<p>Knowledge Graphs (KGs), as an important semantic resource for enhanced recommender systems, contains rich entity relationships that can effectively improve recommendation performance. Despite the significant progress of existing approaches based on Graph Neural Networks and Comparative Learning, not all relationships within the KGs are beneficial for the target recommendation task. In fact, redundant relations in the KGs are prone to introduce semantic noise that interferes with user preference modeling. In addition, comparison-based learning approaches fall short in modeling fine-grained user preferences and cannot further leverage KGs information for prediction, leading to sub-optimal performance. To fill this research gap, we propose a new Diffusion-based Contrastive Knowledge Graph Recommendation (DCKG), which achieves knowledge-aware recommendation optimization by innovatively fusing the generative diffusion model with the contrastive learning mechanism DCKG. Specifically, DCKG designs a diffusion mechanism for KGs, which effectively filters irrelevant relationships by generating task-related subgraphs through forward noise injection and backward denoising recovery; at the same time, it introduces a graph Transformer to integrate global information, and utilizes global information of user-item-entity interactions to guide the process of information aggregation, and obtains the diffused KGs and user-item fusion representation. In addition, a multi-level collaborative comparison mechanism is adopted to compare the attribute importance from global (item attributes) and local (user behavior) perspectives, respectively, to optimize the alignment capability of the user-item representation. Experimental results show that DCKG outperforms state-of-the-art methods on multiple datasets.</p>