Problem

Accurate radio maps are critical for wireless communication, sensing, and network planning, but their acquisition remains a significant challenge. Traditional methods, such as dense measurements or physics-based simulations (e.g., ray tracing), are computationally expensive, time-consuming, and lack the flexibility required for large-scale deployment and real-time adaptation in dynamic environments. Existing generative AI models for radio map generation often fall short in providing fine-grained control and maintaining physical consistency, as radio propagation is heavily influenced by geometry-dependent physical constraints like shadowing and multipath effects.

Method

ControlRadio proposes a controllable diffusion-based framework that redefines radio map generation as a cross-modal conditional synthesis problem. The framework integrates natural language descriptions (textual prompts) with explicit environmental layout constraints to generate physically plausible radio maps. Key components include a Layout-Aware ControlNet, which injects geometric priors like building morphology and transmitter locations into the diffusion process at every denoising step, ensuring strong spatial consistency. A novel Noise Controller replaces the conventional fixed Gaussian initialization with a statistically modulated noise distribution (specifically, μ=-0.1 and σ²=0.001), enhancing generation stability, controllability, and cross-sample consistency. Furthermore, a Decoupled Fine-tuning Strategy adapts pretrained visual-semantic priors from models like CLIP and VAEs to the radio-specific signal domains efficiently, allowing robust performance even with limited annotated data.

Results

ControlRadio demonstrates state-of-the-art accuracy and strong generalization across diverse urban scenarios, consistently outperforming baseline models (RadioUNet, RME-GAN, RadioDiff, RadioDiff-k²) on the RadioMapSeer dataset for both Static Radio Mapping (SRM) and Dynamic Radio Mapping (DRM). It achieves the lowest RMSE (0.0166 for SRM, 0.0180 for DRM) and NMSE, along with the highest PSNR and SSIM. Task-specific diagnostics like LNCErr and BGF1 further confirm its superior performance in maintaining propagation plausibility and structural consistency. The framework reduces radio map generation computation time by more than four orders of magnitude compared to conventional simulation-based methods, achieving sub-second latency per map (0.02-0.44s). Ablation studies confirm the critical roles of the Layout-Aware ControlNet, the Noise Controller, and prompt-driven conditioning in achieving high fidelity and controllability.

Implications

ControlRadio establishes a new paradigm for scalable and controllable wireless environment modeling, with broad implications for next-generation communication systems and data-driven radio sensing. By bridging generative AI with wireless signal modeling, it offers a flexible foundation for tasks like network planning, coverage prediction, anomaly detection, and autonomous wireless operations. The efficiency gains enable the construction of large-scale time-series radio map benchmarks, such as TimeRadioMap, for systematic evaluation of temporal generalization. Future research will focus on evaluating sim-to-real transfer, developing condition-dependent noise priors, and integrating sparse RF observations for measurement-guided synthesis, further enhancing its applicability in practical wireless digital twins.