The paper introduces a novel approach to motion planning using potential-based methods, leveraging diffusion models to learn and optimize motion trajectories. Traditional potential-based motion planning is effective but prone to local minima and requires solving a global optimization problem. The proposed method trains a neural network to learn easily optimizable potentials over motion planning trajectories, avoiding local minima issues. The approach is compositional, allowing different motion constraints to be combined by adding corresponding potentials, enabling the planner to handle a wide range of constraints. The effectiveness of the method is demonstrated through experiments, showing superior performance compared to both classical and recent learned motion planning approaches. The method also generalizes well to new environments and constraints, as shown through its compositional nature and ability to handle complex scenarios. The paper includes detailed evaluations on various simulated and real-world datasets, highlighting the planner's accuracy, efficiency, and robustness.The paper introduces a novel approach to motion planning using potential-based methods, leveraging diffusion models to learn and optimize motion trajectories. Traditional potential-based motion planning is effective but prone to local minima and requires solving a global optimization problem. The proposed method trains a neural network to learn easily optimizable potentials over motion planning trajectories, avoiding local minima issues. The approach is compositional, allowing different motion constraints to be combined by adding corresponding potentials, enabling the planner to handle a wide range of constraints. The effectiveness of the method is demonstrated through experiments, showing superior performance compared to both classical and recent learned motion planning approaches. The method also generalizes well to new environments and constraints, as shown through its compositional nature and ability to handle complex scenarios. The paper includes detailed evaluations on various simulated and real-world datasets, highlighting the planner's accuracy, efficiency, and robustness.