Title: RMPrior: Integrating Propagation Priors with Diffusion Refinement for Streamlined Radio Map Generation

Abstract: While diffusion models excel at generating high-fidelity radio maps via iterative denoising, their high sampling costs hinder practical deployment in dynamic wireless environments that require frequent map updates. Conversely, traditional propagation models possess valuable scene-level insights that are typically ignored by standard diffusion inference, which begins with pure Gaussian noise. To address this, we introduce a "mid-start" sampling approach that integrates propagation priors with diffusion refinement. This method perturbs a matched propagation prior to an intermediate diffusion stage, allowing the pretrained diffusion backbone to perform only the subsequent reverse steps. This strategy directs computational resources toward multipath-aware refinement rather than reconstructing the entire map from noise. We present a theoretical framework that defines an upper bound for the initialization gap, identifies sufficient conditions for enhanced reconstruction fidelity through truncation, and formally characterizes how prior quality sensitivity behaves under aggressive truncation. Evaluations on the IRT4HighRes dataset demonstrate that at $P_{\text{start}}=0.5$, our approach delivers a $2.01\times$ acceleration while outperforming the full-step baseline in NMSE, RMSE, SSIM, and PSNR. Furthermore, an ablation study across three propagation models of varying fidelity reveals that reconstruction accuracy correlates with prior quality, with sensitivity increasing as reverse trajectories shorten, aligning with our theoretical findings. These outcomes imply that mid-start reconstruction performance can function as a reliable metric for ranking the scene-level fidelity of various propagation models.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC