Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Generative Diffusion Model-based Compression of MIMO CSI

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read This paper proposes a fixed-rate CSI codec whose trainable codebook encoder and conditional diffusion decoder achieve rate-distortion gains of more than two-fold over existing neural CSI compression methods.

desk verdict A sensible combination of existing vector-quantized encoding and conditional diffusion decoding applied to MIMO CSI, but the headline rate-distortion gain is not auditable because the codebook size is never reported and the main comparison uses baseline curves taken from other papers. read the letter →

arxiv 2503.03753 v1 pith:5JDXVDR7 submitted 2025-02-07 cs.IT cs.AIeess.SPmath.IT

classification cs.ITcs.AIeess.SPmath.IT
keywords CSIcompressiondiffusionmodelsvectorquantizationfixed-ratecodingsideinformationchannelpredictionMIMOfeedbackneural
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 proposes a fixed-rate neural codec for MIMO channel state information (CSI) feedback in which a trainable vector-quantization codebook compresses the current downlink CSI into a B-bit codeword and a conditional diffusion model decodes it, optionally aided by uplink CSI as side information. It aims to show that this combination is markedly better than existing neural CSI codecs for two tasks: reconstructing the current channel and predicting a future channel from the past one. Across both tasks the reported rate-distortion curves sit far below the baselines, with comparable distortion at less than half the bit rate in several operating points. If the claim holds, diffusion-based decoding moves from image restoration to a practical role in wireless feedback links, where halving the feedback rate at fixed channel quality translates directly into radio resource savings.

What carries the argument

The two load-bearing components are a trainable vector quantization codebook and a conditional denoising diffusion decoder. The codebook turns a neural representation of the CSI into a fixed-length discrete codeword via nearest-neighbor selection, trained with the codebook loss of Eq. (2); the decoder models p(z|c,y) with a denoising diffusion process, using the DDIM update of Eq. (6) to iterate from z_T=0 down to z_0, with the denoising network conditioned on the codeword and, when present, the side information. The training objective in Eq. (7) sums a likelihood-based diffusion loss with the codebook loss, and the claimed behavior follows from these two mechanisms working together.

What would settle it

Measure the gradient of the diffusion loss with respect to the encoder parameters in the training loop of Algorithm 1: if it is exactly zero (as the chain rule forces through the nearest-neighbor argmin without a specified gradient bypass), then the encoder is trained only by the codebook loss, and retraining with the diffusion loss detached from the encoder would give the same codebook, settling whether the joint training objective is the source of the reported gains.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a fixed-rate coding scheme for computing with side information can beat both fixed-rate and entropy-coded neural baselines by pairing a discrete codebook encoder with a diffusion decoder. The encoder maps the input to continuous vectors, replaces each with its nearest neighbor in a trainable codebook, and transmits only the indices; the decoder reconstructs the target by a short T=4 denoising diffusion process conditioned on the codeword and on uplink CSI when available. The paper reports that reaching about -7 dB NMSE on the standard reconstruction benchmark takes fewer than 150 bits with the proposed scheme while all baselines require more than 400 bits, and that in the prediction scenario the scheme attains comparable distortion at less than half the data rate. It also notes that the bit-rate increases from 64 to 128 and 128 to 192 bits keep yielding more than 0.5-1 dB gains, whereas the baselines saturate near -6 dB NMSE.

Load-bearing premise

The reported gains rest on the described joint training actually teaching the encoder through the diffusion reconstruction loss, but the codebook selection is a non-differentiable nearest-neighbor argmin and the paper specifies no straight-through estimator or other gradient path, so the encoder may only be trained by the codebook loss.

Editorial extensions

If this is right

  • For the CSI reconstruction task, the proposed fixed-rate codec achieves around -7 dB NMSE below 150 bits, where the best reported baseline needs more than 400 bits, so fixed-rate coding need not be a disadvantage relative to variable-length entropy coding.
  • For the CSI prediction task with uplink CSI as side information, the method attains comparable distortion at less than half the data rate of the adapted CsiNet and CRNet baselines, supporting the Wyner-Ziv-style setup as a viable feedback design.
  • Because the decoder uses only T=4 denoising steps, the backward-diffusion process is computationally short enough for the authors to argue practical deployment is plausible despite the general cost of diffusion sampling.
  • The rate-distortion curve for the proposed method is still improving at 192 bits, so the gains over baselines are not a one-point artifact but hold across the tested rate range.

Reading between the lines

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

  • If the reported RD curves are correct, the codebook is likely learning to preserve only the information in the current CSI that matters for the target (reconstruction or future channel), with the diffusion decoder supplying the rest from its learned prior; a direct test would be to fix the decoder and vary the codebook size at constant bit rate and see whether distortion improves as expected.
  • The training description in Algorithm 1 may not actually backpropagate the diffusion loss through the nearest-neighbor argmin, so the observed gains might come from the decoder's generative prior and only a codebook-loss-trained encoder; this is testable by checking gradients or ablating the diffusion loss from the encoder update.
  • Because the method is formulated as coding for computing with side information, the same encoder-decoder design should transfer to other tasks where the decoder needs a function of the source (for example, downlink beamforming metrics) rather than the source itself; the paper only demonstrates channel reconstruction and prediction.
  • The 64- to 192-bit regime in the prediction experiment is far below the bits needed by the baselines; an untested extension is whether the approach still dominates at very low rates (below 64 bits) where the codebook's expressiveness may become the bottleneck.
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

4 major / 5 minor

Summary. This paper proposes a fixed-rate neural codec for MIMO CSI compression, in which an encoder maps a CSI tensor to 64 codebook indices via vector quantization, and a conditional diffusion decoder reconstructs or predicts the target, optionally using UL CSI as side information. The training objective combines a VQ-VAE-style codebook loss with a denoising diffusion loss, and decoding uses a deterministic DDIM-style backward process with T=4 steps. Experiments are reported for two settings: future DL CSI prediction on Sionna CDL-C data with UL side information (Fig. 5) and reconstruction on the COST2100 dataset without side information (Fig. 6). The authors claim that the proposed method outperforms CsiNet/CRNet with uniform quantization and several published CSI codecs, in some cases requiring less than half the rate for comparable distortion.

Significance. If the reported experiments are correct, the paper would establish a vector-quantization-plus-diffusion codec as the best reported neural CSI compressor, and its application to prediction with side information is a useful new task formulation. The paper has notable strengths: it targets a practically motivated fixed-rate regime and a Wyner-Ziv-type setting, the method is clearly described at the block-diagram level, the angular-delay preprocessing is standard, and the authors make an explicit effort to compare with strong baselines, including entropy-coded autoencoders. The quantitative claims, however, are not yet auditable: the codebook size is omitted, the training objective's gradient path through quantization is unspecified, and Fig. 6 imports baseline curves from other papers without identical evaluation conditions.

major comments (4)
  1. [Section III-C, Algorithm 1] The training loop in Algorithm 1 minimizes L in Eq. (7) after setting c = fenc(x; θenc), where fenc includes the nearest-neighbor selection of codebook vectors described in Section III-A. This selection is non-differentiable, and the paper never specifies a straight-through estimator, a Gumbel relaxation, or any other mechanism that lets gradients from the diffusion reconstruction term ¯αt/(1−¯αt)∥z0 − Dθ(√¯αtz0 + √1−¯αtε, c, y, t)∥2 reach θenc. In the VQ-VAE loss (2), the first term uses sg[cconti] and therefore contributes no encoder gradient, while the second term pulls cconti toward the frozen selected embedding. Consequently, as written, the diffusion loss trains only the decoder and codebook embeddings, not the encoder; the reported rate-distortion advantage would not follow from the described joint optimization. Please state the gradient estimator explicitly or present evidence that the encoder is trained by a different objective.
  2. [Section IV-A and Section III-A] The rate axis in Figs. 5 and 6 is not determined by the reported information. Section III-A defines the codeword as the concatenation of 64 indices selected from a codebook of size N_v, so the fixed rate is 64·log2 N_v bits only when N_v is a power of two (otherwise ceil(log2 N_v) per index). The value of N_v is never stated in Section IV, nor is it listed among hyperparameters in Section IV-A, so the operating points labeled 64, 128, 192, etc. cannot be verified, and the claim in Section IV-C3 that the proposed method needs fewer than 150 bits to reach −7 dB NMSE is not auditable. Report N_v or the resulting bit count for every curve point.
  3. [Section IV-C3, Fig. 6] The comparison in Fig. 6 uses baseline rate-distortion curves 'sourced directly from' [38], [39], [17]. The proposed method is evaluated on the cropped angular-delay domain (Section IV-C1), while it is not established that the published baseline curves use the same preprocessing, the same NMSE definition, the same rate convention (e.g., fixed-length versus entropy-coded bits, inclusion of side information), and the same train/test split. If any of these differ, the 'fewer than 150 bits versus more than 400 bits' margin in Section IV-C3, which is the basis for the abstract's 'over twofold performance improvement', could be an artifact of incompatible evaluation. Re-run the baselines under identical conditions or provide a point-by-point compatibility audit.
  4. [Sections IV-B3 and IV-C3] The rate-distortion plots report single curves with no error bars, and the text does not state the number of random seeds or the variance over test batches. Given that some headline margins in Fig. 5 are on the order of 0.5–1 dB and that diffusion decoding is stochastic, the authors should provide mean and variance across at least three training runs or explicitly state that the curves are deterministic and reproducible.
minor comments (5)
  1. [Section IV-B3] The phrase 'increasing the bit rate by over 100 bits at 132-bit rate' is inconsistent with the plotted rates in Fig. 5 (which appear to be 64, 128, and 192 bits); clarify which curve and operating point are meant.
  2. [Eq. (4b)] The symbol T is used both as the maximum diffusion step and as the uniformly sampled time index; use a separate symbol such as t for the sampled index in the expectation to avoid ambiguity.
  3. [Algorithm 1] Line 5 uses the already-quantized codeword c, while line 4 computes Lcb from cconti and e; clarifying whether gradient updates to the codebook embeddings are applied after the stop-gradient operations would help reproducibility.
  4. [Section IV-C1] The COST2100 dataset description gives only the number of training and test samples; report the antenna configuration, number of subcarriers, and the angular-delay cropping used for this dataset, since the numbers in Section IV-B1 apply to the Sionna setup only.
  5. [Section I (contributions)] The sentence 'But there may be challenges in the complexity of this approach' is awkwardly placed in the contribution bullet; move this caveat to the discussion in Section V.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the reported rate-distortion gains are measured on held-out test data with an independently specified training objective, and no equation or result reduces to its own input by construction.

full rationale

The paper's derivation chain is self-contained in the relevant sense: the encoder is a trainable codebook vector quantizer (Eq. (2), Algorithm 1) and the decoder is a conditional diffusion backward process (Eqs. (3)-(7)) whose loss is a standard variational upper-bound approximation; neither the codebook loss nor the diffusion loss is defined in terms of the NMSE benchmark curves used for evaluation. The reported rates are set by the number of transmitted indices (64) and the codebook size, not by fitting a parameter to the baseline curves. The only citation to the authors' own prior work ([19], H. Kim et al., 'Learning variable-rate codes for CSI feedback') appears in a background sentence on multi-rate CSI coding and is not load-bearing. The external-baseline comparisons in Sec. IV-C3 ('The baseline performance data are sourced directly from [38], [39], [17]') raise an auditability question if rate/NMSE conventions differ, but that is a measurement-validity concern, not a circularity, because those baseline numbers are not inputs to the proposed training or derivation. Similarly, the non-differentiable codebook selection in Algorithm 1 is a training-pipeline gap, not a circular step. No equation or claimed result is equivalent to its own input by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard diffusion and VQ machinery plus several domain assumptions about channel sparsity and UL-DL correlation. The most fragile item is the unstated gradient pathway through the discrete codebook, which is not a standard axiom but an implicit training assumption unique to this paper's formulation.

free parameters (4)
  • Codebook size N_v = Not reported
    The number of embedding vectors determines the codeword rate (64 indices times log2(N_v) bits). The paper never states N_v, so the rate values in Figs. 5 and 6 cannot be reproduced.
  • Codebook loss weight eta = 4.5e-4
    Hyperparameter chosen by hand in Eq. (7); the balance between codebook alignment and diffusion reconstruction affects the trained representation.
  • Number of denoising steps T = 4
    Choice of diffusion sampling steps; fewer steps reduce complexity but may change reconstruction quality and the rate-distortion tradeoff.
  • Number of quantized vectors = 64
    Encoder output 8x8x128 is quantized into 64 indices; this structural choice sets the code dimension.
assumptions (4)
  • standard math Diffusion variational upper bound (Eq. 4) and DDIM sampling (Eq. 6) are valid for the decoder.
    Taken from [33] and [32]; the paper does not re-derive them, it relies on them.
  • domain assumption The angular-delay representation is sparse and cropping to the first 32 delay taps preserves the information needed for the target NMSE.
    Section IV-B1 applies a 2D IFFT and keeps only 32 delay taps, assuming the remaining taps are negligible; the NMSE is evaluated only in the cropped domain.
  • domain assumption Uplink CSI is a useful side information for downlink CSI because of frequency-invariant channel characteristics, and perfect UL CSI is available at the base station.
    Section I and IV-B1 assume correlation between UL and DL and perfect UL acquisition; if reciprocity is weak, the side-information gain shrinks.
  • ad hoc to paper The encoder can be trained end-to-end through the non-differentiable codebook selection.
    Algorithm 1 updates theta_enc using L, but the codeword c is a discrete argmin; no straight-through or relaxation is described, making this an implicit and unverified training assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Diffusion Model-based Compression of MIMO CSI." pith.science (2026). https://pith.science/paper/5JDXVDR7

@misc{pith2026250303753,
  author       = {Pith},
  title        = {Pith review of: Generative Diffusion Model-based Compression of MIMO CSI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5JDXVDR7}},
  note         = {Machine review of arXiv:2503.03753}
}
read the original abstract

While neural lossy compression techniques have markedly advanced the efficiency of Channel State Information (CSI) compression and reconstruction for feedback in MIMO communications, efficient algorithms for more challenging and practical tasks-such as CSI compression for future channel prediction and reconstruction with relevant side information-remain underexplored, often resulting in suboptimal performance when existing methods are extended to these scenarios. To that end, we propose a novel framework for compression with side information, featuring an encoding process with fixed-rate compression using a trainable codebook for codeword quantization, and a decoding procedure modeled as a backward diffusion process conditioned on both the codeword and the side information. Experimental results show that our method significantly outperforms existing CSI compression algorithms, often yielding over twofold performance improvement by achieving comparable distortion at less than half the data rate of competing methods in certain scenarios. These findings underscore the potential of diffusion-based compression for practical deployment in communication systems.

Figures

Figures reproduced from arXiv: 2503.03753 by the authors.

Figure 1
Figure 1. System Model (Coding for computing with side [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proposed compression framework. The system’s optimization objective is to minimize the expected distortion as follows. minimize θ EpX,Y,Z [d(fdec(fenc(X; θenc), Y; θdec), Z)], (1) where pX,Y,Z denotes the joint probability distribution of the input source, the side information, and the target. For the distortion measure d, we consider the element-wise squared error. III. PROPOSED APPROACH In this section, we introdu… view at source ↗
Figure 4
Figure 4. Magnitude visualization of CSI samples: The task is to predict [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Rate-distortion curves for experiments in Sec. IV-B. [PITH_FULL_IMAGE:figures/full_fig_p005_5.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. Learning-to-Defer with Expert-Conditional Advice

    stat.ML 2026-03 unverdicted novelty 7.0 of 10

    The paper defines Learning-to-Defer with advice, proves inconsistency of separated surrogate losses, and introduces an augmented surrogate with H-consistency and excess-risk bounds that recovers the Bayes-optimal policy.

Reference graph

Works this paper leans on

41 extracted references · 30 canonical work pages · cited by 1 Pith paper

  1. [25]

    Lossy image compression with conditional diffusion models,

    R. Yang and S. Mandt, “Lossy image compression with conditional diffusion models,” in Adv. Neural Inf. Process. Syst. , vol. 36, 2024

  2. [31]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyals et al. , “Neural discrete representation learning,” in Adv. Neural Inf. Process. Syst. , vol. 30, 2017, pp. 1–11

  3. [38]

    An Efficient Network with Novel Quantization Designed for Massive MIMO CSI Feedback

    X. Sun, Z. Zhang, and L. Yang, “An efficient network with novel quantization designed for massive MIMO CSI feedback,” arXiv preprint arXiv:2405.20068 , 2024, [Online]. Available: https://arxiv.org/abs/2405.20068

  4. [39]

    Convolutional neural network- based multiple-rate compressive sensing for massive MIMO CSI feedback: Design, simulation, and analysis,

    J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Convolutional neural network- based multiple-rate compressive sensing for massive MIMO CSI feedback: Design, simulation, and analysis,” IEEE Trans. Wireless Commun., vol. 19, no. 4, pp. 2827–2840, 2020

  5. [17]

    Deep autoencoder-based massive MIMO CSI feedback with quantization and entropy coding,

    S. Ravula and S. Jain, “Deep autoencoder-based massive MIMO CSI feedback with quantization and entropy coding,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , 2021, pp. 1–6

  6. [1]

    Overview of deep learning- based CSI feedback in massive MIMO systems,

    J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Overview of deep learning- based CSI feedback in massive MIMO systems,” IEEE Trans. Commun. , vol. 70, no. 12, pp. 8017–8045, 2022

  7. [2]

    Dahlman, S

    E. Dahlman, S. Parkvall, and J. Skold, 4G: LTE/LTE-Advanced for Mobile Broadband. Cambridge, MA, USA: Academic Press, 2013

  8. [3]

    Zaidi, F

    A. Zaidi, F. Athley, J. Medbo, U. Gustavsson, G. Durisi, and X. Chen, 5G Physical Layer: Principles, Models and Technology Components . Cambridge, MA, USA: Academic Press, 2018

Show all 41 references
  1. [4]

    End-to-end optimized image compression,

    J. Ballé, V . Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2017

  2. [5]

    Variational image compression with a scale hyperprior,

    J. Ballé, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Variational image compression with a scale hyperprior,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2018, pp. 1–23

  3. [6]

    Lossy image compression with compressive autoencoders,

    L. Theis, W. Shi, A. Cunningham, and F. Huszár, “Lossy image compression with compressive autoencoders,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2017, pp. 1–19

  4. [7]

    An introduction to neural data compression,

    Y . Yang, S. Mandt, L. Theis et al. , “An introduction to neural data compression,” F ound. Trends Comput. Graph. Vis. , vol. 15, no. 2, pp. 113–200, 2023

  5. [8]

    The JPEG still picture compression standard,

    G. K. Wallace, “The JPEG still picture compression standard,” Commun. ACM, vol. 34, no. 4, pp. 30–44, 1991

  6. [9]

    On universal quantization,

    J. Ziv, “On universal quantization,” IEEE Trans. Inf. Theory , vol. 31, no. 3, pp. 344–347, 1985

  7. [10]

    Joint autoregressive and hierarchical priors for learned image compression,

    D. Minnen, J. Ballé, and G. D. Toderici, “Joint autoregressive and hierarchical priors for learned image compression,” in Adv. Neural Inf. Process. Syst., vol. 31, 2018

  8. [11]

    Deep learning for massive MIMO CSI feedback,

    C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive MIMO CSI feedback,” IEEE Wireless Commun. Lett. , vol. 7, no. 5, pp. 748–751, 2018

  9. [12]

    Deep learning-based CSI feedback approach for time-varying massive MIMO channels,

    T. Wang, C.-K. Wen, S. Jin, and G. Y . Li, “Deep learning-based CSI feedback approach for time-varying massive MIMO channels,” IEEE Wireless Commun. Lett. , vol. 8, no. 2, pp. 416–419, 2018

  10. [13]

    A novel CSI feedback approach for massive MIMO using LSTM-attention CNN,

    Q. Li, A. Zhang, P. Liu, J. Li, and C. Li, “A novel CSI feedback approach for massive MIMO using LSTM-attention CNN,” IEEE Access , vol. 8, pp. 7295–7302, 2020

  11. [14]

    HyperRNN: Deep learning-aided downlink CSI acquisition via partial channel reciprocity for FDD massive MIMO,

    Y . Liu and O. Simeone, “HyperRNN: Deep learning-aided downlink CSI acquisition via partial channel reciprocity for FDD massive MIMO,” in Proc. IEEE Int. Workshop Signal Process. Adv. Wireless Commun. (SPA WC), 2021, pp. 31–35

  12. [15]

    Multi-resolution CSI feedback with deep learning in massive MIMO system,

    Z. Lu, J. Wang, and J. Song, “Multi-resolution CSI feedback with deep learning in massive MIMO system,” in Proc. IEEE Int. Conf. Commun. (ICC), 2020, pp. 1–6

  13. [16]

    Multi-rate variable-length CSI compression for FDD massive MIMO,

    B. Park, H. Do, and N. Lee, “Multi-rate variable-length CSI compression for FDD massive MIMO,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP) , 2024, pp. 7715–7719

  14. [18]

    Machine learning-based CSI feedback with variable length in FDD massive MIMO,

    M. Nerini, V . Rizzello, M. Joham, W. Utschick, and B. Clerckx, “Machine learning-based CSI feedback with variable length in FDD massive MIMO,” IEEE Trans. Wireless Commun. , vol. 22, no. 5, pp. 2886–2900, 2022

  15. [19]

    Learning variable-rate codes for CSI feedback,

    H. Kim, H. Kim, and G. De Veciana, “Learning variable-rate codes for CSI feedback,” in Proc. IEEE Global Commun. Conf. (GLOBECOM) , 2022, pp. 1435–1441

  16. [20]

    An overview of 5G advanced evolution in 3GPP release 18,

    X. Lin, “An overview of 5G advanced evolution in 3GPP release 18,” IEEE Commun. Stand. Mag. , vol. 6, no. 3, pp. 77–83, 2022

  17. [21]

    Diffusion models: A comprehensive survey of methods and applications,

    L. Yang, Z. Zhang, Y . Song, S. Hong, R. Xu, Y . Zhao, W. Zhang, B. Cui, and M.-H. Yang, “Diffusion models: A comprehensive survey of methods and applications,” ACM Comput. Surv. , vol. 56, no. 4, pp. 1–39, 2023

  18. [22]

    U-Net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent. (MICCAI) . Springer, 2015

  19. [23]

    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. Intell. , vol. 45, no. 9, pp. 10 850–10 869, 2023

  20. [24]

    Lossy compression with gaussian diffusion,

    L. Theis, T. Salimans, M. D. Hoffman, and F. Mentzer, “Lossy compression with gaussian diffusion,” arXiv preprint arXiv:2206.08889 , 2022, [Online]. Available: https://arxiv.org/abs/2206.08889

  21. [26]

    Generalization of the rate-distortion function for wyner-ziv coding of noisy sources in the quadratic-gaussian case,

    D. Rebollo-Monedero and B. Girod, “Generalization of the rate-distortion function for wyner-ziv coding of noisy sources in the quadratic-gaussian case,” in Data Compression Conference . IEEE, 2005, pp. 23–32

  22. [27]

    Eliminating channel feedback in next-generation cellular networks,

    D. Vasisht, S. Kumar, H. Rahul, and D. Katabi, “Eliminating channel feedback in next-generation cellular networks,” in Proc. ACM SIGCOMM Conf., 2016, pp. 398–411

  23. [28]

    FDD massive MIMO without CSI feedback,

    D. Han, J. Park, and N. Lee, “FDD massive MIMO without CSI feedback,” IEEE Trans. Wireless Commun. , vol. 23, no. 5, pp. 4518–4530, 2023

  24. [29]

    NR; physical layer procedures for data (release 15),

    3GPP, “NR; physical layer procedures for data (release 15),” 3GPP, Tech. Rep. TR 38.214, 12 2017

  25. [30]

    Sionna: An open-source library for next-generation physical layer research,

    J. Hoydis, S. Cammerer, F. A. Aoudia, A. Vem, N. Binder, G. Marcus, and A. Keller, “Sionna: An open-source library for next-generation physical layer research,” arXiv preprint arXiv:2203.11854, 2022, [Online]. Available: https://arxiv.org/abs/2203.11854

  26. [32]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2022, pp. 1–22

  27. [33]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Adv. Neural Inf. Process. Syst. , vol. 33, 2020, pp. 6840–6851

  28. [34]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in Proc. Int. Conf. Learn. Represent. (ICLR) , 2022

  29. [35]

    Adam: A method for stochastic opti- mization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic opti- mization,” arXiv preprint arXiv:1412.6980 , 2014, [Online]. Available: https://arxiv.org/abs/1412.6980

  30. [36]

    Study on channel model for frequencies from 0.5 to 100 GHz,

    3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” Technical Specification Group Radio Access Network TR 38.901, 2019

  31. [37]

    Spatial- and frequency- wideband effects in millimeter-wave massive MIMO systems,

    B. Wang, F. Gao, S. Jin, H. Lin, and G. Y . Li, “Spatial- and frequency- wideband effects in millimeter-wave massive MIMO systems,” IEEE Trans. Signal Process. , vol. 66, no. 13, pp. 3393–3406, 2018

  32. [40]

    Malunet: A multi-attention and light-weight unet for skin lesion segmentation,

    J. Ruan, S. Xiang, M. Xie, T. Liu, and Y . Fu, “Malunet: A multi-attention and light-weight unet for skin lesion segmentation,” in 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) . IEEE, 2022, pp. 1150–1156

  33. [41]

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

    X. Liu, C. Gong et al., “Flow straight and fast: Learning to generate and transfer data with rectified flow,” in Proc. Int. Conf. Learn. Represent. (ICLR), 2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.