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

Jul 16, 2026

Angular radio maps model received-power distribution over angle of arrival and are crucial for beam selection and receiver localization in 6G networks. Directly predicting the angular power spectrum from geometry is ill-posed under non-line-of-sight multipath and must generalize to unseen environments. Distortion-minimizing regressors output conditional means that over-smooth spectra and destroy multipath structure. RadioDiff-v2 reframes prediction as a perception-distortion task using a dual-branch 1D diffusion transformer with flow matching to generate the needed angular statistics and enable Bayes-optimal localization.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Abstract

Introduction

RadioDiff-v2: Generative Angular Radio Maps for Multi-Beam Selection and Localization Xiucheng Wang, Junxi Huang, Nan Cheng arXiv :2607 .08045v 1 [ cs .IT] 9 Jul 2026 Abstract—Angular radio maps describe the received-power distribution over the angle of arrival and underpin beam selection and receiver localization in sixth-generation (6G) networks. Predicting the angular power spectrum (APS) from geometry is difficult, because the mapping is ill-posed in non-line-of-sight (NLOS) conditions and must generalize to unseen environments. Distortion-minimizing regressors return the conditional mean, which over-smooths the spectrum and erases the multipath structure that downstream tasks need. We cast the task asa perception-distortion problem and propose RadioDiff-v2, a dual-branch one-dimensional diffusion transformer trained with flow matching. It couples periodic angular encoding, adaptive layer-normalization conditioning, a Fourier angular mixer, and joint velocity and clean-signal heads. A per-metric estimator portfolio reads every deployment quantity from this single model, so that samples carry the distribution, the clean-signal head supplies a regression-grade point estimate, Bayes-optimal rules select beams, and the conditional likelihood localizes the receiver. We prove that a concentrated conditional yields a straight probability-flow trajectory that one step integrates exactly, identifying deterministic transport as the correct inductive bias. On a zero-shot test of 99 environments and one million links, RadioDiff-v2 leads every baseline on every metric, with a 0.39 dB Wasserstein-1 distance, per-bin error below the regression baseline, a 2.43 dB eight-beam NLOS sweep loss, and a 20.6-pixel localization error with four base stations. Code is available at [https://github.com/UNIC-Lab/RadioDiff-v2](https://github.com/UNIC-Lab/RadioDiff-v2). Index Terms—Radio map, angular power spectrum, flow matching, diffusion model, beam selection, localization, 6G. I. INTRODUCTION The angular structure of the wireless channel governs how a base station forms beams and where it places nulls. Sixthgeneration (6G) systems push toward higher carriers and large antenna arrays, where directional transmission is the dominant source of link gain [1], [2] . A station that knows the angular power spectrum (APS) at a candidate receiver location can steer toward the strongest arrival before any pilot exchange. The same angular knowledge underpins beam management, interference avoidance, and device localization in dense cells. A radio map, also called a channel knowledge map, stores such location-dependent channel descriptors over space so that a station can query them offline [3]–[5] . Most radio-map work targets a scalar received power or path loss [6], [7], whereas the angular radio map is the harder and the more useful object This work was supported by the National Key Research and Development Program of China (2024YFB907500) . Xiucheng Wang, Junxi Huang and Nan Cheng are with the State Key Laboratory of ISN and School of Telecommunications Engineering, Xidian University, Xi’an 710071, China (e-mail: {xcwang_1, [24012100067}@stu.xidian.edu.cn](24012100067}@stu.xidian.edu.cn); [dr.nan.cheng@ieee.org](dr.nan.cheng@ieee.org));(Corresponding author: Nan Cheng.) . because it exposes direction rather than a single scalar. Beam selection and angle-based localization both consume the full angular profile rather than a single power value, and this paper therefore studies the prediction of the angular radio map from geometry. This prediction is difficult because multipath propagation governs the mapping from geometry to the angular profile. In a line-of-sight (LOS) link the dominant arrival follows the geometric bearing from receiver to transmitter, so the angular profile is sharply peaked and nearly determined by the receiver position. In a non-line-of-sight (NLOS) link the direct path is blocked, and the received energy arrives throu

FAQs

What problem does RadioDiff-v2 address in angular radio-map prediction?

It addresses the over-smoothing failure of distortion-minimizing regressors under NLOS conditions, which erases multipath structure needed by beam selection and localization.

How does RadioDiff-v2 generate angular radio maps more effectively than standard regression?

It treats angular radio-map construction as a perception-distortion problem and uses a dual-branch 1D diffusion transformer trained with flow matching to generate samples rather than conditional means.

What are the reported performance results of RadioDiff-v2?

On a zero-shot test of 99 environments and one million links, RadioDiff-v2 outperforms all baselines with a 0.39 dB Wasserstein-1 distance, per-bin error below the regression baseline, a 2.43 dB eight-beam NLOS sweep loss improvement, and 20.6-pixel localization error with four base stations.

Jul 16, 2026

Upload by Genevieve