Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Non-Identical Diffusion Models in MIMO-OFDM Channel Generation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper proves that diffusion generation can stay correct when every input element has its own noise schedule, and applies this to wireless channel recovery.

desk verdict Elementwise-time diffusion is a real engineering idea and the experiments are thorough, but Theorem 2 is not proven as written and Algorithm 2's initialization violates the theorem's hypothesis, so the theory needs major revision. read the letter →

arxiv 2509.01641 v3 pith:ARFX2LQZ submitted 2025-09-01 eess.SP cs.AIcs.LG

Yuzhi Yang , Omar Alhussein , Mérouane Debbah This is my paper
classification eess.SPcs.AIcs.LG
keywords non-identicaldiffusionelement-wisetimeembeddingDDIMMIMO-OFDMchannelestimationgenerationpilotreliabilityinterleavedMLP-Mixermodels
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

Standard diffusion models use a single scalar time to represent global noise level, which assumes every element degrades uniformly. This paper introduces non-identical diffusion, where each element of the input, e.g. each OFDM subcarrier or antenna, has its own time indicator and noise schedule. The central theoretical result (Theorem 2) shows that the inverse DDIM generation process preserves the correct data distribution for any valid element-wise schedule and any matching initialization, so generation from uneven-quality estimates still lands on the true channel distribution. The paper then applies this to MIMO-OFDM channel recovery, where pilot placement makes some elements far more reliable than others, and demonstrates numerically that the non-identical scheme outperforms identical diffusion, especially for highly non-uniform noise patterns.

What carries the argument

The central object is the element-wise time indicator: a matrix of noise levels alpha_t, one per input element, which generalizes the scalar time of standard diffusion. Each entry controls how much of the original signal remains in that element at time t, with matching beta_t = sqrt(1−alpha_t^2). The forward SDE dH_t = [log alpha_t]' ∘ H_t dt + sqrt(−2[log alpha_t]') ∘ dW_t yields the closed-form transition H_t = alpha_t ∘ H_0 + beta_t ∘ ξ, and the non-identical DDIM ODE defines a probability flow. Theorem 2, proved via a Fokker–Planck argument and Grönwall's inequality following the proof template of the cited sampling work, is the mechanism that guarantees generation correctness for arbitr

What would settle it

On a synthetic dataset with a known Gaussian mixture, run non-identical DDIM from the Eq. (14) initialization and compare the output distribution to the true mixture via a two-sample test; then run the same algorithm initialized from a point that exactly satisfies G_t0 ~ phi^{alpha}_{T−t0}. If the first output fails the test while the second passes, the initialization mismatch is the deciding factor. Alternatively, measure the empirical variance of the initial estimate and compare it with beta^2 at alpha = M/(M+1) to check whether the schedule is consistent.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that the non-identical diffusion process — defined by a vector-valued noise schedule alpha_t in (0,1]^d replacing the scalar time index — admits a correct DDIM inverse process. Specifically, Theorem 2 states that under the ODE dG_t/dt = [log alpha_{T-t}]' ∘ (G_t − D^{alpha}_{T−t}(G_t)), if the starting point has the marginal law phi^{alpha}_{T−t0}, then for every t the law of G_t equals phi^{alpha}_{T−t}, and at t=T it equals the true data distribution. This holds for any schedule satisfying Definition 1 and any starting point with the required marginal law, provided the data distribution has compact support and the denoiser is the exact conditi

Load-bearing premise

The main guarantee assumes an error-free denoiser and a bounded data distribution, and the algorithm's starting point in Eq. (14) does not exactly match the marginal noise level that Theorem 2 requires.

Editorial extensions

If this is right

  • MIMO-OFDM channel recovery can start from a rough channel estimate whose reliability varies strongly across antennas and subcarriers, and the non-identical DDIM output still follows the training channel distribution.
  • Training and generation no longer need to share the same noise path: the theory permits arbitrary paths, so training can mix several noise patterns while generation uses a more targeted stepping scheme such as water-filling.
  • Water-filling stepping, which first brings the least reliable elements up to the reliability of the better ones, consistently achieved the best final NMSE in the paper's experiments compared with linear or hybrid stepping.
  • The numerical advantage over identical diffusion grows as the initial noise map becomes more non-uniform, which the paper demonstrates with Pilot-Car and Salt-Rec patterns outperforming White and Salt patterns.
  • The training noise pattern strongly affects generation: a model trained on a mix of patterns ('All') generalizes best, while training on one specific pattern can overfit and degrade recovery under other initializations.

Reading between the lines

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

  • A direct corollary of the paper's framework, not explored there, is that the same element-wise reliability-map initialization should apply beyond wireless channels to any inverse problem with a known per-element confidence mask, such as image inpainting, radio map reconstruction, or accelerated MRI-style partial observations.
  • The observed advantage of water-filling stepping suggests that the stepping path is a free design parameter; a testable extension is to optimize the path by minimizing an estimate of the per-element denoising error of the learned network rather than fixing water-filling a priori.
  • The theorem guarantees distribution preservation relative to a matching starting marginal, but the algorithm's initialization in Eq. (14) does not generally satisfy that condition; quantifying this gap by comparing empirical noise variance at initialization with beta^2 at the scheduled alpha would clarify how much of the practical performance rests on the theorem versus the network's robustness.
  • The paper leaves the row/column embedding puzzle open; one plausible explanation worth testing is that column-wise embedding provides a more consistent reliability context per operational dimension, which could be probed with ablations on synthetic noise maps.
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 / 5 minor

Summary. The paper proposes 'non-identical diffusion,' a diffusion model in which the noise level is described by an element-wise time matrix rather than a scalar. It defines a forward SDE with elementwise decay factors α_t, states Theorem 1 giving the marginal law as α_t∘H0+β_t∘ξ, and states Theorem 2 claiming that the DDIM-like ODE with the conditional-expectation denoiser maps the marginal φ_{T−t0} to the data distribution μ. The paper then introduces training and generation algorithms for this framework, a dimension-wise time-embedding scheme for MIMO-OFDM channel matrices, and numerical experiments comparing identical and non-identical diffusion under several pilot-like noise patterns.

Significance. If the theoretical claim is repaired, the paper proposes a natural and potentially useful extension of DDIM to settings where different elements of an observation have different reliabilities; MIMO-OFDM channel recovery with pilot-based nonuniform error is a reasonable application. The paper is also commendably broad empirically: it reports 48 training experiments, several training noise patterns, multiple generation schedulers, and an anonymous code link. However, as written, the central theorem is not proven and the proposed initialization does not satisfy the theorem's hypothesis. These issues are load-bearing and must be corrected before the theoretical or algorithmic claims can be accepted.

major comments (3)
  1. [Appendix A, Eqs. (24)-(26)] The proof of Theorem 2 is invalid as written. After Eq. (23), the expectation of the integral should contain ψ_s, not ψ_t0; the continuity equation for the density under the ODE flow is ψ_t = ψ_t0 − ∫_{t0}^t div(b_s ψ_s) ds. Eq. (25) instead places ψ_t0 inside the s-integral, and Eq. (26) repeats the same freezing for φ. Consequently the Gronwall estimate in Eq. (28) controls a quantity that is not the true φ_t, and the conclusion φ_t=0 does not follow. The result may be repairable via a standard probability-flow/Fokker-Planck argument, but this appendix does not provide that proof.
  2. [Section IV-A and Algorithm 2, Eqs. (14)-(16)] The initialization used in generation does not satisfy Theorem 2's hypothesis. From Eq. (14), eH(0) = Hbar/(M+1) = [M/(M+1)]∘H + [1/(M+1)]∘z with z∼CN(0,I), so its per-element noise variance is 1/(M+1)^2. At the assigned time τ0=γ^{-1}(M/(M+1)), the forward marginal has β^2=1−(M/(M+1))^2=(2M+1)/(M+1)^2. These differ, so G_t0 is not distributed as φ_{T−t0}. A consistent elementwise choice would be to divide by sqrt(M^2+1) and set τ0=γ^{-1}(M/sqrt(M^2+1)), but the manuscript does not make this choice.
  3. [Section V-C, Fig. 7, Tables II-III] The empirical comparison is partly circular. The non-identical configuration used in Fig. 7 is described as 'chosen based on prior experiments,' but no validation/test split is specified, so it is unclear whether the same test initializations were used for model selection. Similarly, Table III evaluates the 'All' training pattern on all test patterns and reports the best final numbers; without an explicit validation procedure, this is close to selecting the model on the test set. The claim that non-identical diffusion 'remarkably outperforms' identical diffusion is therefore not established at the reported level of rigor.
minor comments (5)
  1. [Section II-A, Theorem 1] The proof of Theorem 1 is only outlined. The result is elementary, but since it is foundational for Theorem 2, a full Fokker-Planck verification should be included for self-containedness.
  2. [Section IV-B] The notation around τ_c and τ_a is confusing: τ_c is called 'subcarrier-domain mapping' but is obtained by averaging over subcarriers. Please clarify the terminology and check the placement of the embeddings in the Row-Wise, Column-Wise, and Embedding Together bullets.
  3. [Algorithm 2] The relation between the velocity parameterization and the denoiser, D_α(g)=α∘g−β∘f(g,τ;θ), appears only as a comment in the algorithm. It should be stated and justified in the main text, since the algorithm's correctness depends on it.
  4. [Tables II and III] Please state explicitly that the reported values are NMSE and describe how they are averaged over the test set and over random realizations of the noise initialization.
  5. [Throughout] Minor typos and inconsistencies: Appendix A writes 't ∈ [to, T]' where t0 is meant; Section V-C2 says 'four kinds of noise initialization' while the settings list six patterns; the proof of Theorem 2 assumes an exact denoiser D and compact support, which the paper acknowledges in Section II, but a sentence explicitly stating that Theorem 2 applies to the ideal D_α and not to the trained network would prevent overclaiming.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity in the core derivation; Theorem 2 is a self-contained extension of the DDIM probability-flow argument and does not reduce to its assumptions. Minor self-citations and test-set-based configuration choices are present but are not load-bearing circularity.

full rationale

The central theoretical claim, Theorem 2, is a marginal-preservation statement for the ODE (5) given that the process starts at the forward marginal. The proof derives the conclusion from the Fokker-Planck/continuity equation and an external uniqueness lemma [31]; it does not define the conclusion into the premises. The non-identical construction in Definition 1 is a componentwise generalization of standard DDIM, and Theorem 1 explicitly gives the forward marginal law, so the hypothesis G_t0 ~ phi^alpha_{T-t0} is not the same as the conclusion G_t ~ phi^alpha_{T-t}. The application initialization (15)-(16) uses the reliability map M to set the starting time; this is side information, not a fitted parameter, so the measured NMSE against held-out H is not forced by construction. The empirical comparison does select the non-identical configuration "based on prior experiments" (Sec. V-C2) and evaluates on the same test initializations, which is a test-set-selection concern, but it is not an equation-level reduction and does not make the reported improvement circular in the sense of this analysis. Self-citations [25], [29] are used for the backbone architecture and experimental protocol, not to establish the central correctness claim. Two non-circular correctness caveats are worth explicit flagging: (i) Appendix A's integral equation (25) freezes the density at psi_t0 in the s-integral, which appears to invalidate the Gronwall argument as written; (ii) Algorithm 2's starting point has noise variance (M+1)^{-2}, while the forward process at alpha=M/(M+1) has variance 1-(M/(M+1))^2, so the initialization does not in general satisfy the theorem's marginal hypothesis. These are proof/implementation gaps, not circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The ledger captures the assumptions required for the central claims: standard stochastic calculus, bounded support of the data distribution, exact knowledge of the reliability map M, and NN approximation of the denoiser. The experimental claims also depend on hand-selected training noise patterns and scheduler hyperparameters.

free parameters (5)
  • stochasticity coefficient epsilon = 0.4
    Chosen for generation in Eq. (9); affects the DDIM/DDPM hybrid update and is not derived.
  • hybrid stepping mixing coefficient epsilon_hybrid = 0.3, 0.5, 0.7; water-filling (0) performs best
    Interpolates uniform and water-filling stepping in Eq. (13); tuned per experiment.
  • training noise-pattern distributions = period ranges 4-20 subcarriers, 4-10 antennas; car-only pattern
    Hand-designed sampling schemes for alpha paths; generation results depend strongly on them.
  • default SNR per initialization pattern = -10, -5, 0, 0, 10, 10 dB for White, Exp, Salt, Salt-Rec, Pilot, Pilot-Car
    Test settings chosen without derivation from a channel model.
  • number of generation steps NG = 50
    Fixed for all experiments; no sensitivity analysis is provided.
assumptions (5)
  • standard math Itô SDE and Fokker-Planck theory, including existence and uniqueness of ODE solutions under Lipschitz and linear growth conditions
    Used in the proofs of Theorem 1 and Theorem 2 (Appendix A).
  • domain assumption The data distribution mu is bounded with compact support
    Assumed in Theorem 2 to guarantee that D^alpha_t is Lipschitz and satisfies a linear growth condition.
  • domain assumption The initial channel estimate satisfies Hbar - M circle H ~ CN(0, I) with M known exactly
    Eq. (14); the paper explicitly treats M and Hbar as provided inputs without discussing how they are obtained.
  • domain assumption The trained neural network f approximates the conditional expectation D^alpha_t through velocity parameterization
    Algorithms 1 and 2 substitute f for the exact denoiser; no approximation-error bound is provided.
  • domain assumption alpha_t is componentwise non-increasing, differentiable, alpha_0 = 1, and lim_{t -> T} alpha_t = 0
    Definition 1 imposes these regularity conditions on the diffusion schedule.
invented entities (2)
  • element-wise time matrix tau
    purpose: Represents the per-element noise level in the diffusion process instead of a scalar time index.
    A new modeling variable used in the SDE, training algorithms, and generation algorithms. It has no external falsifiable handle outside the proposed framework.
  • non-identical diffusion process
    purpose: Generalizes identical diffusion by allowing different noise schedules for different elements.
    A formal construct defined in Definition 1; its usefulness is measured through the numerical experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Non-Identical Diffusion Models in MIMO-OFDM Channel Generation." pith.science (2026). https://pith.science/paper/ARFX2LQZ

@misc{pith2026250901641,
  author       = {Pith},
  title        = {Pith review of: Non-Identical Diffusion Models in MIMO-OFDM Channel Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ARFX2LQZ}},
  note         = {Machine review of arXiv:2509.01641}
}
read the original abstract

We propose a novel diffusion model, termed the non-identical diffusion model, and investigate its application to wireless orthogonal frequency division multiplexing (OFDM) channel generation. Unlike the standard diffusion model that uses a scalar-valued time index to represent the global noise level, we extend this notion to an element-wise time indicator to capture local error variations more accurately. Non-identical diffusion enables us to characterize the reliability of each element (e.g., subcarriers in OFDM) within the noisy input, leading to improved generation results when the initialization is biased. Specifically, we focus on the recovery of wireless multi-input multi-output (MIMO) OFDM channel matrices, where the initial channel estimates exhibit highly uneven reliability across elements due to the pilot scheme. Conventional time embeddings, which assume uniform noise progression, fail to capture such variability across pilot schemes and noise levels. We introduce a matrix that matches the input size to control element-wise noise progression. Following a similar diffusion procedure to existing methods, we show the correctness and effectiveness of the proposed non-identical diffusion scheme both theoretically and numerically. For MIMO-OFDM channel generation, we propose a dimension-wise time embedding strategy. We also develop and evaluate multiple training and generation methods and compare them through numerical experiments.

Figures

Figures reproduced from arXiv: 2509.01641 by the authors.

Figure 1
Figure 1. A conceptual illustration of non-identical diffusion. Extending this idea to wireless transceivers, similar gen￾eration problems from rough initializations are widespread. Unlike traditional neural network (NN)-based mapping frame￾works, generative methods are gaining increased attention due to their high resilience to the ubiquitous noise in wireless systems [19]–[21]. Instead of generating solely from white noise,… view at source ↗
Figure 3
Figure 3. A conceptual illustration of the proposed two-dimensional time embedding method the system and the specific method of obtaining H¯ and M, treating them as originally provided inputs. We note that although the channel here is a two-dimensional complex matrix instead of the real vectors discussed above, the difference only lies in the design of the backbone NN, and the algorithms described above can be directly applie… view at source ↗
Figure 4
Figure 4. Training convergence curve with different embedding meth￾ods [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Training convergence curve with different normalization and dimensional averaging methods. independent training experiments corresponding to two repre￾senting methods of hbi as explained in Section III A, three time embeddings, and two time averaging methods as introdu…
Figure 6
Figure 6. Figure 6: An illustration of different initial noise patterns. The noise power are normalized and only the power is shown. from one of the following distributions, completing the power map. The noise is independently Gaussian-distributed with the generated power map. We note tha…
Figure 7
Figure 7. Figure 7: The comparison between identical and non-identical diffusion models. Initialization Exp Salt Salt-Rec Pilot Pilot-Car τ-Linear 0.098 0.202 0.172 0.386 0.276 τ-Hybrid (0.7) 0.095 0.189 0.155 0.367 0.225 τ-Hybrid (0.5) 0.094 0.182 0.147 0.358 0.205 τ-Hybrid (0.3) 0.093 0…
Figure 8
Figure 8. Figure 8: The generation results of different embedding schemes. I, where “baseline” indicates the column-embedding, α-avg, power-normalization result, and other names indicate their minimum differences to the baseline. These results generally look similar to those in Table I. H…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages

  1. [1]

    Denoising diffusion probabilis- tic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilis- tic models,” in Proc. 34th Int. Conf. Neural Inform. Process. Syst. (NeurIPS) , Red Hook, NY , USA, 2020, pp. 1–12

    work page 2020

  2. [2]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502 , 2020

    arXiv 2010

  3. [3]

    Versatile diffusion: Text, images and variations all in one diffusion model,

    X. Xu, Z. Wang, G. Zhang, et al. , “Versatile diffusion: Text, images and variations all in one diffusion model,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV) , 2023, pp. 7754– 7765

    work page 2023

  4. [4]

    IRDM: A generative diffusion model for indoor radio map interpolation,

    K. Qiu, S. Bakirtzis, I. Wassell, et al. , “IRDM: A generative diffusion model for indoor radio map interpolation,” in IEEE Global Commun. Conf. (Globecom) , IEEE, 2023, pp. 01–06

    work page 2023

  5. [5]

    RF-Diffusion: Radio signal generation via time-frequency diffusion,

    G. Chi, Z. Yang, C. Wu, et al. , “RF-Diffusion: Radio signal generation via time-frequency diffusion,” in Proc. 30th Annu. Int. Conf. Mobile Comput. Netw. (MobiCom) , 2024, pp. 77–92

    work page 2024

  6. [6]

    Diffusion-based gen- erative prior for low-complexity MIMO channel estimation,

    B. Fesl, M. Baur, F. Strasser, et al. , “Diffusion-based gen- erative prior for low-complexity MIMO channel estimation,” IEEE Wireless Commun. Lett. , vol. 13, no. 12, pp. 3493–3497, Dec. 2024, ISSN : 2162-2345. DOI: 10.1109/LWC.2024.34745 70

    work page arXiv 2024

  7. [7]

    MIMO channel estimation using score-based generative models,

    M. Arvinte and J. I. Tamir, “MIMO channel estimation using score-based generative models,” IEEE Trans. Wireless Com- mun., vol. 22, no. 6, pp. 3698–3713, Jun. 2023, ISSN : 1558-

    work page 2023

  8. [8]

    Diffusion models: A comprehensive survey of methods and applications,

    L. Yang, Z. Zhang, Y . Song, et al. , “Diffusion models: A comprehensive survey of methods and applications,” ACM Comput. Surv., vol. 56, no. 4, pp. 1–39, 2023

    work page 2023

Show all 32 references
  1. [9]

    An overview of diffusion models: Applications, guided generation, statistical rates and optimization,

    M. Chen, S. Mei, J. Fan, et al. , “An overview of diffusion models: Applications, guided generation, statistical rates and optimization,” arXiv preprint arXiv:2404.07771 , 2024

    2024 arXiv

  2. [10]

    Repaint: Inpainting using denoising diffusion probabilistic models,

    A. Lugmayr, M. Danelljan, A. Romero, et al. , “Repaint: Inpainting using denoising diffusion probabilistic models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 11 461–11 471

    2022

  3. [11]

    Latentpaint: Image inpainting in latent space with diffusion models,

    C. Corneanu, R. Gadde, and A. M. Martinez, “Latentpaint: Image inpainting in latent space with diffusion models,” in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV) , 2024, pp. 4334–4343

    2024

  4. [12]

    Image inpainting via tractable steering of diffusion models,

    A. Liu, M. Niepert, and G. V . d. Broeck, “Image inpainting via tractable steering of diffusion models,” arXiv preprint arXiv:2401.03349, 2023

    2023 arXiv

  5. [13]

    Imagic: Text-based real image editing with diffusion models,

    B. Kawar, S. Zada, O. Lang, et al. , “Imagic: Text-based real image editing with diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 6007– 6017

    2023

  6. [14]

    Diffusion model-based image editing: A survey,

    Y . Huang, J. Huang, Y . Liu, et al. , “Diffusion model-based image editing: A survey,” IEEE Trans. Pattern Anal. Mach. Intell., 2025

    2025

  7. [15]

    Inversion-based style transfer with diffusion models,

    Y . Zhang, N. Huang, F. Tang, et al. , “Inversion-based style transfer with diffusion models,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2023, pp. 10 146– 10 156

    2023

  8. [16]

    RADiff: Controllable diffusion models for radio astronomical maps generation,

    R. Sortino, T. Cecconello, A. DeMarco, et al. , “RADiff: Controllable diffusion models for radio astronomical maps generation,” IEEE Trans. Artif. Intell. , 2024

    2024

  9. [17]

    RM-Gen: Conditional diffusion model-based radio map generation for wireless net- works,

    X. Luo, L. Zhizhen, Z. Peng, et al. , “RM-Gen: Conditional diffusion model-based radio map generation for wireless net- works,” in IFIP Networking Conference (IFIP Networking) , IEEE, 2024, pp. 543–548

    2024

  10. [18]

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

    X. Wang, K. Tao, N. Cheng, et al. , “RadioDiff: An effective generative diffusion model for sampling-free dynamic radio map construction,” IEEE Trans. Cogn. Commun. Netw. , 2024

    2024

  11. [19]

    An introduction to deep learning for the physical layer,

    T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,” IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 563–575, 2017. DOI: 10.1109/TCCN.2017.2758370

    2017

  12. [20]

    A comprehensive survey of knowledge-driven deep learning for intelligent wireless network optimization in 6G,

    R. Sun, N. Cheng, C. Li, et al. , “A comprehensive survey of knowledge-driven deep learning for intelligent wireless network optimization in 6G,” IEEE Commun. Surv. Tutor . , pp. 1–1, 2025. DOI: 10.1109/COMST.2025.3574765

    2025

  13. [21]

    F. Zhu, X. Wang, X. Li, et al. , Wireless large AI model: Shaping the AI-native future of 6G and beyond , 2025. arXiv: 2504.14653. [Online]. Available: https://arxiv.org/abs/2504.1 4653

    2025 arXiv

  14. [22]

    High dimensional channel estimation using deep generative networks,

    E. Balevi, A. Doshi, A. Jalal, et al., “High dimensional channel estimation using deep generative networks,” IEEE J. Sel. Areas Commun., vol. 39, no. 1, pp. 18–30, 2020

    2020

  15. [23]

    Diffusion model based channel estimation,

    X. Ma, Y . Xin, Y . Ren, et al., “Diffusion model based channel estimation,” in IEEE Int. Conf. Commun. (ICC Workshops) , Jun. 2024, pp. 1159–1164. DOI: 10.1109/ICCWorkshops5955 1.2024.10615282

    2024

  16. [24]

    Joint channel estimation and data detection in massive MIMO systems based on diffusion models,

    N. Zilberstein, A. Swami, and S. Segarra, “Joint channel estimation and data detection in massive MIMO systems based on diffusion models,” in IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Apr. 2024, pp. 13 291–13 295. DOI: 10.1109/ICASSP48485.2024.10446413

    2024

  17. [25]

    Generative diffusion receivers: Achieving pilot-efficient MIMO-OFDM communi- cations,

    Y . Yang, O. Alhussein, A. Arani, et al. , “Generative diffusion receivers: Achieving pilot-efficient MIMO-OFDM communi- cations,” arXiv preprint arXiv:2506.18419 , 2025

    2025

  18. [26]

    Channel mapping based on interleaved learning with complex-domain MLP-Mixer,

    Z. Chen, Z. Zhang, Z. Yang, et al., “Channel mapping based on interleaved learning with complex-domain MLP-Mixer,” IEEE Wireless Commun. Lett. , vol. 13, no. 5, pp. 1369–1373, 2024

    2024

  19. [27]

    Improved denoising diffu- sion probabilistic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffu- sion probabilistic models,” in Proc. Int. Conf. Mach. Learn. (ICML), PMLR, 2021, pp. 8162–8171

    2021

  20. [28]

    Progressive distillation for fast sam- pling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sam- pling of diffusion models,” in Proc. Int. Conf. Neural Inform. Process. Syst. (NeurIPS) , vol. 35, Curran Associates, Inc., 2022, pp. 17 870–17 881

    2022

  21. [29]

    A hybrid inference archi- tecture incorporating neural network with belief propagation for AI receivers,

    Y . Yang, Z. Zhang, Z. Chen, et al., “A hybrid inference archi- tecture incorporating neural network with belief propagation for AI receivers,” IEEE Trans. Wireless Commun., 2025, Early Access. DOI: 10.1109/TWC.2025.3552818

    2025

  22. [30]

    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

    1902 arXiv

  23. [31]

    New algorithms for sampling and diffusion mod- els,

    X. Zhang, “New algorithms for sampling and diffusion mod- els,” arXiv preprint arXiv:2406.09665 , 2024

    2024 arXiv

  24. [2248]

    DOI: 10.1109/TWC.2022.3220784

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.