An arXiv preprint reports that RFWM, a physics-guided model for generating three-dimensional radio-frequency (RF) fields, delivered the best reported scores among scene-specific methods on held-out routes in training scenes and retained substantial gains on scenes absent from training. The work asks whether a shared model can transfer propagation knowledge and predict a scene’s changing 3D RF field without target-scene RF measurements.
The benchmark behind the result
The benchmark contained 7,715 synchronized physical–RF sequences from 115 3D environments. Training used 7,045 sequences and 23,139 clips. The in-distribution (ID) test used 293 sequences from 74 held-out routes in 37 scenes, while the out-of-distribution (OOD) test used 377 sequences from 14 scenes absent from training.
The data drew on 3D-FRONT environments, human-walking trajectories generated by DIMOS and spatiotemporal RF fields generated by AutoMS.
RFWM uses a two-stage training strategy with a Friis-guided prior, ControlNet-based physical-to-RF conditioning, six physics-guided regularizers—additional constraints during training—and one-pass generation across receiver heights.
The paper compared RFWM with NeRF2, WRF-GS, RadCloudSplat, RFCanvas and a 3D-retrained Diffusion2. The first four were scene-specific methods, while Diffusion2 used a shared model with receiver height as a condition. The evaluation used four measures: MSE, PSNR, SSIM and LPIPS.
Strong scores on both tests
On the ID test, RFWM reported an MSE of −18.4664 dB, a PSNR of 22.3067 dB, an SSIM of 0.8145 and an LPIPS of 0.0698—the best reported results among the scene-specific methods.
Against the strongest scene-specific result, the reported gains were 5.61 dB in MSE, 7.62 dB in PSNR, 76.9% in LPIPS and 0.1250 in SSIM.
On OOD scenes, RFWM reported an MSE of −10.6225 dB, a PSNR of 11.1471 dB, an SSIM of 0.6427 and an LPIPS of 0.3406. The reported improvements over the best corresponding baseline were 3.86 dB in MSE, 4.18 dB in PSNR, 49.7% in LPIPS and 0.2381 in SSIM.
Physics checks and speed
On OOD predictions, RFWM showed stronger physical compliance than Diffusion2 across six propagation criteria, with the most pronounced reported advantage in far-field attenuation.
In this study, physical compliance was a surrogate based on sampled propagation constraints rather than a direct measurement of the physical world.
Compared with RFWM without physics-guided regularization, the full model reported MSE improvements of 0.66 dB on ID and 0.75 dB on OOD. PSNR improvements were 0.86 dB and 1.03 dB, respectively.
The paper also reports an average rendering time of 0.7697 seconds for one complete 3D RF-field frame, while noting that the model has approximately 8 billion parameters.
What remains untested
The findings are limited to a constructed computational benchmark. Its RF fields were generated by AutoMS, and validation on real measured RF fields or physical deployments was not reported.
Although the OOD test used scenes absent from training, its 14 scenes came from the same benchmark-generation pipeline. That does not establish performance in arbitrary real environments. The ID test held out routes while retaining scenes seen during training, so it was not a scene-level unseen-environment test.
The comparison combines scene-specific fitting methods, shared models and a retrained Diffusion2 setup, so the reported numbers apply to the evaluated configurations.
The reported comparisons were point estimates: no confidence intervals or significance tests were provided, and the ablation did not report repeated-run variability. Rendering time is also dependent on hardware and implementation.
Independent testing on measured RF data and across other environments, motion patterns, wireless configurations and propagation conditions is still needed. The document is an arXiv preprint dated 20 August 2026.
Paper data and sources
Original title: RFWM: Physics-Guided World Model for Dynamic Wireless Radiance Field Generation
Authors: Zijiu Yang, Qianqian Yang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text