Multimodal recommender systems leverage diverse information sources, such as visual, textual, and auditory features, to better model user preferences. However, existing diffusion-based methods often face challenges of semantic drift during sampling and inadequate alignment in the target space, which limit their ability to accurately capture complex user interests. To address these issues, we propose C \(^2\) -DiffMM, a novel multimodal recommendation framework that introduces a Conditioned Denoising Module (CDM) and a Cross-Denoising Alignment (CDA) mechanism. The CDM explicitly incorporates multimodal semantic conditions into the reverse diffusion process, allowing the denoiser to perceive item-level multimodal content and effectively mitigate semantic drift. Furthermore, the CDA employs a contrastive alignment strategy that operates naturally within the diffusion process without requiring manual data perturbations, thereby enforcing global semantic consistency and enhancing the discriminability of item embeddings. Extensive experiments on three publicly available datasets demonstrate that C \(^2\) -DiffMM consistently outperforms several strong baselines, achieving notable improvements in both recommendation accuracy and semantic coherence. These results validate the effectiveness and superiority of the proposed framework for multimodal recommendation tasks.