DiffuserLite is a lightweight diffusion planning framework that significantly improves decision-making frequency and performance on D4RL benchmarks. It employs a planning refinement process (PRP) to generate coarse-to-fine-grained trajectories, reducing redundant information and enhancing planning efficiency. DiffuserLite achieves a decision-making frequency of 122Hz, which is 112.7 times faster than previous frameworks. It outperforms existing methods in terms of inference time, model size, and performance on three popular D4RL benchmarks. DiffuserLite is designed as a flexible plugin that can be integrated into other diffusion planning algorithms, offering a structural design reference for future works. The framework reduces the complexity of the fitted distribution, enabling the use of lighter neural networks and fewer denoising steps. It also reduces the plan-search space, allowing the planner to focus on distant key points and immediate steps. DiffuserLite's architecture includes three variations based on different backbones: diffusion models, rectified flow, and rectified flow with an additional reflow step. Experiments show that DiffuserLite achieves significant performance improvements across various domains, maintaining high decision-making frequency while achieving state-of-the-art results. It is also effective in sparse reward tasks when using value-assisted guidance. The framework's lightweight design and efficient planning process make it suitable for real-time applications. However, it currently faces limitations due to the classifier-free guidance (CFG) mechanism, which requires adjusting the target condition. Future work may focus on improving guidance mechanisms and simplifying the framework. DiffuserLite contributes to increasing decision-making frequency and performance in diffusion planning.DiffuserLite is a lightweight diffusion planning framework that significantly improves decision-making frequency and performance on D4RL benchmarks. It employs a planning refinement process (PRP) to generate coarse-to-fine-grained trajectories, reducing redundant information and enhancing planning efficiency. DiffuserLite achieves a decision-making frequency of 122Hz, which is 112.7 times faster than previous frameworks. It outperforms existing methods in terms of inference time, model size, and performance on three popular D4RL benchmarks. DiffuserLite is designed as a flexible plugin that can be integrated into other diffusion planning algorithms, offering a structural design reference for future works. The framework reduces the complexity of the fitted distribution, enabling the use of lighter neural networks and fewer denoising steps. It also reduces the plan-search space, allowing the planner to focus on distant key points and immediate steps. DiffuserLite's architecture includes three variations based on different backbones: diffusion models, rectified flow, and rectified flow with an additional reflow step. Experiments show that DiffuserLite achieves significant performance improvements across various domains, maintaining high decision-making frequency while achieving state-of-the-art results. It is also effective in sparse reward tasks when using value-assisted guidance. The framework's lightweight design and efficient planning process make it suitable for real-time applications. However, it currently faces limitations due to the classifier-free guidance (CFG) mechanism, which requires adjusting the target condition. Future work may focus on improving guidance mechanisms and simplifying the framework. DiffuserLite contributes to increasing decision-making frequency and performance in diffusion planning.