<p>Denoising-based planning, which gradually refines noisy initial trajectories, has shown strong performance in generating safe and optimized ones. However, such approaches often suffer from slow convergence to an optimal trajectory, which can be problematic in real-world applications where computational delays may lead to planning failures or unsafe behavior. In this paper, we propose a novel learning-based path-planning method that leverages a normalizing flow framework to adaptively adjust the trajectory length, thereby enabling the planner to operate within a given computational time budget. To achieve this, our network refines a single step in the trajectory at each iteration, sequentially progressing along the planning horizon, rather than refining the entire trajectory as in diffusion-based methods. Consequently, our proposed method improves the flexibility of path planning by allowing the receding horizon length to be dynamically adjusted according to computational constraints. Furthermore, the inference results of the network can serve as initial guidance within the model predictive path integral framework, thereby reducing the number of optimization iterations required for path convergence and improving overall optimization efficiency. We validate the efficiency of our method through comparisons with diffusion-like approaches in simulation and demonstrate its practicality through hardware experiments.</p>