{"id":"b9bbe286-18e5-47f2-a81a-080986fe5a34","arxiv_id":"2501.01138","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SGD-JSCC guides a diffusion denoiser with transmitted text or edge maps to clean the received latent features of a DeepJSCC image codec, improving perceptual quality at low SNR without pilot-based CSI.","lead":"Wireless image transmission can be made far more robust at very low signal-to-noise ratio by sending a text caption or edge map alongside the image and using a diffusion model, guided by that semantic hint, to clean the received signal. The method targets a practical problem in deep joint source-channel coding: keeping images recognizable when the channel is extremely noisy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Text side channel is not free/perfect at SNR=-15 dB: capacity math shows a short caption would consume more than the whole CBR budget, so the text-guided gains in Fig. 15 are not deployable in the low-SNR regime.","rationale":"The reader's weakest assumption is the perfect/free text side channel, and I agree this is the most load-bearing concern. The abstract presents text description as a headline example of semantic guidance, and the ablation (Fig. 15) shows text guidance contributes substantial gains at low SNR. The paper's own footnote admits transmission at SNR=-15 dB is challenging, and a capacity calculation shows the text is not negligible: at -15 dB, even a short caption requires more channel uses than the entire image budget at CBR=1/20. The edge-map variant is a partial mitigation, but the paper does not report edge-map reconstruction accuracy at -15 dB, so the 'satisfactory reconstruction' claim in that regime rests on an unverified premise. Other potential concerns - asymmetric training data, absence of error bars, the internal consistency of the diffusion step matching and water-filling equations - were checked and are not fatal; the mathematics is internally consistent and the ablation study validates the benefit of semantic guidance within the proposed framework. The reader's CONDITIONAL verdict remains appropriate, with the condition that the side-channel assumption be addressed or the claim appropriately scoped.","tokens_in":37464,"tokens_out":14867,"duration_ms":136356,"concrete_test":"Transmit each COCO validation caption over the same -15 dB AWGN channel using a capacity-approaching code (or a DeepJSCC text code), within the total CBR budget. Use the received (possibly corrupted) text as guidance in the text-guided DM, and compare LPIPS/FID/CLIP against the clean-text upper bound in Fig. 15. Also measure decoded edge-map Dice/BCE at -15 dB and ablate with clean vs. received edge maps. If the clean-vs-received gap exceeds the guidance gain in Fig. 15, the low-SNR claim is not deployable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of semantics guidance improving perceptual quality at extremely poor channel conditions leans on two side information modes. Text guidance (Sec. IV-A) assumes 'the transmission cost of text descriptions is negligible' and 'can be transmitted to the receiver perfectly' (footnote 2). At the SNR=-15 dB regime highlighted in the abstract, this is not merely imperfect; it is capacity-infeasible. For a 128x128 image, CBR=1/20 gives roughly 2,458 complex channel symbols total. At -15 dB, Shannon capacity is about 0.045 bits per complex dimension, so a typical 50-character BLIP2 caption (~400 bits) needs on the order of 8,900 channel uses - more than three times the whole budget. The footnote handwaves that 'specialized DeepJSCC techniques' could preserve semantic meaning, but no such scheme is provided or evaluated. The reported text-guidance gains (Fig. 15, CLIP/FID) therefore cannot be realized at -15 dB. The edge-map variant avoids the free-text assumption by counting the edge map against the CBR, but the paper never measures decoded edge-map accuracy at -15 dB; if the edge map is too corrupted, the diffusion guidance is misled. Thus the low-SNR 'satisfactory reconstruction' claim is only as strong as the unmeasured edge-map decoding at that SNR.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SGD-JSCC, a deep joint source-channel coding scheme in which a diffusion model is used to denoise the JSCC channel output before decoding, guided by semantic side information in the form of either a text description or an edge map. For slow fading channels the scheme estimates the instantaneous SNR and phase directly from the equalized channel output, so that the diffusion starting step is matched continuously to the channel state without pilot overhead; for fast fading channels it introduces a water-filling-inspired denoising procedure that equalizes per-symbol noise levels so that a DM trained under AWGN can be applied without fine-tuning. The numerical section compares SGD-JSCC with ADJSCC, JSCCformer, DeepJSCC-Diff, JSCCDiff, and VAEJSCC on COCO2017 and Kodak, reporting gains in perceptual metrics (LPIPS, FID, CLIP) at low SNR, and includes ablations of text versus edge-map guidance, blind CSI estimation accuracy, masking strategies, and fast-fading behavior.","tokens_in":37680,"tokens_out":9066,"duration_ms":80692,"significance":"If the reported results hold, the paper makes a useful step toward making diffusion-based JSCC practical: the continuous step-matching formulation (Eq. (20)) is a clean, parameter-free mapping from SNR to diffusion noise level, and the water-filling extension in Section VI is an interesting training-free way to reuse a slow-fading DM in fast-fading channels. The manuscript also contains extensive comparisons and ablations, and the authors promise public code and checkpoints, which would facilitate reproducibility. The main limitation is that the text-guidance variant assumes an ideal, cost-free side channel at exactly the low-SNR regime where the paper claims its largest perceptual gains; this assumption is capacity-infeasible at SNR = -15 dB and needs to be reframed as an upper bound or accompanied by an evaluated transmission scheme. The edge-map variant is within the rate budget but its reliability at very low SNR is not directly measured.","major_comments":[{"comment":"The perfect, cost-free text side channel is capacity-infeasible in the regime the paper highlights. For a 128x128x3 image at CBR = 1/20, the system has roughly 2,458 complex channel symbols in total. At SNR = -15 dB, the Shannon capacity is approximately 0.045 bits per complex dimension, giving about 110 bits per image for the whole transmission. A typical 50-character BLIP2 caption (~400 bits) would require on the order of 8,900 channel uses, more than three times the entire CBR budget. Since Fig. 15 attributes the best low-SNR CLIP/FID performance to text guidance, those specific gains are not deployable at -15 dB. The footnote's appeal to 'specialized DeepJSCC techniques' is not backed by a scheme, simulation, or capacity analysis. Please either count the text overhead in the CBR and evaluate with corrupted text reception, or present the text-guided results explicitly as an upper bound and make the edge-map-guided variant the primary deployment claim.","section":"Section IV-A, footnote 2"},{"comment":"The reliability of the decoded edge map at very low SNR is not measured, even though edge-map guidance is the variant used in the main -15 dB comparisons (footnote 6). The only edge-map reconstruction example, Fig. 16, is shown at SNR = -10 dB, and the paper itself states in Section VII-C1 that edge-map transmission errors can mislead the diffusion model and degrade performance at high SNR. Without a quantitative evaluation of decoded edge-map fidelity at -15 dB (e.g., BCE/Dice on the COCO validation set) or an ablation that compares ground-truth edge guidance with decoded edge guidance, the claim that semantic guidance is responsible for the -15 dB perceptual gains is not fully supported. Please add such measurements or soften the corresponding claims.","section":"Section VII-C1 and Fig. 15"},{"comment":"The use of CLIP score as a reconstruction metric is biased in favor of the text-guided variant because the receiver has access to the ground-truth caption. A text-conditioned diffusion model can obtain a high CLIP score by producing any image that is semantically consistent with the caption, including hallucinated content that the paper itself acknowledges can occur (Section VII-C1). The FID metric has a related realism bias. To support the perceptual-quality claims fairly, please report reconstruction-oriented metrics (LPIPS with the decoded edge map, or a faithfulness metric that penalizes caption-consistent but content-incorrect reconstructions) alongside CLIP/FID, and discuss the bias explicitly.","section":"Section VII, Performance Metrics and Fig. 15"}],"minor_comments":[{"comment":"The benchmark naming is inconsistent: the text alternates between 'DeepJSCC-Diff', 'JSCCDiff', and 'DiffJSCC' for what appear to be the same or closely related baselines; please align the names in the text, figures, and table captions.","section":"Section II-C and Section VII-A"},{"comment":"The sentence 'The ADJSCC scheme refers to the DeepJSCC architecture in [14]' appears to cite the wrong reference, since ADJSCC is introduced in [16]; please correct the citation.","section":"Section VII-A"},{"comment":"The text says the edge-map JSCC model uses a ViT architecture, while Fig. 2 is labeled 'Swin Transformer'; please clarify which architecture is actually used.","section":"Section IV-B and Fig. 2"},{"comment":"As typeset, the noise coefficient in Eq. (16) appears to contain an extra factor; please verify that it matches the standard variance-preserving conditional distribution q(ft|fs).","section":"Eq. (16)"},{"comment":"Table I mixes layer architecture with training hyperparameters (batch size, epochs, optimizer); please separate the network specification from the training configuration.","section":"Table I"},{"comment":"The abstract and introduction emphasize 'no additional pilot transmissions for channel estimation', but the fast-fading extension in Section VI assumes perfect CSI at the receiver; please state this assumption explicitly in the contributions and conclusions.","section":"Abstract and Section VI"}],"recommendation":"major_revision","confidential_remarks":"The free-text side-channel assumption is the main correctness risk and is likely to be a sticking point for the cs.IT readership. The edge-map variant is within the rate budget, so the core DeepJSCC+DM contribution is defensible; the authors should be asked to strengthen the edge-map evaluation at -15 dB and to present the text-guided results as an upper bound unless they provide a concrete and evaluated low-rate text transmission scheme. The paper is otherwise technically solid and contains enough reproducible detail that I would not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2501.01138. The paper is a genuine step forward for diffusion-based JSCC. The continuous timestep matching and water-filling adaptation are sound, and the main slow-fading results use edge-map guidance that is counted in the CBR budget. The weak spot is the text side channel: the paper assumes text is transmitted perfectly for free, and at -15 dB that assumption fails by an order of magnitude under Shannon capacity. That undercuts the text-guided ablation in Fig. 15, but it does not sink the main comparison.\n\nWhat is actually new: semantics-guided denoising (text or edge map) for channel denoising, continuous timestep matching instead of discrete step matching, pilot-free joint SNR/phase estimation, and a water-filling denoising strategy for fast fading. The step-matching math is internally consistent. Equation (9) gives a variance-preserving corruption with noise level 1/(1+gamma), and Eq. (20) inverts the sigmoid schedule to select a valid starting point. I see no circularity: the DM is trained independently of the final metrics, and the fitted components are the SNR/phase estimators and hyperparameters, not the headline numbers. The ablations and sensitivity analysis are extensive.\n\nThe free-text assumption is the one load-bearing flaw. Footnote 2 concedes the problem and handwaves about specialized DeepJSCC, but the capacity math is not kind: at -15 dB, a 50-character caption needs several times the entire CBR budget of 2458 complex symbols. So the text-guidance gains in Fig. 15 are not deployable in the regime where the abstract promises \"satisfactory reconstruction.\" The edge-map variant avoids the free-rider issue, yet the paper never measures decoded edge-map accuracy at -15 dB, leaving open the possibility that the guidance itself is corrupt exactly when it matters most.\n\nTwo lesser issues. First, no error bars anywhere, so we cannot judge whether the low-SNR differences are noise. Second, the proposed model trained on roughly 14 million image-text pairs while the baselines trained on COCO train, so the head-to-head is not fully apples-to-apples. Code and checkpoints are promised but not yet released.\n\nOn balance, the central claim — edge-guided diffusion preprocessing beats post-processing baselines at low SNR — looks credible and the engineering is careful. This paper deserves a serious referee. I would send it out with a request for: (a) error bars; (b) either costing the text side channel within CBR or repositioning text-guidance claims as an upper bound with a clean side channel; (c) measuring decoded edge-map accuracy at low SNR; and (d) releasing the code. Anyone working in semantic communication or diffusion-aided JSCC will want to read this; I would cite it for the water-filling trick.","headline":"A solid, referee-worthy JSCC-diffusion paper whose main edge-guided results hold up; the free-text side channel fails at the low-SNR regime where text guidance is claimed to shine, but that flaw mostly hits an ablation, not the headline comparison.","tokens_in":38294,"tokens_out":3088,"would_cite":true,"duration_ms":30536,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SGD-JSCC transmits a text caption or edge map alongside deep joint source–channel coded image features and uses a semantics-guided diffusion model to denoise the channel output, claiming perceptual gains over prior DeepJSCC schemes at SNR…","keywords":["semantics-guided diffusion","deep joint source-channel coding","wireless image transmission","channel denoising diffusion model","semantic side information","pilot-free CSI estimation","fast fading adaptation","perceptual image quality"],"falsifier":"Run SGD-JSCC at SNR = −15 dB while transmitting the text caption through the same fading channel, or at a finite rate, rather than assuming perfect delivery, and compare the text-guided variant against the edge-map-only and unconditional variants on LPIPS, FID, and CLIP score; if the text-guided gains disappear or reverse, the semantics-guidance claim at extreme low SNR collapses.","tokens_in":37201,"feed_emoji":"📡","tokens_out":9143,"duration_ms":81532,"temperature":0.7,"pith_summary":"This paper aims to move deep joint source–channel coding (DeepJSCC) from transmitting image features alone to transmitting image features plus a small amount of semantic side information—a text caption or an edge map—that is used to guide a diffusion model acting as a channel denoiser. The central proposal, SGD-JSCC, treats the noisy channel output as an intermediate state of a diffusion process and iteratively removes channel noise before the JSCC decoder, rather than using diffusion to post-process a distorted reconstruction. The authors argue that this semantic conditioning turns a generative model from a source of plausible but arbitrary images into a faithful reconstructor, and that it makes the system channel-adaptive without pilots or fine-tuning. If the claim holds, wireless image transmission can preserve recognizable semantics at signal-to-noise ratios as low as −15 dB, where conventional DeepJSCC reconstructs images that lose key content. The paper reports gains over prior diffusion-based and attention-based DeepJSCC schemes on perceptual metrics (LPIPS, FID, CLIP score) at a bandwidth compression ratio of 1/20, while acknowledging that pixel-level PSNR remains below a leading baseline.","feed_headline":"Diffusion plus semantic cues revive wireless images at −15 dB","feed_subtitle":"A caption or edge map guides channel denoising, beating DeepJSCC baselines on perceptual quality without pilots.","key_machinery":"The mechanism that carries the argument is the placement and conditioning of the diffusion denoiser. Instead of refining the decoder output, the denoiser operates on the equalized channel output before the JSCC decoder, viewing the channel output as one intermediate noisy state in a variance-preserving forward diffusion with a continuous sigmoid noise schedule; the denoiser predicts the clean latent from that state. Conditioning comes from semantic side information: text embeddings serve as keys and values in cross-attention layers, while edge maps enter through a frozen copy of the base network plus a zero-initialized linear fusion branch. The same diffusion schedule is used for channel adaptation: a learned estimator maps the received signal to the equivalent SNR and phase, the inverse schedule maps SNR to a continuous starting timestep, and the fast-fading extension equalizes per-symbol noise levels by adding noise to over-clean elements, then updates only elements whose noise level is above the next target level.","core_discovery":"SGD-JSCC's core discovery is that the channel noise can be matched to the forward noise of a diffusion model and removed under semantic guidance, so that the JSCC encoder and decoder can be trained once for one channel state while the diffusion denoiser absorbs all channel variation. Two semantic modalities are supported: coarse text descriptions, extracted by an image-captioning model and injected through cross-attention, and fine edge maps, transmitted through a dedicated JSCC encoder and injected through a zero-initialized structural-control branch. In slow fading, the receiver estimates the instantaneous SNR and phase directly from the normalized channel output, converts the estimated SNR to a continuous noise level via the inverse scheduling function, and starts denoising at the matched timestep, eliminating pilot-based channel estimation. In fast fading, a water-filling-inspired procedure adds carefully chosen Gaussian noise to cleaner symbols to equalize their noise levels to the diffusion schedule, so a denoiser trained under AWGN can be applied without retraining. The reported result is that this scheme outperforms ADJSCC, JSCCformer, DeepJSCC-Diff, JSCCDiff, and a non-diffusion VAE variant on LPIPS, FID, and CLIP score over the tested conditions, with visual examples showing semantic content such as a bear in grass preserved at −10 dB where unconditional denoising reconstructs an unidentified animal.","pith_inferences":["The paper's text-based variant assumes the caption is received perfectly and costs nothing; transmitting the caption through the same channel at finite rate is a testable extension that would reveal whether the reported low-SNR gains survive when the guidance itself is noisy.","Because the diffusion prior is trained on open-domain image–text data, the receiver can generate content that is plausible but absent from the source; a natural next evaluation is to measure object-level consistency, not just global perceptual scores, between the reconstruction and the original.","The continuous noise-level matching and water-filling update are channel-agnostic in principle; the paper targets point-to-point image links, but the same mechanism is a plausible starting point for MIMO, OFDM, or non-image modalities, which the paper does not test."],"forward_implications":["In slow fading, the system can operate without pilot transmissions: SNR and phase are estimated from the received latent, and the paper reports these estimates are accurate enough that reconstruction quality is nearly the same with and without receiver SNR knowledge.","The JSCC encoder–decoder is channel-agnostic: once trained under a fixed AWGN SNR, it can be reused across channel conditions and semantic metrics, with only the diffusion denoiser and guidance modules tailored.","In fast fading, a diffusion denoiser trained for AWGN can be applied without fine-tuning via the water-filling update rule, and the paper reports only a slight performance decline as fading block length increases from 1 to 1024.","Masking parts of the latent feature during diffusion training enables rate-adaptive transmission, where dropping the lowest-L2-norm tokens degrades quality less than random dropping, suggesting an importance-based resource allocation rule.","The semantic side information is what preserves meaning at extreme low SNR: at −15 dB the scheme retains recognizable objects, whereas conventional DeepJSCC baselines lose semantics even when their PSNR is higher."],"supporting_citations":[{"why":"The channel-denoising diffusion model that the paper extends, providing the preprocessing placement and the step-matching idea.","marker":"[15]"},{"why":"The base denoising diffusion formulation whose forward and reverse processes are adapted to a continuous noise schedule and deterministic sampling.","marker":"[17]"},{"why":"Provides the non-Markovian sampling distribution used to derive the deterministic reverse denoising step in the proposed DM.","marker":"[39]"},{"why":"The structural-control approach adapted to inject edge-map guidance into the diffusion model.","marker":"[21]"},{"why":"The captioning model used to generate the text descriptions that serve as coarse semantic guidance.","marker":"[37]"},{"why":"The edge detector used to produce the fine structural semantic guidance.","marker":"[38]"},{"why":"A diffusion-based post-processing baseline that the proposed preprocessing approach is designed to improve upon.","marker":"[28]"},{"why":"A diffusion-aided JSCC baseline compared in the numerical evaluation.","marker":"[30]"}],"fun_headline_variants":["Semantic-guided diffusion cleans up wireless images at −15 dB","Text or edges guide diffusion to fix noisy wireless image transmission","Diffusion denoising with semantic cues beats DeepJSCC on bad channels","One diffusion model adapts to all channel conditions, no pilots needed","Semantic diffusion rescues image quality at extremely low SNR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the text description reaches the receiver perfectly and costs no channel resources; if that side channel is noisy or rate-limited, the text-guided gains reported at very low SNR may not be achievable in practice.","fun_headline_variants_meta":{"raw":{"variants":["Semantic-guided diffusion cleans up wireless images at −15 dB","Text or edges guide diffusion to fix noisy wireless image transmission","Diffusion denoising with semantic cues beats DeepJSCC on bad channels","One diffusion model adapts to all channel conditions, no pilots needed","Semantic diffusion rescues image quality at extremely low SNR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000346,"raw_usage":{"total_tokens":1991,"prompt_tokens":1132,"completion_tokens":859,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":748,"completion_tokens_details":{"reasoning_tokens":771}},"tokens_in":748,"tokens_out":859,"duration_ms":7381,"temperature":1.0,"reasoning_tokens":771,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:34:46.222241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SGD-JSCC at SNR = −15 dB while transmitting the text caption through the same fading channel, or at a finite rate, rather than assuming perfect delivery, and compare the text-guided variant against the edge-map-only and unconditional variants on LPIPS, FID, and CLIP score; if the text-guided gains disappear or reverse, the semantics-guidance claim at extreme low SNR collapses.","supporting_citations":[{"cited_title":"CDDM: Channel denoising diffusion models for wireless semantic communications,","cited_arxiv_id":null,"evidence_quote":"The channel-denoising diffusion model that the paper extends, providing the preprocessing placement and the step-matching idea."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"The base denoising diffusion formulation whose forward and reverse processes are adapted to a continuous noise schedule and deterministic sampling."},{"cited_title":"Adding conditional control to text-to-image diffusion models,","cited_arxiv_id":null,"evidence_quote":"The structural-control approach adapted to inject edge-map guidance into the diffusion model."},{"cited_title":"Blip-2: Bootstrapping language- image pre-training with frozen image encoders and large language models,","cited_arxiv_id":null,"evidence_quote":"The captioning model used to generate the text descriptions that serve as coarse semantic guidance."},{"cited_title":"Muge: Multiple granularity edge detection,","cited_arxiv_id":null,"evidence_quote":"The edge detector used to produce the fine structural semantic guidance."}],"review_version":1}