Test-Time Guided Diffusion for Real-Time and Fine-Grained Radio Map Estimation With Ultra-Sparse Samples
Abstract
Accurate radio maps are essential for wireless network planning and optimization, while acquiring dense measurements in practical deployments is often costly and time-consuming. Recent diffusion-based methods provide a promising solution for reconstructing high-resolution radio maps from extremely sparse observations. Nevertheless, under ultra-low sampling rates, one-pass generation may suffer from unstable structures and inconsistent propagation patterns due to limited measurements and uncertainty in the generative process. To address this issue, this paper develops a Test-Time Guided Diffusion framework for efficient radio map reconstruction. The proposed method follows an “adapt-and-refine” strategy that steers a single denoising trajectory from random noise toward a propagation-consistent radio map. In the Adapt phase, a lightweight network produces a coarse estimate and mines sample-specific environmental cues, including source-related high-power anchors and obstacle-sensitive locations. In the Refine phase, these cues are injected into different stages of the diffusion process to stabilize global structure, recover fine spatial details, and enforce consistency with known measurements. Simulation results show that the proposed framework achieves accurate and fast radio map reconstruction when the sampling rate is below 0.1%, providing competitive reconstruction fidelity with substantially reduced inference latency compared with existing diffusion-based approaches.