Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Latent Diffusion Model Based Denoising Receiver for 6G Semantic Communication: From Stochastic Differential Theory to Application

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A closed-form formula maps channel SNR to the optimal diffusion timestep for semantic communication.

desk verdict Plausible system idea, but the closed-form SNR-to-timestep mapping rests on an incorrect derivation and does not correspond to any real LDM schedule. read the letter →

arxiv 2506.05710 v3 pith:FLR23ZXZ submitted 2025-06-06 cs.LG cs.ITcs.SYeess.SYmath.IT

Xiucheng Wang , Honggang Jia , Nan Cheng This is my paper
classification cs.LGcs.ITcs.SYeess.SYmath.IT
keywords semanticcommunicationlatentdiffusionmodelSNR-to-timestepmappingstochasticdifferentialequationszero-shotdenoisingout-of-distributionrobustnessjointsource-channelcoding6Gwireless
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 show that a pretrained latent diffusion model, with no fine-tuning, can serve as the denoising receiver in a semantic communication system, and that the only information the receiver needs is the channel signal-to-noise ratio. The theoretical centerpiece is a closed-form mapping from SNR to the diffusion timestep $t$. When the received latent feature is rescaled by the companion factor $\alpha$, its second-order statistics match the diffusion model's training inputs, so reverse diffusion starting from that rescaled observation acts as the optimal denoiser. If this is right, neural semantic systems no longer need to be retrained for each noise level or data distribution: a generic generative prior replaces the task-specific channel decoder. The paper supports the claim with image-transmission experiments showing the zero-shot receiver outperforms learned joint source-channel coding baselines, especially at low SNR and on out-of-distribution content.

What carries the argument

The carrying object is the SNR-matching identity $g(t;\phi)=(1-t)^2-\phi t$, whose unique root in $(0,1]$ is the closed-form timestep $t^\star=(2+\phi-\sqrt{\phi^2+4\phi})/2$. It translates the channel condition (measured through $\phi$) into the diffusion schedule of a pretrained latent diffusion model, so the receiver knows where along the reverse trajectory to start. The companion object is the scalar scaling factor $\alpha$ from Theorem 2, which equalizes $\mathbb{E}[\|\alpha y\|^2]$ and $\mathbb{E}[\|x_t\|^2]$, making the rescaled received feature statistically indistinguishable, to second order, from the diffusion model's training-time inputs. Together, $t^\star$ and $\alpha$ let an unconditional pretrained diffusion model act as a drop-in denoiser: it sees $\alpha y$ as if it were a forward-diffused sample at timestep $t^\star$, and its reverse process removes the channel noise.

What would settle it

Sweep the denoising timestep $t$ on a fixed pretrained latent diffusion model at a fixed channel SNR, keeping $\alpha$ fixed to its Theorem 2 value, and locate the empirically best $t$ under RMSE, PSNR, or LPIPS. If that optimum differs from the Theorem 1 root $t=(2+\phi-\sqrt{\phi^2+4\phi})/2$ by more than the $\pm5\%$ stability margin the paper reports, the closed-form mapping is not the optimal timestep. A stronger test would use a diffusion model with a known schedule other than $x_t=(1-t)x_0+\sqrt{t}\epsilon$ and check whether the formula still gives the best reconstruction at each SNR.

Watch

Extended reading notes

Core claim

The central claim is that channel denoising and diffusion denoising are the same statistical operation once the receiver matches their signal-to-noise ratios. Writing the diffusion forward process as $x_t=(1-t)x_0+\sqrt{t}\epsilon$ with $\epsilon\sim\mathcal{N}(0,I)$, the diffusion variable at timestep $t$ carries signal energy $(1-t)^2\gamma$ and noise energy $t$, where $\gamma=\mathbb{E}[\|x_0\|^2]$. The channel output $y=\eta z+n$ carries SNR $(\mathbb{E}[\|y\|^2]-\sigma^2)/\sigma^2$. Equating these two ratios yields the quadratic $(1-t)^2=\phi t$ with $\phi=(\mathbb{E}[\|y\|^2]-\sigma^2)/(\gamma\sigma^2)$, whose admissible root is $t=(2+\phi-\sqrt{\phi^2+4\phi})/2$ (Theorem 1). Theorem 2 then gives $\alpha=\sqrt{\big((1-t)^2\gamma+t\big)/\mathbb{E}[\|y\|^2]}$, so that $\alpha y$ has the same mean energy as $x_t$. The paper claims that feeding $\alpha y$ into a pretrained latent diffusion model and running reverse diffusion for $t$ steps therefore yields maximum-a-posteriori-style recovery of the clean latent $z$, without any training.

Load-bearing premise

The closed-form timestep rests on assuming the diffusion forward process is exactly $x_t=(1-t)x_0+\sqrt{t}\epsilon$ with unit-variance noise, and that the channel noise power is $\sigma^2$ in every latent dimension; if a pretrained model's actual schedule differs, or the latent dimension changes the noise energy, the formulas for $t$ and $\alpha$ are not the correct calibration.

Editorial extensions

If this is right

  • A receiver can set its denoising schedule from two measured quantities—received power and noise variance—without any SNR-specific training or model selection.
  • The same pretrained diffusion prior works zero-shot at very low SNR (down to $-20$ dB in the paper's experiments), where conventional JSCC decoders collapse, by preserving high-level semantics even when fine details are lost.
  • Because the diffusion model is decoupled from the encoder–decoder pair, replacing it with a more powerful pretrained model improves the whole communication system without retraining the transceiver.
  • The closed-form parameters are robust to small estimation errors: the paper reports that perturbing $t$ or $\alpha$ by $\pm5\%$ leaves performance nearly unchanged, so coarse SNR estimation suffices.
  • Out-of-distribution content (landscapes, animals) is handled by the generative prior rather than by the task-specific decoder, avoiding the overfitting artifacts seen in discriminative JSCC baselines.

Reading between the lines

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

  • Beyond the paper: the same SNR-matching calibration should apply to any denoiser defined by an affine forward process $x_t=a_t x_0+b_t\epsilon$, not only latent diffusion models; the quadratic would simply be replaced by $a_t^2\gamma=\phi b_t^2$.
  • Beyond the paper: the formula's validity is tied to the assumed schedule $x_t=(1-t)x_0+\sqrt{t}\epsilon$; for a pretrained model trained with a different schedule (e.g., the standard DDPM variance-preserving form), a corrected $t$ would need to be re-derived, and it is an open question how much the mismatch degrades performance.
  • Beyond the paper: the framework suggests a generic recipe for using generative priors as blind channel denoisers in other inverse problems—if the noise is Gaussian and the forward operator is known, the same two-step matching (timestep then scale) could be used for compressive sensing or deblurring without task-specific training.
  • Beyond the paper: the OOD gains depend on the breadth of the pretrained model's training distribution; for data far outside that distribution (e.g., medical or satellite imagery), the zero-shot claim would need to be re-tested, and the method might need a domain-adapted prior.
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. The paper proposes a latent diffusion model (LDM)-based denoising receiver for semantic communication. The system transmits VAE-encoded latent vectors over an AWGN channel and uses a pretrained LDM at the receiver to denoise the corrupted latent. The central theoretical contribution is a claimed closed-form mapping between the channel SNR and an 'optimal' diffusion timestep, obtained by matching second-order signal-to-noise ratios between the diffusion variable and the channel output, along with a scaling factor that aligns the received vector's distribution with the LDM's training distribution. Experiments compare the method with Deep JSCC, Swin JSCC, and a Stable Diffusion baseline on CelebA-HQ and out-of-distribution images, reporting superior pixel-level and perceptual metrics without fine-tuning.

Significance. If the theoretical claims were sound, the paper would provide a principled, training-free mechanism for adapting pretrained diffusion models to varying channel conditions, which is practically valuable for 6G semantic communication. The empirical results suggest that the SNR-dependent timestep heuristic can work well in practice, and the perturbation studies in Section V.D give some evidence that the computed t and alpha are reasonable operating points. However, the theoretical foundation contains load-bearing errors: the simplification to Eq. (17) is algebraically wrong as stated, the noise-energy accounting ignores the latent dimension d, and the linear schedule used for the derivation is not the schedule of the pretrained LDM used in Section V. These issues invalidate the claimed closed-form optimal timestep for pretrained LDMs, so the paper's main contribution is not established.

major comments (4)
  1. [§IV.A, Eqs. (14)–(17)] The simplification leading to Eq. (17) is not correct. Substituting h_t = x0 into (14) gives ∫_0^t h_s ds = t x0 and hence xt = (1+t)x0 + √t ε, not (1-t)x0. To obtain (1-t)x0 one must instead set h_t = -x0, which contradicts the sentence 'by setting the ht = x0'. This sign error is load-bearing because Eqs. (17) and (18) are the basis of Theorems 1 and 2.
  2. [§IV.B, Theorem 1 and Eq. (21)] The proof of Theorem 1 uses E[||n||^2] = σ^2 for n ~ N(0, σ^2 I) in R^d. The correct value is d σ^2. Consequently Eq. (21) and the definition φ = (E[||y||^2] - σ^2)/(γσ^2) are only correct for d = 1; for the vector latents of an LDM the estimated φ does not equal the true SNR, so the closed-form t is miscalibrated. The same omission of d appears in the proof of Theorem 2, where E[||√t ε||^2] is taken as t instead of t·d. If the authors intend σ^2 as per-dimension variance, the total noise power in the SNR ratio must be dσ^2; if they intend σ^2 as total power, Eq. (8) should specify covariance (σ^2/d)I.
  3. [§IV.B, Theorem 1 and Remark 2] Theorem 1 establishes only that a t satisfying the second-moment matching condition (1-t)^2 γ = φ t exists; it does not prove that this t minimizes any end-to-end distortion metric. The abstract and introduction nevertheless call t 'optimal'. The empirical perturbation study in Section V.D is supportive evidence, but it is not a proof. The theorem should be reworded to claim an 'SNR-matching timestep' unless a formal optimality argument with respect to a reconstruction objective is supplied.
  4. [§IV.A–§IV.C, Eq. (17) vs. Eq. (3)] The mapping is derived for the linear attenuation schedule xt = (1-t)x0 + √t ε. Standard LDMs use the marginal xt = √ᾱ_t x0 + √(1-ᾱ_t) ε with ᾱ_t determined by a specific noise schedule, as the paper itself states in Eq. (3). The t computed from Theorem 1 is therefore not the diffusion timestep of a pretrained LDM, and the paper provides no conversion between its SNR-matched t and the actual ᾱ_t of the model used in Section V. Feeding the scaled observation as 'xt' at this t is unjustified; this disconnect invalidates the central claim that Eq. (19) gives the correct timestep for a pretrained LDM.
minor comments (5)
  1. [§II.B, Eq. (1)] The text describing the Gaussian distribution contains typographical errors: 'the mean if µ' should be 'the mean µ', and the covariance in Eq. (1) is β_t I, not σI.
  2. [§VI and §IV.C headings] The heading 'Sematic Communciation' should be 'Semantic Communication'; the same typo appears in the Section IV.C heading.
  3. [§IV.B, Theorem 1] The definition of φ is typeset ambiguously; it should be written as φ = (E[||y||^2] - σ^2)/(γσ^2).
  4. [§IV.B, Remark 2, Eq. (30)] Equation (30) is malformed and dimensionally inconsistent as printed: the terms mix σ^2 and η without a clear derivation. This expression should be re-derived carefully after correcting the issues in Theorem 1.
  5. [§V.A] The experiments should specify the exact pretrained latent diffusion model (architecture and noise schedule) used for the proposed method; the current text only says it was trained on CelebA-HQ and ImageNet, which is also ambiguous relative to the Stable Diffusion baseline.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'optimal' SNR-to-timestep mapping reduces to solving the SNR-matching equation that defines optimality, and the simplified forward schedule on which it rests is imported by ansatz from a cited paper rather than from the pretrained model.

  1. self definitional [Section IV-B, Theorem 1, Eq. (19)-(22)]
    "we propose that the optimal denoising performance can be achieved when the energy ratio between signal and noise in xt is matched to that in y ... The t can be calculated as Theorem 1 ... By introducing ϕ ≜ SNRobs/γ >0, the matching condition of (20) can be obtained as follows. (1 − t)2 = ϕt. (22)"

    Theorem 1's t formula is obtained by solving Eq. (22), which is exactly the SNR-matching condition in Eq. (20) after dividing by γ. The theorem therefore proves existence of a timestep satisfying the matching equation, while the word 'optimal' is imported from the paper's unproved proposal that matching the signal-noise energy ratio is optimal. No independent optimality argument is given. The numerical experiment in Section V-D perturbs t around this same root and shows degradation, which is a local consistency check of the defining equation rather than external confirmation of optimality.

  2. ansatz smuggled in via citation [Section IV-A, Eq. (17) (transition from Eq. (14))]
    "Remarkably, since the ht holds R 1 0 ht + x0 = 0, similar to [35] by setting the ht = x0, a simplified expression of (14) and (16) can be obtained as follows . xt = (1 − t)x0 + √tϵ, (17)"

    Equation (17) is not derived from the stated SDE: substituting ht=x0 into (14) yields x0(1+t)+√tϵ, so (17) requires ht=-x0, contradicting the text. The schedule is imported by ansatz via citation [35], whereas the pretrained LDM used in the experiments follows the DDPM marginal (3), xt=√ᾱt x0+√(1−ᾱt)ϵ. Consequently the t computed from Theorem 1 is a corruption level of the toy process, not of the deployed model, and the claimed closed-form mapping is not a prediction from the pretrained model's actual forward dynamics.

full rationale

The central claimed derivation is partially self-definitional. Theorem 1's closed-form t is the explicit solution of (1−t)^2=ϕt, which is a rearrangement of the SNR-matching condition that the paper itself proposes as the definition of optimal denoising. Thus the 'closed-form mapping between SNR and the optimal denoising timestep' is algebra on an assumed criterion rather than an independently derived optimality result. In addition, the simplified forward process (17) that underlies the whole mapping is imported from [35] by ansatz and is inconsistent with both the preceding SDE (14) and the actual DDPM schedule (3) of the pretrained LDM used in experiments. These issues mean one central 'prediction' reduces by construction. However, the experimental comparison against Deep JSCC, Swin JSCC, and SD baselines is external and provides real engineering evidence for the proposed zero-shot receiver architecture, and no load-bearing self-citation chain was found. Score 6 reflects partial circularity in the theoretical mapping while the system evaluation retains substantive independent content.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The central claim depends on a simplified diffusion SDE, an unproved optimality criterion, and a dimensionally inconsistent noise-energy accounting. There are no new physical entities. The mapping uses one data statistic gamma and the channel noise variance sigma^2 as inputs.

free parameters (1)
  • gamma = E[||x0||^2]
    Appears in Theorems 1 and 2 and in the timestep and scaling formulas; the paper assumes it is known at training time but does not report its value or normalization.
assumptions (6)
  • standard math The reverse-time SDE and DDPM theory from [19] and [33] is accepted as background.
    Used to justify the diffusion denoising framework; this is appropriate background and is not invented by the paper.
  • domain assumption The pretrained latent diffusion forward process is exactly xt=(1-t)x0+sqrt(t)epsilon (Eq. 17).
    Inherited from [35] and used to derive Theorems 1 and 2; the paper does not verify that the actual LDM checkpoint follows this schedule, and the derivation from Eq. (14) by setting ht=x0 has a sign error.
  • ad hoc to paper Matching the second-order signal-to-noise ratio of the diffusion variable to that of the channel observation gives the optimal denoising timestep.
    Theorems 1 and 2 establish an energy equality, not optimality; the claim that this t is optimal is assumed without proof.
  • domain assumption The channel noise and diffusion noise have total variance independent of latent dimension (E||n||^2=sigma^2, E||epsilon||^2=1).
    Used throughout the proofs of Theorems 1 and 2; for d-dimensional latents, E||n||^2=d sigma^2 and E||epsilon||^2=d, so the formulas are not valid as written.
  • domain assumption The latent signal energy gamma is known and normalized so that the timestep mapping depends only on sigma^2.
    Remark 2 assumes normalized transmit power; this is standard in diffusion training but is not stated as an explicit channel-model assumption.
  • ad hoc to paper The reverse update in Eq. (18) is a valid discretization of the reverse diffusion process.
    Used for the multi-step run and for the single-step approximation; no derivation or comparison with the actual learned score is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Latent Diffusion Model Based Denoising Receiver for 6G Semantic Communication: From Stochastic Differential Theory to Application." pith.science (2026). https://pith.science/paper/FLR23ZXZ

@misc{pith2026250605710,
  author       = {Pith},
  title        = {Pith review of: Latent Diffusion Model Based Denoising Receiver for 6G Semantic Communication: From Stochastic Differential Theory to Application},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLR23ZXZ}},
  note         = {Machine review of arXiv:2506.05710}
}
read the original abstract

In this paper, a novel semantic communication framework empowered by generative artificial intelligence (GAI) is proposed, to enhance the robustness against both channel noise and transmission data distribution shifts. A theoretical foundation is established using stochastic differential equations (SDEs), from which a closed-form mapping between any signal-to-noise ratio (SNR) and the optimal denoising timestep is derived. Moreover, to address distribution mismatch, a mathematical scaling method is introduced to align received semantic features with the training distribution of the GAI. Built on this theoretical foundation, a latent diffusion model (LDM)-based semantic communication framework is proposed that combines a variational autoencoder for semantic features extraction, where a pretrained diffusion model is used for denoising. The proposed system is a training-free framework that supports zero-shot generalization, and achieves superior performance under low-SNR and out-of-distribution conditions, offering a scalable and robust solution for future 6G semantic communication systems. Experimental results demonstrate that the proposed semantic communication framework achieves state-of-the-art performance in both pixel-level accuracy and semantic perceptual quality, consistently outperforming baselines across a wide range of SNRs and data distributions without any fine-tuning or post-training.

Figures

Figures reproduced from arXiv: 2506.05710 by the authors.

Figure 1
Figure 1. Illustration of the proposed semantic communication framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the denoising timestep calculation. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the scaling method. Remark 2. In the training procedure, since the purpose of denoise is to solve z, thus if the channel fading factor η can be obtained through some channel estimation methods, the x0 is set equal to z. Therefore, the t can be simplified as follows. t = 2σ 2 + η − p η 2 + 4η 2σ 2 2σ 2 . (30) The absence of an explicit z term in Remark 2 follows directly from the standard training pro… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Performance evaluation of different methods on various SNRs in the CelebA-HQ datasets. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison of image transmission performance at different SNR. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The image transmission performance of the proposed method on low SNRs. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Performance demonstration on OOD data. A key advantage of our architecture is its modularity: the LDM component can be readily replaced with publicly available weights trained on large-scale datasets such as ImageNet or OpenImages. Without any task-specific fine-tuning…
Figure 8
Figure 8. Figure 8: Performance evaluation on different t [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Performance evaluation on different α. both SNR degradation and distributional shift underscores the practical utility and scalability of our approach for next￾generation communication systems. D. Numerical Proof of Theorem 1 and 2 To substantiate the theoretical claim…

Discussion (0). Sign in 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. Inclusion Arena: An Open Platform for Evaluating Large Foundation Models with Real-World Apps

    cs.AI 2025-08 unverdicted novelty 4.0 of 10

    Inclusion Arena ranks large AI models by embedding pairwise comparisons into real app usage and stabilizing the ratings with placement matches and proximity sampling.

Reference graph

Works this paper leans on

41 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    What should 6G be?

    S. Dang, O. Amin, B. Shihada, and M.-S. Alouini, “What should 6G be?” Nature Electron., vol. 3, no. 1, pp. 20–29, 2020

    work page 2020

  2. [2]

    Intellicise wireless networks from semantic communications: A survey, research issues, and challenges,

    P. Zhang, W. Xu, Y . Liu, X. Qin, K. Niu, S. Cui, G. Shi, Z. Qin, X. Xu, F. Wang et al. , “Intellicise wireless networks from semantic communications: A survey, research issues, and challenges,” IEEE Commun. Surv. Tutorials, 2024

    work page 2024

  3. [3]

    Toward immersive communications in 6G,

    X. Shen, J. Gao, M. Li, C. Zhou, S. Hu, M. He, and W. Zhuang, “Toward immersive communications in 6G,”Front. Comput. Sci, vol. 4, p. 1068478, 2023

    work page 2023

  4. [4]

    Industrial internet of things: Challenges, opportunities, and directions,

    E. Sisinni, A. Saifullah, S. Han, U. Jennehag, and M. Gidlund, “Industrial internet of things: Challenges, opportunities, and directions,” IEEE Trans Ind. Informat., vol. 14, no. 11, pp. 4724–4734, 2018

    work page 2018

  5. [5]

    Toward enhanced reinforcement learning-based resource management via digital twin: Opportunities, applications, and challenges,

    N. Cheng, X. Wang, Z. Li, Z. Yin, T. H. Luan, and X. Shen, “Toward enhanced reinforcement learning-based resource management via digital twin: Opportunities, applications, and challenges,” IEEE Network, vol. 39, no. 1, pp. 189–196, 2025

    work page 2025

  6. [6]

    Semantic communications: Principles and challenges,

    Z. Qin, X. Tao, J. Lu, W. Tong, and G. Y . Li, “Semantic communications: Principles and challenges,” arXiv preprint arXiv:2201.01389 , 2021

    arXiv 2021

  7. [7]

    The road towards 6g: A comprehensive survey,

    W. Jiang, B. Han, M. A. Habibi, and H. D. Schotten, “The road towards 6g: A comprehensive survey,” IEEE Open J. Commun. Soc. , vol. 2, pp. 334–366, 2021

    work page 2021

  8. [8]

    Joint source–channel coding: Fundamentals and recent progress in practical designs,

    D. G ¨und¨uz, M. A. Wigger, T.-Y . Tung, P. Zhang, and Y . Xiao, “Joint source–channel coding: Fundamentals and recent progress in practical designs,” Proc. IEEE, 2024

    work page 2024

Show all 41 references
  1. [9]

    Semantic communication: A survey of its theoretical development,

    G. Xin, P. Fan, and K. B. Letaief, “Semantic communication: A survey of its theoretical development,” Entropy, vol. 26, no. 2, p. 102, 2024

    2024

  2. [10]

    Deep learning in physical layer: Review on data driven end-to-end communication systems and their enabling semantic applications,

    N. Islam and S. Shin, “Deep learning in physical layer: Review on data driven end-to-end communication systems and their enabling semantic applications,” IEEE Open J. Commun. Soc. , 2024

    2024

  3. [11]

    V . N. Vapnik, V . Vapniket al., Statistical learning theory . Wiley New York, 1998

    1998

  4. [12]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, and A. Courville, Deep learning. MIT press, 2016

    2016

  5. [13]

    Commu- nication algorithms via deep learning,

    H. Kim, Y . Jiang, R. Rana, S. Kannan, S. Oh, and P. Viswanath, “Commu- nication algorithms via deep learning,” arXiv preprint arXiv:1805.09317 , 2018

    2018 arXiv

  6. [14]

    Deep joint source- channel coding for wireless image transmission,

    E. Bourtsoulatze, D. B. Kurka, and D. G ¨und¨uz, “Deep joint source- channel coding for wireless image transmission,” IEEE Trans. Cognit. Commun. Networking., vol. 5, no. 3, pp. 567–579, 2019

    2019

  7. [15]

    Joint source–channel codes for mimo block- fading channels,

    D. Gunduz and E. Erkip, “Joint source–channel codes for mimo block- fading channels,” IEEE Trans. Inf. Theory , vol. 54, no. 1, pp. 116–134, 2008

    2008

  8. [16]

    Diffusion models in vision: A survey,

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

    2023

  9. [17]

    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

  10. [18]

    Auto-encoding variational bayes,

    D. P. Kingma, M. Welling et al. , “Auto-encoding variational bayes,” 2013

    2013

  11. [19]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” arXiv preprint arXiv:2011.13456 , 2020

    2011 arXiv

  12. [20]

    Simultaneous image-to-zero and zero-to-noise: Diffusion models with analytical image attenuation,

    Y . Huang, Z. Qin, X. Liu, and K. Xu, “Simultaneous image-to-zero and zero-to-noise: Diffusion models with analytical image attenuation,” 2024

    2024

  13. [21]

    Semantic communications: Overview, open issues, and future research directions,

    X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications: Overview, open issues, and future research directions,” IEEE Wireless Commun., vol. 29, no. 1, pp. 210–219, 2022

    2022

  14. [22]

    Semantics-empowered communications: A tutorial-cum-survey,

    Z. Lu, R. Li, K. Lu, X. Chen, E. Hossain, Z. Zhao, and H. Zhang, “Semantics-empowered communications: A tutorial-cum-survey,” IEEE Commun. Surv. Tutorials, vol. 26, no. 1, pp. 41–79, 2024

    2024

  15. [23]

    A contemporary survey on semantic communications: Theory of mind, generative ai, and deep joint source-channel coding,

    L. X. Nguyen, A. D. Raha, P. S. Aung, D. Niyato, Z. Han, and C. S. Hong, “A contemporary survey on semantic communications: Theory of mind, generative ai, and deep joint source-channel coding,” arXiv preprint arXiv:2502.16468, 2025

    2025

  16. [24]

    Semantic communication empowered 6G networks: Techniques, applications, and challenges,

    Y . Wang, H. Han, Y . Feng, J. Zheng, and B. Zhang, “Semantic communication empowered 6G networks: Techniques, applications, and challenges,” IEEE Access, 2025

    2025

  17. [25]

    Semantic edge computing and semantic communications in 6G networks: A unifying survey and research challenges,

    M. Zhang, M. Abdi, V . R. Dasari, and F. Restuccia, “Semantic edge computing and semantic communications in 6G networks: A unifying survey and research challenges,” arXiv preprint arXiv:2411.18199 , 2024

    2024 arXiv

  18. [26]

    A survey on semantic communications in internet of vehicles,

    S. Ye, Q. Wu, P. Fan, and Q. Fan, “A survey on semantic communications in internet of vehicles,” Entropy, vol. 27, no. 4, p. 445, 2025

    2025

  19. [27]

    Modeling and performance analysis for semantic communications based on empirical results,

    S. Ma, B. Shen, C. Zhang, Y . Wu, H. Li, S. Li, G. Shi, and N. Al-Dhahir, “Modeling and performance analysis for semantic communications based on empirical results,” arXiv preprint arXiv:2504.21055 , 2025

    2025 arXiv

  20. [28]

    Generative adversarial networks: An overview,

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

    2018

  21. [29]

    Diffbir: Towards blind image restoration with generative diffusion prior,

    X. Lin, J. He, Z. Chen, Z. Lyu, B. Fei, B. Dai, W. Ouyang, Y . Qiao, and C. Dong, “Diffbir: Towards blind image restoration with generative diffusion prior,” arXiv preprint arXiv:2308.15070 , 2023

    2023 arXiv

  22. [30]

    Structured denoising diffusion models in discrete state-spaces,

    J. Austin, D. D. Johnson, J. Ho, D. Tarlow, and R. Van Den Berg, “Structured denoising diffusion models in discrete state-spaces,” Advances in Neural Information Processing Systems (NeurIPS), vol. 34, pp. 17 981– 17 993, 2021

    2021

  23. [31]

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

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

    2025

  24. [32]

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

    X. Wang, Z. Fang, and N. Cheng, “Radiodiff-inverse: Diffusion enhanced bayesian inverse estimation for isac radio map construction,” arXiv preprint arXiv:2504.14298, 2025

    2025 arXiv

  25. [33]

    Denoising diffusion probabilistic models,

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

    2020

  26. [34]

    Stimulating diffusion model for image denoising via adaptive embedding and ensembling,

    T. Li, H. Feng, L. Wang, L. Zhu, Z. Xiong, and H. Huang, “Stimulating diffusion model for image denoising via adaptive embedding and ensembling,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 46, no. 12, 2024

    2024

  27. [35]

    Decoupled diffusion models: Simultaneous image to zero and zero to noise,

    Y . Huang, Z. Qin, X. Liu, and K. Xu, “Decoupled diffusion models: Simultaneous image to zero and zero to noise,” 2024

    2024

  28. [36]

    V . A. Zorich and O. Paniagua, Mathematical analysis II. Springer, 2016, vol. 220

    2016

  29. [37]

    Swinjscc: Taming swin transformer for deep joint source-channel coding,

    K. Yang, S. Wang, J. Dai, X. Qin, K. Niu, and P. Zhang, “Swinjscc: Taming swin transformer for deep joint source-channel coding,” IEEE Trans. Cognit. Commun. Networking. , 2024

    2024

  30. [38]

    Progressive growing of gans for improved quality, stability, and variation,

    T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of gans for improved quality, stability, and variation,” arXiv preprint arXiv:1710.10196, 2017

    2017 arXiv

  31. [39]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition (CVPR) . IEEE, 2009, pp. 248–255

    2009

  32. [40]

    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

  33. [41]

    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 (CVPR), 2018, pp. 586–595

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.