Pith. sign in

REVIEW 2 major objections 7 minor 46 references

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

T0 review · 2 major / 7 minor · reviewed 2026-07-10 · glm-5.2

Pith's one-line read One generative model replaces regressors for 6G beam selection and localization

desk verdict Flow matching for angular radio maps: strong engineering, thin theory, self-authored benchmark read the letter →

arxiv 2607.08045 v1 pith:FTFPBYZD submitted 2026-07-09 cs.IT cs.LGeess.SPmath.IT

classification cs.ITcs.LGeess.SPmath.IT
keywords angularpowerspectrumflowmatchingdiffusiontransformerbeamselectionlocalizationradiomapperception-distortiontradeoff6Gnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Predicting how radio power arrives from different angles at a receiver is essential for steering beams and locating devices in 6G networks, but the problem is ill-posed when the direct line of sight is blocked. Existing methods minimize per-angle error, which returns the average of all plausible multipath patterns—a broad blur that points at no real arrival. This paper argues that the task is a perception–distortion trade-off: you can match the average or reproduce the sharp multi-lobe structure, but not both. The authors build a generative model—trained with flow matching on a one-dimensional diffusion transformer—that learns the full conditional distribution of the angular power spectrum given building geometry, rather than collapsing it to its mean. From this single model, different readouts serve different needs: raw samples carry the distribution for diverse beam sweeps, a clean-signal head provides a regression-grade point estimate, Bayes-optimal rules select beams, and the conditional likelihood localizes the receiver via maximum-a-posteriori inference. The key theoretical claim is that because the angular spectrum is nearly deterministic once geometry is fixed, the flow-matching transport follows a straight line that a single integration step solves exactly, making deterministic transport the correct inductive bias over noise-injecting stochastic samplers. On 99 unseen environments and one million links, the model leads every baseline on every metric, achieving a 0.39 dB distributional distance, 0.184 per-bin error, 2.43 dB eight-beam sweep loss in NLOS, and 20.6-pixel localization error with four base stations.

What carries the argument

The machinery is a dual-branch one-dimensional diffusion transformer trained with rectified flow matching. Five components distinguish it: (1) periodic angular positional encoding that respects the circular azimuth axis; (2) adaptive layer normalization that injects geometry into every transformer block; (3) an adaptive Fourier transform angular mixer at the bottleneck that exploits the spectral sparsity of multipath lobes; (4) coupled velocity and clean-signal decoder heads fused by a learnable gate, keeping intermediate samples on the manifold of valid spectra; (5) classifier-free guidance whose scale is selected by Wasserstein-1 distance rather than per-bin error, so the sampler preserves

What would settle it

If a distortion-minimizing regressor equipped with the same clean-signal readout and the same condition encoder matched RadioDiff-v2 on the per-bin metrics, or if the Wasserstein-1 advantage of the sampling readout did not translate into better beam-sweep or localization performance, the central claim that distribution matching is necessary for downstream tasks would be undermined. The paper addresses the first by giving the prior diffusion baseline the same readout (it still trails), and the second by reporting sweep and localization gains directly.

Watch

Extended reading notes

Core claim

The central discovery is that angular radio-map prediction is a perception–distortion problem where distortion-minimizing regression structurally fails by returning the conditional mean—a blur that erases multipath structure—and that a single flow-matching generative model, by learning the conditional distribution rather than its mean, simultaneously supplies distributional fidelity, per-bin point estimates, Bayes-optimal beam selection, and likelihood-based localization, all from one trained network with no per-task retraining. The theoretical backbone is the proof that a concentrated conditional yields a straight-line probability-flow trajectory integrable in one Euler step, which bothjust

Load-bearing premise

The theoretical justification for deterministic flow-matching transport over noise-injecting diffusion rests on the assumption that the angular power spectrum given building geometry is a Dirac mass—perfectly deterministic. The paper itself acknowledges that in NLOS conditions, where the method's advantages are largest, the conditional retains genuine residual uncertainty and multi-modality, meaning the clean one-step integration guarantee applies precisely where the problem

Editorial extensions

If this is right

  • If the conditional distribution of the angular spectrum is learnable from coarse geometry alone, then environment-aware 6G beam management can operate without pilot signals or per-environment calibration, reducing overhead in dense deployments.
  • The generative-MAP localization framework—scoring candidate positions by the flow-matching loss of an observed spectrum under each position's condition—provides a likelihood that discriminative regressors cannot, and multi-station triangulation improves monotonically rather than saturating as fingerprint-based methods do.
  • The per-metric estimator portfolio principle (one model, multiple readouts, each Bayes-optimal for its metric) generalizes to any perception–distortion task where distributional fidelity and point accuracy are both needed, such as channel state information prediction or spatial path-loss mapping.
  • The theoretical result that concentrated conditionals yield straight-line flow trajectories integrable in one step suggests that near-deterministic physical quantities are a natural fit for flow matching over stochastic diffusion, which may apply beyond radio to other geometry-conditioned field prediction problems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The one-step integration result applies cleanly only when the conditional is a Dirac mass (LOS regime); in NLOS where the method's gains are largest, the conditional is genuinely multi-modal, so the theoretical guarantee degrades to an approximation whose tightness is not formally bounded.
  • The Wasserstein-1 sampler selection criterion replaces the conventional NMSE-tuned guidance scale, but the paper does not establish that W1-optimal guidance generalizes across environments with different multipath statistics; it may need re-tuning if deployment environments differ substantially from validation.
  • The multi-station localization requires 15–20% of receivers to be heard by three or more base stations, which constrains practical applicability in sparse deployments or cell-edge scenarios.
  • Extending the conditional density to wideband or elevation-resolved spectra, as the authors propose, would increase the target dimensionality substantially; whether the one-dimensional transformer and Fourier mixer scale efficiently to 2D angular-frequency or 3D angular-elevation-frequency tensors remains an open question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. The paper proposes RadioDiff-v2, a dual-branch one-dimensional diffusion transformer trained with flow matching (rectified flow) to predict angular power spectra (APS) from building geometry. The central insight is that APS prediction is a perception-distortion problem: distortion-minimizing regressors return the conditional mean, which over-smooths multipath structure in NLOS conditions. The model matches the conditional distribution p(x0|c) rather than its mean, enabling a per-metric estimator portfolio that serves distributional fidelity (sampling readout), per-bin accuracy (clean-signal head + posterior mean blend), Bayes-optimal beam selection (greedy submodular sweep), and generative-MAP localization (flow-matching score as likelihood). The backbone combines periodic angular encoding, adaLN-zero conditioning, an AFT angular mixer, and coupled velocity/clean-signal heads with a learnable fusion gate. A theoretical result (Proposition 1) shows that under a Dirac conditional, the flow-matching ODE follows a straight-line trajectory integrable in one Euler step. Experiments span 99 environments with a zero-shot protocol, comparing against four baselines.

Significance. The reframing of angular radio-map prediction as a perception-distortion problem is well-motivated and the per-metric estimator portfolio is a principled contribution: one trained model serving distributional, regression, beam-selection, and localization readouts is architecturally elegant. The code release and reproducible zero-shot protocol (79/20 environment split, identical links for all methods) are commendable. The generative-MAP localization approach—scoring candidate positions by flow-matching loss to obtain a Bayesian likelihood that regressors cannot provide—is a genuinely novel capability. The experimental gains are substantial, particularly the 13x improvement in W1 (0.39 vs 1.97 dB) and the 8-beam NLOS sweep loss (2.43 vs 4.60 dB). Proposition 1 is correct as stated, though its scope requires clarification (see major comments).

major comments (2)
  1. §IV-D, Proposition 1 and Assumption 1: The theoretical justification for deterministic ODE transport over noise-injecting SDEs rests on Assumption 1 (Dirac conditional p(x0|c) = δ(x0 − μ(c))). The paper itself acknowledges in §III that 'the NLOS APS retains a small but genuine residual uncertainty' and admits 'limited multi-modality for NLOS links.' However, the empirical advantages are concentrated in NLOS (Table II: LOS single-beam loss is 0.02 dB for RadioDiff-v2 vs 0.00 dB for COST231—near-optimal for all methods—while NLOS sweep loss shows the largest gains). The theory thus provides formal support for the easy regime while the hard regime relies on empirical validation. This is not an internal inconsistency, but it weakens the claim that Proposition 1 'identifies deterministic transport as the correct inductive bias' (Abstract). The authors should either (a) soften the claim to an
  2. Abstract and Table II: The headline claim that RadioDiff-v2 'leads every baseline on every metric' requires qualification. The sampling readout (Table II, row 4) has NMSE = 0.351, worse than all baselines (RME-GAN 0.217, MS-Areg 0.199, RadioDiff 0.312). The portfolio readout (row 5) achieves NMSE = 0.184, but this is a different estimator using the clean-signal head and posterior-mean blending with hyperparameters selected per-metric on validation data. The claim is accurate for the portfolio, but the paper should clarify that no single readout dominates on all metrics—the portfolio's advantage comes from per-metric estimator selection, which is a design choice rather than a property of the model alone.
minor comments (7)
  1. The Map2APS benchmark [40] used for evaluation is co-authored by the paper's authors. This is disclosed in the reference but a brief note in §V-B acknowledging this relationship would improve transparency.
  2. Table I lists Pmin, Pmax as fixed dynamic-range limits applied across the dataset, but their values are not reported. These affect the dB normalization in Eq. (1) and should be stated.
  3. §IV-C, Eq. (12): The W1-based guidance-scale selection is an inference-time contribution, but the computational cost of evaluating W1 across candidate w values on validation data is not reported. A brief note on this overhead would be helpful.
  4. Fig. 1 is dense and the sub-panels (a)-(f) contain substantial detail. Consider splitting into multiple figures or enlarging individual panels, particularly (c) and (d), to improve readability.
  5. §IV-E, Eq. (15): The clean-signal readout uses antithetic draws with tr 'close to one' and R draws, but the specific values of tr and R used are not reported. These should be stated for reproducibility.
  6. The paper cites several self-references [7, 11, 19, 20, 21, 22, 23, 27] from the same group. While each appears relevant, the authors should ensure novelty is clearly distinguished from prior RadioDiff family work, particularly [7] and [11].
  7. §V-F: The single-station generative-MAP error (62.6 px) is reported on the full test set while the multi-station fusion (Table IV, lower block) uses 577 fixed queries. The different evaluation subsets should be clarified earlier to avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: theory is self-contained, self-citations are contextual not load-bearing

full rationale

The paper's central theoretical contribution (Proposition 1, §IV-D) is derived entirely from first principles: it combines Assumption 1 (Dirac conditional) and Assumption 2 (optimal velocity) to show that the flow-matching ODE trajectory is a straight line with zero discretization error. The proof substitutes the Dirac assumption into the rectified-flow interpolant (Eq. 6), derives the constant target velocity (Eq. 7), and concludes the ODE has a constant right-hand side. No step in this derivation depends on any self-cited result. The self-citations present ([7, 11, 19-23, 27, 40]) are contextual: [40] provides the evaluation benchmark/dataset, [7] is the prior diffusion baseline being compared against, and others extend the RadioDiff family. None of these citations are invoked to justify the core theoretical claim or to establish uniqueness that would foreclose alternatives. The benchmark [40] (Map2APS, co-authored by present authors Huang, Wang, Cheng) is used for empirical evaluation against external baselines (COST231, RME-GAN, MS-Areg), not to derive a theoretical result. The perception-distortion framing (citing [12], Blau & Michaeli) is an externally established framework. The estimator portfolio (§IV-E) uses standard decision-theoretic rules (MMSE averaging, greedy submodular beam selection) that are independently derived. While the reader correctly notes that Assumption 1 applies cleanly only to LOS while empirical gains concentrate in NLOS, this is a correctness/applicability concern, not circularity: the theory is not defined in terms of its own outputs, and no prediction reduces to a fitted input by construction. The paper is self-contained against external benchmarks with independently reproducible code. Score 2 reflects the presence of multiple self-citations that, while not load-bearing for the central claim, create an ecosystem of related work that could warrant scrutiny on independence of evaluation.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

No new physical entities, particles, or forces are introduced. The model architecture components (AFT mixer, adaLN, dual heads) are engineering constructs, not new physical postulates.

free parameters (5)
  • Guidance scale w = w* = 1 (W1-optimal)
    Selected on validation data by minimizing Wasserstein-1 distance (Eq. 12). Controls the perception-distortion tradeoff at inference time.
  • Fusion gate α = learnable, value not reported
    Learnable parameter controlling the blend between velocity-implied and direct clean-signal estimates (Eq. 10).
  • Portfolio hyperparameters (tr, R, K, λ) = selected on validation, specific values not all reported
    Control the clean-signal readout (tr, R), posterior mean averaging (K), and per-metric blend (λ). All selected on held-out validation links and frozen before testing.
  • Pmin, Pmax (dynamic range limits) = fixed across dataset, values not reported
    Used in the dB normalization of the APS target (Eq. 1). Affect the target distribution.
  • ε (logarithm bound constant) = small constant, value not reported
    Bounds the logarithm in deep nulls in the APS normalization (Eq. 1).
assumptions (3)
  • ad hoc to paper Assumption 1: The conditional law of the APS given the condition is a Dirac mass p(x0|c) = δ(x0 − μ(c)).
    Invoked in §IV-D to prove Proposition 1. The paper itself acknowledges this is only approached in the LOS regime and that NLOS retains residual uncertainty, making this an idealization that does not hold in the regime where the method's advantages are largest.
  • domain assumption Assumption 2: The network attains the minimizer of the flow-matching objective, so vϕ equals the conditional expectation of the target velocity.
    Invoked in §IV-D. Standard optimal-regression assumption in flow matching theory, but assumes perfect training convergence which is never achieved in practice.
  • domain assumption Per-station observations are conditionally independent given the position.
    Invoked in §V-A to justify additive log-likelihoods across base stations (Eq. 19). This is a standard assumption in multi-station localization but may not hold exactly due to shared scatterers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RadioDiff-v2: Generative Angular Radio Maps for Multi-Beam Selection and Localization." pith.science (2026). https://pith.science/paper/FTFPBYZD

@misc{pith2026260708045,
  author       = {Pith},
  title        = {Pith review of: RadioDiff-v2: Generative Angular Radio Maps for Multi-Beam Selection and Localization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTFPBYZD}},
  note         = {Machine review of arXiv:2607.08045}
}
read the original 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 as a 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.

Figures

Figures reproduced from arXiv: 2607.08045 by the authors.

Figure 1
Figure 1. Overview of RadioDiff-v2. (a) The condition is built from the receiver-centred polar occupancy, the geometry, and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of predicted angular power spectra on NLOS links, shown in the dB domain with each curve [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Multi-base-station likelihood maps over space for one [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages

  1. [40]

    Map2APS: A Physically Grounded Benchmark for Direct Angle Power Spectrum Prediction from Urban Geometry

    J. Huang, X. Wang, N. Cheng, K. Wang, R. Sun, and Z. Yin, “Map2aps: A physically grounded benchmark for direct angle power spectrum prediction from urban geometry,”arXiv preprint arXiv:2605.14989, 2026

  2. [7]

    Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,

    X. Wang, K. Tao, N. Cheng, Z. Yin, Z. Li, Y . Zhang, and X. Shen, “Radiodiff: An effective generative diffusion model for sampling-free dynamic radio map construction,”IEEE Trans. Cognit. Commun. Net- working, vol. 11, no. 2, pp. 738–750, 2025

  3. [1]

    6G omni-scenario on-demand services provisioning: vision, technology and prospect(in chinese),

    N. Cheng, F. Chen, W. Chen, Z. Cheng, Q. Yang, C. Li, and X. Shen, “6G omni-scenario on-demand services provisioning: vision, technology and prospect(in chinese),”Sci Sin Inform, vol. 54, no. 5, pp. 1025–1054, 2024

  4. [2]

    5g channel model for bands up to 100 ghz,

    N. Docomoet al., “5g channel model for bands up to 100 ghz,” Technical report, Tech. Rep., 2016

  5. [3]

    RadioNet: Robust deep-learning based radio fingerprinting,

    H. Li, K. Gupta, C. Wang, N. Ghose, and B. Wang, “RadioNet: Robust deep-learning based radio fingerprinting,” inProceedings of the 2022 IEEE Conference on Communications and Network Security (CNS), 2022, pp. 190–198

  6. [4]

    Generative ai on spectrumnet: An open benchmark of multiband 3d radio maps,

    S. Zhang, S. Jiang, W. Lin, Z. Fang, K. Liu, H. Zhang, and K. Chen, “Generative ai on spectrumnet: An open benchmark of multiband 3d radio maps,”IEEE Trans. Cognit. Commun. Networking, vol. 11, no. 2, pp. 886–901, 2025. 12

  7. [5]

    Generative ai for deep reinforcement learning: Framework, analysis, and use cases,

    G. Sun, W. Xie, D. Niyato, F. Mei, J. Kang, H. Du, and S. Mao, “Generative ai for deep reinforcement learning: Framework, analysis, and use cases,”IEEE Wireless Commun., vol. 32, no. 3, pp. 186–195, 2025

  8. [6]

    RadioUNet: Fast radio map estimation with convolutional neural networks,

    R. Levie, Ç. Yapar, G. Kutyniok, and G. Caire, “RadioUNet: Fast radio map estimation with convolutional neural networks,”IEEE Trans. Wireless Commun., vol. 20, no. 6, pp. 4001–4015, 2021

Show all 46 references
  1. [8]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. ...

  2. [9]

    RME-GAN: A learning framework for radio map estimation based on conditional generative adversarial network,

    S. Zhang, A. Wijesinghe, and Z. Ding, “RME-GAN: A learning framework for radio map estimation based on conditional generative adversarial network,”IEEE Internet Things J., vol. 10, no. 20, pp. 18 016–18 027, 2023

  3. [10]

    Generative adversarial networks: An overview,

    A. Creswell, T. White, V . Dumoulin, K. Arulkumaran, B. Sengupta, and A. A. Bharath, “Generative adversarial networks: An overview,”IEEE Signal Process. Mag., vol. 35, no. 1, pp. 53–65, 2018

  4. [11]

    Radiodiff-k2: Helmholtz equation informed generative diffusion model for multi-path aware radio map construction,

    X. Wang, Q. Zhang, N. Cheng, R. Sun, Z. Li, S. Cui, and X. Shen, “Radiodiff-k2: Helmholtz equation informed generative diffusion model for multi-path aware radio map construction,”IEEE J. Sel. Areas Commun., vol. 44, pp. 2318–2333, 2026

  5. [12]

    The perception-distortion tradeoff,

    Y . Blau and T. Michaeli, “The perception-distortion tradeoff,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 6228–6237

  6. [13]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inProc. Int. Conf. Learn. Represent. (ICLR), 2023

  7. [14]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” inProc. Int. Conf. Learn. Represent. (ICLR), 2023

  8. [15]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023, pp. 4172–4182

  9. [16]

    Locunet: Fast urban positioning using radio maps and deep learning,

    Ç. Yapar, R. Levie, G. Kutyniok, and G. Caire, “Locunet: Fast urban positioning using radio maps and deep learning,” inICASSP 2022- 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 4063–4067

  10. [17]

    Indoor radio map construction and localization with deep gaussian processes,

    X. Wang, X. Wang, S. Mao, J. Zhang, S. C. Periaswamy, and J. Patton, “Indoor radio map construction and localization with deep gaussian processes,”IEEE Internet Things J., vol. 7, no. 11, pp. 11 238–11 249, 2020

  11. [18]

    Toward environment-aware 6G communications via channel knowledge map,

    Y . Zeng and X. Xu, “Toward environment-aware 6G communications via channel knowledge map,”IEEE Wireless Commun., vol. 28, no. 3, pp. 84–91, 2021

  12. [19]

    A tutorial on learning-based radio map construction: Data, paradigms, and physics-awareness,

    X. Wang, Y . Pan, N. Cheng, Ç. Yapar, R. Sun, Z. Yin, C. Zhou, W. Xu, Y . Zhang, J. Zhang, S. Cui, and X. Shen, “A tutorial on learning-based radio map construction: Data, paradigms, and physics-awareness,”arXiv preprint arXiv:2603.17499, 2026

  13. [20]

    Radiodiff-inverse: Diffusion enhanced bayesian inverse estimation for isac radio map construction,

    X. Wang, Z. Fang, N. Cheng, R. Sun, H. Zhou, Z. Su, Z. Li, and X. Shen, “Radiodiff-inverse: Diffusion enhanced bayesian inverse estimation for isac radio map construction,”IEEE Trans. Wireless Commun., vol. 25, pp. 14 611–14 626, 2026

  14. [21]

    iRadioDiff: Physics-informed diffusion model for indoor radio map construction and localization,

    X. Wang, T. Yuan, Y . Cao, N. Cheng, R. Sun, and W. Zhuang, “iRadioDiff: Physics-informed diffusion model for indoor radio map construction and localization,”arXiv preprint arXiv:2511.20015, 2025

  15. [22]

    RadioDiff-FS: Physics-informed manifold alignment in few-shot diffusion models for high-fidelity radio map construction,

    X. Wang, Z. Guo, and N. Cheng, “RadioDiff-FS: Physics-informed manifold alignment in few-shot diffusion models for high-fidelity radio map construction,”arXiv preprint arXiv:2603.18865, 2026

  16. [23]

    RadioDiff-flux: Efficient radio map construction via generative denoise diffusion model trajectory midpoint reuse,

    X. Wang, P. Zheng, H. Jia, N. Cheng, R. Sun, C. Zhou, and X. Shen, “RadioDiff-flux: Efficient radio map construction via generative denoise diffusion model trajectory midpoint reuse,”IEEE Trans. Cognit. Com- mun. Networking, vol. 12, pp. 4882–4895, 2026

  17. [24]

    Radiogat: A joint model-based and data-driven framework for multi-band radiomap reconstruction via graph attention networks,

    X. Li, S. Zhang, H. Li, X. Li, L. Xu, H. Xu, H. Mei, G. Zhu, N. Qi, and M. Xiao, “Radiogat: A joint model-based and data-driven framework for multi-band radiomap reconstruction via graph attention networks,”IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 17 777–17 792, 2024

  18. [25]

    Radio map estimation–an open dataset with directive transmitter antennas and initial experiments,

    F. Jaensch, G. Caire, and B. Demir, “Radio map estimation–an open dataset with directive transmitter antennas and initial experiments,”arXiv preprint arXiv:2402.00878, 2024

  19. [26]

    Ckmimagenet: A comprehensive dataset to enable channel knowledge map construction via computer vision,

    D. Wu, Z. Wu, Y . Qiu, S. Fu, and Y . Zeng, “Ckmimagenet: A comprehensive dataset to enable channel knowledge map construction via computer vision,” in2024 IEEE/CIC International Conference on Communications in China (ICCC Workshops). IEEE, 2024, pp. 114– 119

  20. [27]

    RadioDiff-3D: A 3D×3D radio map dataset and generative diffusion based benchmark for 6G environment-aware communication,

    X. Wang, Q. Zhang, N. Cheng, J. Chen, Z. Zhang, Z. Li, S. Cui, and X. Shen, “RadioDiff-3D: A 3D×3D radio map dataset and generative diffusion based benchmark for 6G environment-aware communication,” IEEE Trans. Netw. Sci. Eng., vol. 13, pp. 3773–3789, 2026

  21. [28]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems (NeurIPS), vol. 33, pp. 6840–6851, 2020

  22. [29]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inProc. Int. Conf. Learn. Represent. (ICLR), 2021, pp. 1–12

  23. [30]

    Score-based generative modeling through stochastic differ- ential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,” inProc. Int. Conf. Learn. Represent. (ICLR), 2021

  24. [31]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 10 674–10 685

  25. [32]

    Glide: Towards photorealistic image gener- ation and editing with text-guided diffusion models,

    A. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. McGrew, I. Sutskever, and M. Chen, “Glide: Towards photorealistic image gener- ation and editing with text-guided diffusion models,” inProc. Int. Conf. Mach. Learn. (ICML), ser. PMLR, vol. 162, 2022, pp. 16 784–16 804

  26. [33]

    Decoupled diffusion models with explicit transition probability,

    Y . Huang, Z. Qin, X. Liu, and K. Xu, “Decoupled diffusion models with explicit transition probability,”arXiv preprint arXiv:2306.13720, 2023

  27. [34]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems (NeurIPS), vol. 30, 2017

  28. [35]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,”IEEE Trans. Pattern Anal. Mach, vol. 45, no. 9, pp. 10 850–10 869, 2023

  29. [36]

    Exploiting radio fingerprints for simultaneous localization and mapping,

    R. Liu, B. P. L. Lau, K. Ismail, A. Chathuranga, C. Yuen, S. X. Yang, Y . L. Guan, S. Mao, and U.-X. Tan, “Exploiting radio fingerprints for simultaneous localization and mapping,”IEEE Pervasive Comput., vol. 22, no. 3, pp. 38–46, 2023

  30. [37]

    Confidence-regulated generative diffusion models for reliable ai agent migration in vehicular metaverses,

    Y . Kang, J. Kang, J. Wen, T. Zhang, Z. Yang, D. Niyato, and Y . Zhang, “Confidence-regulated generative diffusion models for reliable ai agent migration in vehicular metaverses,”arXiv preprint arXiv:2505.12710, 2025

  31. [38]

    Electromagnetic scattering laws in weyl systems,

    M. Zhou, L. Ying, L. Lu, L. Shi, J. Zi, and Z. Yu, “Electromagnetic scattering laws in weyl systems,”Nature Commun., vol. 8, no. 1, p. 1388, 2017

  32. [39]

    Ray techniques in electromagnetics,

    G. A. Deschamps, “Ray techniques in electromagnetics,”Proc. IEEE, vol. 60, no. 9, pp. 1022–1035, 1972

  33. [41]

    Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,

    A. Alkhateeb, “Deepmimo: A generic deep learning dataset for millimeter wave and massive mimo applications,”arXiv preprint arXiv:1902.06435, 2019

  34. [42]

    Dominant path prediction model for urban scenarios,

    R. Wahl, G. Wölfle, P. Wertz, P. Wildbolz, and F. Landstorfer, “Dominant path prediction model for urban scenarios,” in14th IST mobile and wireless communications summit, 2005, pp. 1–5

  35. [43]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” 2022

  36. [44]

    Computational optimal transport: With appli- cations to data science,

    G. Peyré and M. Cuturi, “Computational optimal transport: With appli- cations to data science,”Found. Trends Mach. Learn., vol. 11, no. 5-6, pp. 355–607, 2019

  37. [45]

    Digital mobile radio towards future gen- eration systems—COST action 231 final report,

    E. Damosso and L. M. Correia, “Digital mobile radio towards future gen- eration systems—COST action 231 final report,” European Commission, Brussels, Belgium, Tech. Rep. EUR 18957, 1999

  38. [46]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Trans. Image Processing, vol. 13, no. 4, pp. 600–612, 2004

Pith tools

Reviewed July 10, 2026 · model on record in the stance chip above.