FOMRadio: Nonstationary Operator Flow Matching for Radio Map Generation
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Generating radio maps from scene geometry and transmitter configurations requires combining long-range propagation with local, scene-dependent interactions. We propose FOMradio, a conditional flow-matching model with a geometry-conditioned nonstationary Fourier operator. Its velocity network couples distant locations spectrally, while an amplitude–phase gate and a residual spatial branch adapt the retained modes and recover structures beyond them. A condition encoder integrates building heights, analytical physical parameters, and transmitter configurations, making the effective kernel. \method transports a multiscale Gaussian random-field source along a straight conditional path, separating spatially structured uncertainty from environment-dependent transport without access to the target radio map at inference. Experiments on RadioMapSeer and U6G demonstrate improved pixelwise accuracy across static and dynamic scenario and variations in cross-configurations, cross-frequency bands, and beam direction.