Artificial intelligence has dramatically promoted and prospered research in many fields, such as recommendation systems (RSs). Due to privacy concerns, many data-sharing algorithms in RSs have been proposed to alleviate the data sparsity problem while ensuring privacy through differential privacy (DP). However, they face challenges with complex sensitivity calculations and reduced performance under strict privacy budgets. To confront these issues, we present a differentially private recommendation scheme based on the diffusion model (DM) and Rényi divergence. First, we introduce a redesigned DM (DPDM) to generate item rating frequency matrices that straightforwardly yield DP and manage utility without harming privacy. Subsequently, these perturbed matrices can be safely released to overcome data scarcity. Then, we construct a DP-shared recommendation method based on Rényi divergence (DPDM-RA), which is grounded in the frequency distribution sampled by the DPDM to meet (ϵ,δ)-DP and perform robustly in tiny privacy budgets. We further develop a skew-symmetric factor for DPDM-RA to enhance reliability. Theoretical proofs and experimental results indicate our model achieves high-quality recommendations while providing solid privacy protection.