This paper introduces a potential-based diffusion motion planning approach that leverages diffusion models to learn and optimize motion planning potentials. The method addresses the limitations of traditional potential-based motion planning, such as local minima and the need for implicit obstacle representations. By using diffusion models, the approach can learn easily optimizable potentials over motion planning trajectories, enabling the generation of collision-free paths. The method is trained on a dataset of solved motion planning problems, allowing it to learn an energy-based model that can be used to generate motion plans. The approach is also compositional, allowing the combination of multiple potential functions to handle different motion constraints. The method is tested on various environments, including 2D and 14D motion planning tasks, and is shown to outperform classical and learned motion planning approaches in terms of success rate, planning time, and collision checks. The method is also able to generalize to new environments with more obstacles and different constraints. The approach is further shown to be effective in real-world scenarios, such as navigating through crowded environments with multiple agents. The method is probabilistically complete, ensuring that it can find a valid trajectory with high probability. The paper also discusses the limitations of the approach, including the need for additional computation power when combining multiple potential functions. Overall, the proposed method provides a promising approach to motion planning in high-dimensional spaces, with the ability to generalize to new environments and handle complex motion constraints.This paper introduces a potential-based diffusion motion planning approach that leverages diffusion models to learn and optimize motion planning potentials. The method addresses the limitations of traditional potential-based motion planning, such as local minima and the need for implicit obstacle representations. By using diffusion models, the approach can learn easily optimizable potentials over motion planning trajectories, enabling the generation of collision-free paths. The method is trained on a dataset of solved motion planning problems, allowing it to learn an energy-based model that can be used to generate motion plans. The approach is also compositional, allowing the combination of multiple potential functions to handle different motion constraints. The method is tested on various environments, including 2D and 14D motion planning tasks, and is shown to outperform classical and learned motion planning approaches in terms of success rate, planning time, and collision checks. The method is also able to generalize to new environments with more obstacles and different constraints. The approach is further shown to be effective in real-world scenarios, such as navigating through crowded environments with multiple agents. The method is probabilistically complete, ensuring that it can find a valid trajectory with high probability. The paper also discusses the limitations of the approach, including the need for additional computation power when combining multiple potential functions. Overall, the proposed method provides a promising approach to motion planning in high-dimensional spaces, with the ability to generalize to new environments and handle complex motion constraints.