Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

BS-1-to-N: Diffusion-Based Environment-Aware Cross-BS Channel Knowledge Map Generation for Cell-Free Networks

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a generative diffusion model, trained only on the channel knowledge maps and locations of existing base stations in an area, can synthesize accurate channel knowledge maps for any number of new base-station…

desk verdict A plausible diffusion-based cross-BS CKM generator with a novel variable-size attention architecture, but the feasibility argument is heuristic and the experiments are less convincing than the text implies. read the letter →

arxiv 2507.23236 v1 pith:2ZQQ6CKW submitted 2025-07-31 eess.SP eess.IV

Zhuoyin Dai , Di Wu , Yong Zeng , Xiaoli Xu , Xinyi Wang , Zesong Fei This is my paper · ORCID
classification eess.SPeess.IV
keywords channelknowledgemapcross-BSCKMinferencecell-freenetworkslatentdiffusionmodelconditionalgenerationbasestationlocationembeddingattentionmechanismdeploymentoptimization
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

The paper tries to establish that a base station's channel knowledge map (CKM), a grid of location-tagged channel values such as channel gain, can be generated for any new base-station location from the CKMs and locations of existing base stations, with no measurement or ray tracing at the new site and no physical environment map. The plausibility argument is that all base stations in one area share the same wireless environment, so each CKM is a different view of that environment. The authors build BS-1-to-N, a conditional latent-diffusion model that takes any number of source CKMs and locations and returns CKMs for any number of target locations, using polar-coordinate base-station location embedding and cross- and self-attention. On city-scale radio maps, it produces target CKMs that preserve shadowing and blockage structure and improves as more source CKMs are supplied. If the claim holds, constructing and updating CKMs in dense cell-free networks would no longer require visiting every base station.

What carries the argument

The load-bearing mechanism is the learned composition $h=f_c\circ f_e$: the network maps a set of source CKM-location pairs directly to a target CKM instead of explicitly reconstructing the wireless environment. The named components are the BS location embedding (BSLE), which encodes each BS position in polar coordinates and applies a rotary matrix so that attention dot-products depend only on relative BS locations, and the inter-CKM attention blocks, where cross-attention aligns target CKM tokens with source CKM tokens and self-attention enforces consistency among multiple target CKMs. These components sit inside a latent diffusion backbone, so generation proceeds by iterative denoising in a compressed latent space conditioned on the source tokens.

What would settle it

Find two physical environments that produce identical channel knowledge maps at all source BS locations but different maps at a target BS location; if such a pair exists, no deterministic function can infer the target map from source maps alone, so any model that succeeds on one environment must fail on the other.

Watch

Extended reading notes

Core claim

The paper's central claim is that the function $f:(\{M_i,l_i\}_{i=1}^I, l_0)\mapsto M_0$ is learnable as a conditional generative model $p(M_0\mid \{M_i,l_i\}, l_0)$, and that a diffusion-based network can implement it for arbitrary numbers of source and target BSs without reconstructing the physical environment. Feasibility is argued through the decomposition $[M_i]_{j,k}=f_c(E,l^u_{j,k};l_i)$, where $E$ is the shared abstract wireless environment: sufficiently many source CKMs can in principle reconstruct $E$, and composing that reconstruction with $f_c$ yields the target CKM, so an explicit environment representation can be bypassed. BS-1-to-N realizes this by running latent diffusion conditioned on source CKM tokens, embedding each BS location through a rotary polar-coordinate scheme so attention depends only on relative BS locations, and using cross-attention between source and target BSs plus self-attention among target BSs. The experiments show that, given five to ten source CKMs and no environment map, the inferred target CKMs retain building-blocking and attenuation structure, and the deployment study shows the inferred maps can guide selection of new BS locations.

Load-bearing premise

A finite set of source base-station maps must carry enough information about the shared wireless environment that the target base-station map is a well-defined function of those inputs, independent of which environment produced them.

Editorial extensions

If this is right

  • For a newly deployed base station, its channel knowledge map can be synthesized from the maps of neighboring base stations and their locations, skipping per-site measurement or ray tracing.
  • The approach handles a flexible count of sources and targets in one model, so an operator can add or remove cooperating base stations without redesigning a fixed-input architecture.
  • Because inference cost grows slowly with the number of target base stations, generating maps for many candidate locations at once becomes practical, which is exactly what site-selection needs.
  • Accuracy increases with the number of source CKMs, with diminishing marginal gains, so operators can trade source count against inference fidelity.
  • The deployment-optimization study shows inferred CKMs can rank candidate base-station locations by coverage of target regions under a channel-gain threshold.

Reading between the lines

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

  • Inference: the conditioning mechanism is not tied to channel gain; if training maps for delay, angle-of-arrival, or angle-of-departure were available, the same location-embedding plus attention structure could generate those CKMs across BSs as well.
  • Inference: because the environment is learned only implicitly, performance should depend on source-BS geometry; a useful stress test would hold out validation environments where all source BSs lie on one side of the target and measure error growth.
  • Inference: the same model could serve as a fast re-planning engine when the environment changes, updating all BS CKMs from a small set of re-measured source maps, though the paper does not test dynamic environments.
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

3 major / 6 minor

Summary. The paper proposes BS-1-to-N, a latent diffusion model for cross-BS channel knowledge map (CKM) inference in cell-free networks. Given the CKMs and locations of an arbitrary number of source BSs, the model infers CKMs for arbitrary target BS locations without using the physical environment map. The architecture adapts Stable Diffusion v1.5 and the multi-view-synthesis method EscherNet by introducing a polar-coordinate BS location embedding (BSLE) with rotary position embedding, a self-attention mechanism among target BSs, and a cross-attention mechanism between source and target BSs. The method is trained and evaluated on the RadioMapSeer dataset, with quantitative metrics (RMSE, SSIM, LPIPS, PSNR), comparisons to a fixed-UNet baseline and a distance-based weighting scheme, and a case study on BS deployment optimization.

Significance. If the approach is empirically sound, it addresses a real scalability problem: constructing CKMs for each distributed BS via measurement or ray tracing is costly, and cross-BS inference from existing CKMs avoids per-BS traversal. The paper's architecture is a thoughtful adaptation of multi-view synthesis and latent diffusion to CKM inference, and it targets a setting (arbitrary numbers of source and target BSs) not handled by prior fixed-input UNet methods. The paper ships algorithmic descriptions (Algorithms 1–3), uses a public dataset, evaluates on held-out environments, and reports a deployment-optimization use case, which are concrete strengths. The main weaknesses are that the theoretical feasibility argument in Section III-A is a composition of assumed functions rather than a proof of identifiability, and that the central benchmark comparison is performed under unequal training conditions.

major comments (3)
  1. [Section III-A, Eqs. (2)–(5)] The feasibility analysis does not establish that the cross-BS CKM inference function h in Eq. (5) is well-defined. The argument presupposes the existence of an environment reconstruction map fe in Eq. (4) that uniquely identifies the abstract wireless environment E from a finite set of source CKMs {Mi, li}. The paper states this is 'theoretically possible' with sufficient BSs, but no identifiability condition or injectivity proof is given for the forward map E → {fc(E,·; li)}. If two different environments produce identical source CKMs but different target CKMs, then the target CKM is not a deterministic function of the inputs, and the conditional distribution p(M0 | {Mi,li}, l0) in Eq. (6) has no stable ground truth independent of the architecture. This is a load-bearing issue for the central claim: it justifies the well-posedness of the learning problem itself. The paper should either provide a formal identifiability argument for the class of environments and channel-gain maps considered, or add an empirical test that quantifies target-CKM variability across environments that produce near-identical source CKMs.
  2. [Section V-B.2, Table IV and accompanying text] The comparison with the UNet baseline is not on equal footing. The paper states that 'UNet was trained with CKMs of 79 source BSs as inputs, while BS-1-to-N used only 10 source BSs during the training.' Yet Table IV shows that the UNet baseline outperforms BS-1-to-N on all four reported metrics (RMSE 0.0354 vs 0.0649, SSIM 0.9449 vs 0.8980, PSNR 29.85 vs 24.07, LPIPS 0.0672 vs 0.1342). The qualitative discussion in Fig. 7 is not a substitute for a fair quantitative comparison. To support the claim that BS-1-to-N is effective for flexible source-BS counts, the UNet baseline should be trained and evaluated with the same number of source BSs (e.g., 5 or 10), or the experiment should include a matched setting. As reported, the numbers undermine the conclusion that BS-1-to-N is competitive with state-of-the-art cross-BS inference.
  3. [Section V-B, Tables III–IV and Figures 6–8] All reported metrics appear to be from a single evaluation run, with no error bars, confidence intervals, or significance tests. This is particularly problematic for Table III, where the difference between 5 and 10 source BSs is small (RMSE 0.0649 vs 0.0615; SSIM 0.8980 vs 0.9025) and could be within run-to-run variability. Similarly, the inference-time comparison in Fig. 8 and the deployment-optimization results in Fig. 9 are single illustrative trials. The paper should report mean and variance over multiple training seeds and/or evaluation runs, or at least justify why single-run results are sufficient to support the quantitative claims about scaling with the number of source BSs.
minor comments (6)
  1. [Section V-A, Table II] The 'Max training step 2000' value is surprisingly low for training a latent diffusion model on a dataset of this size; please clarify whether this is 2000 iterations, 2000 epochs, or a typo, and discuss convergence.
  2. [Section V-A] The text says 'the building region is removed in the metric calculation,' but it is not explained how the building region is identified given that the physical environment map is not used during inference; please specify the procedure and whether it uses ground-truth environment information.
  3. [Section V-B.2, Eq. (34)] The UNet baseline formula in Eq. (34) uses a weighting parameter ω, but the value of ω is not stated; please report the exact configuration used to generate the results in Table IV.
  4. [Section V-B.1 and Fig. 6] The caption and text for Fig. 6 state that the inferred CKMs are compared with ground truths, but the figure itself does not indicate which rows are which; adding clear labels and a color scale would improve readability.
  5. [Section IV-A] The architecture description would benefit from explicitly listing which modules are frozen vs trainable (e.g., the VAE encoder/decoder versus the denoising UNet and image encoder τ), since Fig. 5 shows 'Frozen module' and 'Trainable module' without a clear key for all components.
  6. [Section III-A] In Eq. (5), the notation fc(fe({Mi, li}), lu_j,k; l0) treats fe as returning a full environment E, but fe in Eq. (4) is defined as producing an estimate E-hat; using distinct notation for the estimate would avoid confusion between the true and reconstructed environment.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical evaluation uses disjoint training and validation environments, and the theoretical feasibility argument is an unproven premise rather than a fitted input disguised as a prediction.

full rationale

The paper's central claim—that a conditional diffusion model can infer target-BS CKMs from source-BS CKMs and locations—is not circular. Training and validation are performed on disjoint physical environments (630 vs. 71 RadioMapSeer environments), and the source and target BS sets are disjoint; target CKMs are never used to fit the BSLE, attention weights, or diffusion parameters. The feasibility argument in Section III-A is a conditional existence statement: Eq. (2) defines the forward physical map, Eq. (4) assumes an inverse map fe without proof, and Eq. (5) composes fe and fc. This is tautological as a proof of feasibility, but it is not a fitted input disguised as a prediction; the learned model's success is established against unseen validation environments and external benchmarks (UNet, distance-based weighting). Self-citations ([7], [16], [22], [23]) supply background and baselines, but the load-bearing empirical comparison is external and the architecture components (Stable Diffusion, RoFormer, Eschernet) are independently published. The absence of an injectivity or identifiability proof for fe is a correctness risk about ground-truth well-posedness, not a circularity in the derivation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central mapping is learned, not derived, so the main free parameters are the network weights and the hand-chosen training recipe. The feasibility argument assumes rather than proves that the wireless environment can be recovered from source CKMs. No new physical entities are introduced.

free parameters (3)
  • Learned weights of the denoising UNet, image encoder tau, and attention projection matrices = unknown
    The central mapping h in Eq. (5) is implemented entirely by these trained weights; no closed form or released weights are provided.
  • Training recipe (|I_e|=10, |I_t|=5, max steps 2000, batch 10, inference steps 50) = |I_e|=10, |I_t|=5, steps=2000, batch=10, inference=50
    Hand-chosen hyperparameters in Table II; no sensitivity analysis or ablations are reported.
  • Distance-weighting decay gamma in baseline = 0.1
    Used only in benchmark Eq. (35); arbitrary and not part of BS-1-to-N.
assumptions (6)
  • domain assumption There exist functions fc and fe such that channel knowledge equals fc(E, l_u; l_i) and the environment E can be reconstructed as fe({M_i,l_i}).
    Section III-A, Eqs. (2) and (4): the feasibility of cross-BS inference assumes the shared environment E is identifiable from source CKMs; no uniqueness or sufficiency proof is given.
  • domain assumption CKMs of different BSs are multi-view projections of one shared wireless environment.
    Section III-A and Fig. 3: this analogy is the conceptual basis for using view synthesis; it is asserted, not derived.
  • domain assumption The RadioMapSeer dominant path model dataset is representative of real CKM data.
    Section V-A: all experiments use simulated CGMs from 701 city maps; no real measurements validate transfer.
  • standard math Rotary position embedding makes attention depend only on relative BS locations.
    Section III-B, Eq. (12): rotation matrix dot-product invariance, from RoFormer; this is a standard property.
  • domain assumption Stable Diffusion v1.5 autoencoder and UNet priors transfer from natural images to CKM images.
    Section IV-B and V-A: pretrained VAE and LDM are used without adaptation analysis for radio maps.
  • domain assumption Validation environments are i.i.d. samples from the same distribution as training environments.
    Section V-A: 630/71 split of RadioMapSeer; no out-of-distribution or temporal generalization test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BS-1-to-N: Diffusion-Based Environment-Aware Cross-BS Channel Knowledge Map Generation for Cell-Free Networks." pith.science (2026). https://pith.science/paper/2ZQQ6CKW

@misc{pith2026250723236,
  author       = {Pith},
  title        = {Pith review of: BS-1-to-N: Diffusion-Based Environment-Aware Cross-BS Channel Knowledge Map Generation for Cell-Free Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ZQQ6CKW}},
  note         = {Machine review of arXiv:2507.23236}
}
read the original abstract

Channel knowledge map (CKM) inference across base stations (BSs) is the key to achieving efficient environmentaware communications. This paper proposes an environmentaware cross-BS CKM inference method called BS-1-to-N based on the generative diffusion model. To this end, we first design the BS location embedding (BSLE) method tailored for cross-BS CKM inference to embed BS location information in the feature vector of CKM. Further, we utilize the cross- and self-attention mechanism for the proposed BS-1-to-N model to respectively learn the relationships between source and target BSs, as well as that among target BSs. Therefore, given the locations of the source and target BSs, together with the source CKMs as control conditions, cross-BS CKM inference can be performed for an arbitrary number of source and target BSs. Specifically, in architectures with massive distributed nodes like cell-free networks, traditional methods of sequentially traversing each BS for CKM construction are prohibitively costly. By contrast, the proposed BS-1-to-N model is able to achieve efficient CKM inference for a target BS at any potential location based on the CKMs of source BSs. This is achieved by exploiting the fact that within a given area, different BSs share the same wireless environment that leads to their respective CKMs. Therefore, similar to multi-view synthesis, CKMs of different BSs are representations of the same wireless environment from different BS locations. By mining the implicit correlation between CKM and BS location based on the wireless environment, the proposed BS-1-to-N method achieves efficient CKM inference across BSs. We provide extensive comparisons of CKM inference between the proposed BS-1-to-N generative model versus benchmarking schemes, and provide one use case study to demonstrate its practical application for the optimization of BS deployment.

Figures

Figures reproduced from arXiv: 2507.23236 by the authors.

Figure 1
Figure 1. An illustration of cross-BS CKM inference problem. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An Illustration of the feasibility of cross-BS CKM inference. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Illustration of the BS location encoding (BSLE). [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Architecture of the proposed BS-1-to-N generative diffusion model for cross-BS CKM inference. Gaussian noise to M with a variance sequence of {βt} during the forward process as Mt ∼ q(Mt|Mt−1) = N (Mt; p 1 − βtMt−1, βtI). (16) Based on the properties of joint probabili…
Figure 6
Figure 6. Figure 6: Comparison of the inferred CKMs of BS-1-to-N under different numbers of source BSs. via BSLE and further guides the CKM inference across BSs via the attention mechanism. Compared with the benchmarks, the trained BS-1-to-N has a significantly better awareness of the wir…
Figure 7
Figure 7. Figure 7: Comparison of the inferred CKMs of BS-1-to-N with other benchmarks under 5 source BSs. 1 2 3 4 5 6 7 8 9 10 Numbers of the target BSs 0 5 10 15 Time required for inference (s) UNet BS 1-to-N Distance-based weighting [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Comparison of the inference time under different numbers of target [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Tutorial on Learning-Based Radio Map Construction: Data, Paradigms, and Physics-Awareness

    eess.SY 2026-03 unverdicted novelty 5.0 of 10

    Learning-based radio map construction is taxonomized as source-aware forward prediction versus source-agnostic inverse reconstruction, spanning five neural families, optics-inspired continuous fields, and a three-leve...

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [22]

    Generating CKM Using Others' Data: Cross-AP CKM Inference with Deep Learning

    Z. Dai, D. Wu, X. Xu, and Y . Zeng, “Generating CKM using others’ data: Cross-AP CKM inference with deep learning,” arXiv preprint arXiv:2411.17716, 2024

    work page Pith review arXiv 2024

  2. [23]

    Generative CKM construction using partially observed data with diffusion model,

    S. Fu, Z. Wu, D. Wu, and Y . Zeng, “Generative CKM construction using partially observed data with diffusion model,” in 2025 IEEE 101st V ehicular Technology Conference (VTC2024-Spring), Jun. 2025

    work page 2025

  3. [1]

    Towards 6G wireless communication networks: Vision, enabling technologies, and new paradigm shifts,

    X. You, C.-X. Wang, J. Huang, X. Gao, Z. Zhang, M. Wang, Y . Huang, C. Zhang, Y . Jiang, J. Wanget al., “Towards 6G wireless communication networks: Vision, enabling technologies, and new paradigm shifts,” Sci. China Inf. Sci. , vol. 64, no. 1, pp. 1–74, 2021

    work page 2021

  4. [2]

    Cell-free massive MIMO versus small cells,

    H. Q. Ngo, A. Ashikhmin, H. Yang, E. G. Larsson, and T. L. Marzetta, “Cell-free massive MIMO versus small cells,” IEEE Trans. Wirel. Commun., vol. 16, no. 3, pp. 1834–1850, Mar. 2017

    work page 2017

  5. [3]

    Rate-region characterization and channel estimation for cell-free symbiotic radio communications,

    Z. Dai, R. Li, J. Xu, Y . Zeng, and S. Jin, “Rate-region characterization and channel estimation for cell-free symbiotic radio communications,” IEEE Trans. Commun. , vol. 71, no. 2, pp. 674–687, Feb. 2023

    work page 2023

  6. [4]

    Ultra-dense networks: A survey,

    M. Kamel, W. Hamouda, and A. Youssef, “Ultra-dense networks: A survey,” IEEE Commun. Surv. Tutor . , vol. 18, no. 4, pp. 2522–2545, Fourthquarter 2016

    work page 2016

  7. [5]

    User-centric ultra-dense networks for 5G: challenges, methodologies, and directions,

    S. Chen, F. Qin, B. Hu, X. Li, and Z. Chen, “User-centric ultra-dense networks for 5G: challenges, methodologies, and directions,” IEEE Wirel. Commun., vol. 23, no. 2, pp. 78–85, Apr. 2016

    work page 2016

  8. [6]

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

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

    work page 2021

Show all 40 references
  1. [7]

    A tutorial on environment-aware communications via channel knowledge map for 6G,

    Y . Zeng, J. Chen, J. Xu, D. Wu, X. Xu, S. Jin, X. Gao, D. Gesbert, S. Cui, and R. Zhang, “A tutorial on environment-aware communications via channel knowledge map for 6G,” IEEE Commun. Surv. Tutor . , pp. 1–1, Feb. 2024

    2024

  2. [8]

    Environment-aware hybrid beamforming by leveraging channel knowledge map,

    D. Wu, Y . Zeng, S. Jin, and R. Zhang, “Environment-aware hybrid beamforming by leveraging channel knowledge map,” IEEE Trans. Wirel. Commun., vol. 23, no. 5, pp. 4990–5005, 2024

    2024

  3. [9]

    You may use the same channel knowledge map for environment-aware NLos sensing and communication,

    D. Wu, Z. Dai, and Y . Zeng, “You may use the same channel knowledge map for environment-aware NLos sensing and communication,” arXiv preprint arXiv:2507.03589, 2025

    2025 arXiv

  4. [10]

    IMNet: Interference-aware channel knowledge map construction and localiza- tion,

    L. Zhao, Z. Fei, X. Wang, J. Huang, Y . Li, and Y . Zhang, “IMNet: Interference-aware channel knowledge map construction and localiza- tion,” IEEE Wirel. Commun. Lett. , vol. 14, no. 3, pp. 856–860, Mar. 2025

    2025

  5. [11]

    RadioUNet: Fast radio map estimation with convolutional neural networks,

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

    2021

  6. [12]

    Sionna RT: Differentiable ray tracing for radio propagation modeling,

    J. Hoydis, F. A. Aoudia, S. Cammerer, M. Nimier-David, N. Binder, G. Marcus, and A. Keller, “Sionna RT: Differentiable ray tracing for radio propagation modeling,” in 2023 IEEE Globecom Workshops (GC Wkshps), Dec 2023, pp. 317–321

    2023

  7. [13]

    Channel knowledge map for environment-aware communications: EM algorithm for map construc- tion,

    K. Li, P. Li, Y . Zeng, and J. Xu, “Channel knowledge map for environment-aware communications: EM algorithm for map construc- tion,” in 2022 IEEE Wireless Communications and Networking Confer- ence (WCNC) . IEEE, 2022, pp. 1659–1664

    2022

  8. [14]

    Model-free radio map estimation in massive MIMO systems via semi- parametric Gaussian regression,

    N. Dal Fabbro, M. Rossi, G. Pillonetto, L. Schenato, and G. Piro, “Model-free radio map estimation in massive MIMO systems via semi- parametric Gaussian regression,” IEEE Wirel. Commun. Lett. , vol. 11, no. 3, pp. 473–477, Mar. 2022

    2022

  9. [15]

    3D-RadioDiff: An altitude-conditioned diffusion model for 3D radio map construction,

    L. Zhao, Z. Fei, X. Wang, J. Luo, and Z. Zheng, “3D-RadioDiff: An altitude-conditioned diffusion model for 3D radio map construction,” IEEE Wirel. Commun. Lett. , vol. 14, no. 7, pp. 1969–1973, Jul. 2025

    1969

  10. [16]

    CKMDiff: A generative diffusion model for CKM construction via inverse problems with learned priors,

    S. Fu, Y . Zeng, Z. Wu, D. Wu, S. Jin, C.-X. Wang, and X. Gao, “CKMDiff: A generative diffusion model for CKM construction via inverse problems with learned priors,” arXiv preprint arXiv:2504.17323, 2025

    2025 arXiv

  11. [17]

    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, Oct. 2023

    2023

  12. [18]

    Deep learning-based CKM construction with image super-resolution,

    S. Wang, X. Xu, and Y . Zeng, “Deep learning-based CKM construction with image super-resolution,” in 2025 IEEE 101st V ehicular Technology Conference (VTC2024-Spring), Jun. 2025

    2025

  13. [19]

    An I2I inpainting approach for efficient channel knowledge map construction,

    Z. Jin, L. You, J. Wang, X.-G. Xia, and X. Gao, “An I2I inpainting approach for efficient channel knowledge map construction,” IEEE Trans. Wirel. Commun. , vol. 24, no. 2, pp. 1415–1429, Feb. 2025

    2025

  14. [20]

    Learning-based remote channel inference: Feasibility analysis and case study,

    S. Chen, Z. Jiang, S. Zhou, Z. Niu, Z. He, A. Marinescu, and L. A. DaSilva, “Learning-based remote channel inference: Feasibility analysis and case study,” IEEE Trans. Wirel. Commun. , vol. 18, no. 7, pp. 3554– 3568, July. 2019

    2019

  15. [21]

    Deep learning for TDD and FDD massive MIMO: Mapping channels in space and frequency,

    M. Alrabeiah and A. Alkhateeb, “Deep learning for TDD and FDD massive MIMO: Mapping channels in space and frequency,” in 2019 53rd Asilomar Conference on Signals, Systems, and Computers , Nov. 2019, pp. 1465–1470

    2019

  16. [24]

    Zero-1-to-3: Zero-shot one image to 3D object,

    R. Liu, R. Wu, B. Van Hoorick, P. Tokmakov, S. Zakharov, and C. V ondrick, “Zero-1-to-3: Zero-shot one image to 3D object,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2023, pp. 9298–9309

    2023

  17. [25]

    Eschernet: A generative model for scalable view synthesis,

    X. Kong, S. Liu, X. Lyu, M. Taher, X. Qi, and A. J. Davison, “Eschernet: A generative model for scalable view synthesis,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 9503–9513

    2024

  18. [26]

    Nearest neighbor pattern classification,

    T. Cover and P. Hart, “Nearest neighbor pattern classification,” IEEE Trans. Inf. Theory , vol. 13, no. 1, pp. 21–27, Jan. 1967

    1967

  19. [27]

    How much data is needed for channel knowledge map construction?

    X. Xu and Y . Zeng, “How much data is needed for channel knowledge map construction?” IEEE Trans. Wirel. Commun. , pp. 1–1, May. 2024

    2024

  20. [28]

    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. Networking, pp. 1–1, 2024

    2024

  21. [29]

    Denoising diffusion probabilistic mod- els,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic mod- els,” in Proceedings of the 34th International Conference on Neural Information Processing Systems , ser. NIPS ’20. Red Hook, NY , USA: Curran Associates Inc., 2020

    2020

  22. [30]

    Diffusion models beat GANs on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat GANs on image synthesis,” in Proceedings of the 35th International Conference on Neural Information Processing Systems , ser. NIPS ’21. Red Hook, NY , USA: Curran Associates Inc., 2021

    2021

  23. [31]

    Palette: Image-to-image diffusion models,

    C. Saharia, W. Chan, H. Chang, C. Lee, J. Ho, T. Salimans, D. Fleet, and M. Norouzi, “Palette: Image-to-image diffusion models,” in ACM SIGGRAPH 2022 Conference Proceedings , ser. SIGGRAPH ’22. New York, NY , USA: Association for Computing Machinery, 2022

    2022

  24. [32]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,” arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  25. [33]

    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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2022, pp. 10 684–10 695

    2022

  26. [34]

    RoFormer: Enhanced transformer with rotary position embedding,

    S. Jianlin, A. Murtadha, L. Yu, P. Shengfeng, B. Wen, and L. Yunfeng, “RoFormer: Enhanced transformer with rotary position embedding,” Neurocomputing, vol. 568, p. 127063, Feb. 2024

    2024

  27. [35]

    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 , vol. 30, 2017

    2017

  28. [36]

    Perceiver: General perception with iterative attention,

    A. Jaegle, F. Gimeno, A. Brock, O. Vinyals, A. Zisserman, and J. Carreira, “Perceiver: General perception with iterative attention,” in International conference on machine learning . PMLR, 2021, pp. 4651– 4664

    2021

  29. [37]

    Convnext v2: Co-designing and scaling convnets with masked autoencoders,

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, and S. Xie, “Convnext v2: Co-designing and scaling convnets with masked autoencoders,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 16 133–16 142

    2023

  30. [38]

    Dataset of pathloss and ToA radio maps with localization application,

    C ¸ . Yapar, R. Levie, G. Kutyniok, and G. Caire, “Dataset of pathloss and ToA radio maps with localization application,” arXiv preprint:2212.11777, 2022

    2022 arXiv

  31. [39]

    Image quality assessment: from error visibility to structural similarity,

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

    2004

  32. [40]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.