This paper proposes a novel Plug-in Diffusion Model for Recommendation (PDRec) framework that leverages diffusion models as a flexible plugin to jointly utilize diffusion-generated user preferences on all items. The main goal is to address the data sparsity issue in recommendation by fully utilizing the diffusion-based preferences on both observed and unobserved items. PDRec introduces three key components: Historical Behavior Reweighting (HBR), Diffusion-based Positive Augmentation (DPA), and Noise-free Negative Sampling (NNS). HBR identifies high-quality behaviors and suppresses noisy interactions, DPA leverages top-ranked unobserved items as potential positive samples to alleviate data sparsity, and NNS selects stable negative samples to mitigate false negative sampling issues. The framework is designed to be model- and task-agnostic, enabling its application across different recommendation scenarios. Extensive experiments on four real-world datasets demonstrate that PDRec outperforms state-of-the-art baselines in various recommendation tasks, including sequential recommendation and cross-domain sequential recommendation. The code is available at https://github.com/hulkima/PDRec.This paper proposes a novel Plug-in Diffusion Model for Recommendation (PDRec) framework that leverages diffusion models as a flexible plugin to jointly utilize diffusion-generated user preferences on all items. The main goal is to address the data sparsity issue in recommendation by fully utilizing the diffusion-based preferences on both observed and unobserved items. PDRec introduces three key components: Historical Behavior Reweighting (HBR), Diffusion-based Positive Augmentation (DPA), and Noise-free Negative Sampling (NNS). HBR identifies high-quality behaviors and suppresses noisy interactions, DPA leverages top-ranked unobserved items as potential positive samples to alleviate data sparsity, and NNS selects stable negative samples to mitigate false negative sampling issues. The framework is designed to be model- and task-agnostic, enabling its application across different recommendation scenarios. Extensive experiments on four real-world datasets demonstrate that PDRec outperforms state-of-the-art baselines in various recommendation tasks, including sequential recommendation and cross-domain sequential recommendation. The code is available at https://github.com/hulkima/PDRec.