Multimodal Recommender Systems (MMRS) alleviate data sparsity and enhance item understanding by integrating auxiliary information, such as images and text, making them a prominent research direction. However, they face two limitations. Traditional models typically statically concatenate an item’s modal and behavioral features, restricting the effective utilization of its key physical attributes for improving recommendation accuracy. Additionally, graph propagation can induce representation over-smoothing; after several rounds of information aggregation, the representations of modally-similar items homogenize. This process masks critical personalized features that attract specific users, thereby diminishing the discriminative power necessary for precise recommendations. To address these limitations, we propose a novel framework for multimodal recommendation, the Dual Calibration and Sharpening Network (DCSN). This framework operates on a two-stage calibration-sharpening paradigm to optimize feature fusion and representation quality. For calibration, DCSN employs a collaborative interaction gating module where an item’s physical attributes serve as dynamic contexts to reshape its core behavioral representation, thus overcoming static fusion limitations. To counteract representation over-smoothing, the sharpening stage introduces a Decoupled Modal Contrastive Control (DMCC) module. This module utilizes a combined contrastive and reconstruction loss to refine the modal representations, maximizing their discriminative power while enhancing robustness. Extensive experiments conducted on two public datasets validate the effectiveness of our method.