To tackle this challenge, we propose PCD-Net, an end-to-end dual-branch network for image manipulation localization (IML). The architecture incorporates a dedicated noise branch to exploit distributional inconsistencies between tampered and authentic regions, thereby improving localization granularity. To efficiently adapt large pre-trained vision backbones to IML tasks without incurring high computational overhead, we introduce a tunable prompt module, enabling frozen backbones to extract task-relevant multi-view features while preserving their generalization ability. Additionally, we design a Feature Enhancement and Fusion Module (FEFM) that employs channel and spatial attention along with a bidirectional cross-attention mechanism to integrate complementary multi-modal, multi-scale features and suppress irrelevant information. The results demonstrate that, compared with existing approaches in the same category, PCD-Net achieves significantly higher localization accuracy across multiple datasets and exhibits superior robustness in complex manipulation scenarios.</p>