RadioDiff-v2: Generative Angular Radio Maps for Multi-Beam Selection and Localization

Predicting the angular power spectrum (APS) from building geometry is a fundamental challenge for 6G networks, as it enables precise beam selection and receiver localization. Traditional regression models, which minimize per-pixel reconstruction loss, tend to produce the conditional mean of the angular profile. This results in over-smoothed maps that erase the distinct multipath lobes necessary for downstream tasks. This paper investigates whether a generative approach can better capture the sharp, multi-modal nature of the APS in non-line-of-sight (NLOS) conditions.

The authors frame the APS prediction as a perception-distortion problem and introduce RadioDiff-v2, a dual-branch one-dimensional diffusion transformer. Unlike stochastic diffusion models that inject noise during sampling, RadioDiff-v2 is trained with flow matching to learn a deterministic ordinary differential equation (ODE). This approach transports a noise prior to the data along a straight-line trajectory, which the authors prove is the optimal inductive bias for the nearly deterministic mapping between geometry and angular power. The model includes a per-metric estimator portfolio that extracts diverse samples for beam selection, a regression-grade point estimate for pointing, and a conditional likelihood for localization.

RadioDiff-v2 significantly outperforms existing baselines across all tested metrics in a zero-shot evaluation spanning 99 environments and one million links. It achieves a 0.39 dB Wasserstein-1 distance (compared to 1.97 dB for prior diffusion baselines) and maintains a lower per-bin error than standard regressors. Furthermore, the model's ability to provide a conditional likelihood enables generative-MAP localization, achieving a 20.6-pixel error with four base stations, whereas regressor-based methods saturate at much higher error rates. The deterministic transport mechanism successfully preserves the dynamic range and angular separation of multipath components, which are critical for reliable beamforming.