{"id":"6f6cdf10-6339-4645-a020-3229132f6660","arxiv_id":"2504.21466","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A parallel-stream image transmission scheme sends JPEG/LDPC-coded images together with a residual-conditioned deep semantic stream and merges them with SNR-dependent weights, outperforming separate and learned baselines across most SNR and rate settings.","lead":"This paper combines a conventional JPEG-and-LDPC image transmission path with a small learned 'semantic' stream that carries the details JPEG discards, then blends both at the receiver based on channel noise. The result is a practical image transmission scheme that stays compatible with existing separated coding hardware while gaining resilience at low signal-to-noise ratios.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training SNR for ParaSC is never stated, although Section IV.A fixes the channel state during training; the claimed robustness across the 2–12 dB SNR sweep is therefore unverified and may be extrapolation, especially versus DynaJSCC which was trained over the full range.","rationale":"The reader's weakest assumption and my own stress-test converge: the paper's central claim of superiority over a large SNR range depends on the model's ability to operate at SNRs different from the training condition, yet the training SNR is never reported and Section IV.A explicitly says the channel state is fixed during training. This is not a mere reproducibility nit; the PAGNet aggregation weights are conditioned on SNR, and the semantic stream is sent without channel coding, so the entire claimed robustness to SNR fluctuations is learned behavior. If the model was trained at a single SNR, the reported curves across 2–12 dB are extrapolation, and the comparison against DynaJSCC (trained over the full range) is not apples-to-apples. The paper has real strengths: a coherent hybrid architecture, a plausible rate-conditioning mechanism, and consistent improvements in the reported experiments, plus a useful ablation and complexity table. But none of these compensate for the missing training SNR, because the central claim is explicitly about robustness across a range of channel states. The proposed test directly settles whether the concern lands: retrain under fixed versus mixed SNR conditions and compare to the published curves. I therefore keep the reader's CONDITIONAL verdict; the paper should be accepted only after this is resolved, not rejected outright, since the flaw is an unverified condition rather than a demonstrated contradiction.","tokens_in":24338,"tokens_out":7627,"duration_ms":82018,"concrete_test":"Obtain the released training configuration (or contact the authors for the exact SNR value or SNR sampling used in the three-stage training). Then retrain ParaSC under three controlled conditions on CVRG-Pano/AWGN: (i) fixed SNR=6 dB, (ii) fixed SNR=10 dB, and (iii) SNR uniformly sampled from 2–12 dB, using identical hyperparameters. Regenerate Fig. 6 at CBR=0.1. If the fixed-SNR models lose more than ~1 dB in PSNR at off-training SNRs relative to the reported ParaSC curve, the claimed SNR robustness is not supported and the comparison with DynaJSCC must be redone with matched training SNR conditions. Also report at least 3 seeds with mean plus or minus standard deviation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central performance claim is that ParaSC's SNR-conditioned components (PAGNet weights, semantic decoder, and the rate-adaptive codec) generalize across the SNR range over which superiority is claimed. The paper does not establish this. Section IV.A states 'the channel state is fixed during the training stage' when justifying the decoder condition, and the training details in Section V.A.3 give the loss, optimizer, augmentation, and three-stage schedule but never report the training SNR or whether SNR was sampled from a distribution. The evaluation in Figs. 6-9 sweeps SNR from 2 to 12 dB on both AWGN and Rayleigh channels, and the abstract and contributions specifically advertise robustness to SNR fluctuations. The semantic stream is transmitted uncoded, so its resilience to channel noise is entirely learned; if training used a single SNR, all results at other SNRs are extrapolations, and the PAGNet's dynamic weights may simply memorize the training SNR. The comparison is also asymmetric: DynaJSCC was explicitly trained over 2–12 dB, while ParaSC's training SNR is undisclosed. Without this information, the claimed large-range superiority is not falsifiable from the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ParaSC, a parallel-stream image transmission framework that combines a conventional JPEG/LDPC/QPSK image stream with a learned deep semantic stream. The semantic encoder uses the JPEG residual as attention to enhance extracted features, a conditional entropy model adapts the semantic transmission rate, and a decoder-side PAGNet fuses the two streams with SNR-conditioned pixel-level weights. The authors claim that ParaSC outperforms both conventional SSCC and DL-based JSCC baselines across a wide range of SNR and CBR settings while keeping the semantic stream lightweight. The paper includes rate-distortion and complexity analyses, plus experiments on AWGN and Rayleigh channels with CVRG-Pano and Kodak24 datasets.","tokens_in":24563,"tokens_out":4710,"duration_ms":52681,"significance":"If the claims are substantiated, the hybrid parallel-stream architecture is a practically relevant contribution: it shows how a learned JSCC component can be retrofitted onto a deployed separated coding system without discarding the conventional stack, with a demonstrated mitigation of the cliff effect and a tunable semantic rate. The paper provides a thorough ablation of the rate-adaptation mechanism, a complexity table, and visualizations of the learned aggregation weights. The main significance hinges on the empirical superiority claim and on the correctness of the variational derivation that justifies the training loss, both of which currently require additional support.","major_comments":[{"comment":"The channel SNR(s) used for training ParaSC are never reported, although Section IV.A states that 'the channel state is fixed during the training stage' and Section V.A.3 gives training details without mentioning any SNR distribution. The evaluation sweeps SNR from 2 to 12 dB on AWGN and Rayleigh channels, where the uncoded semantic stream's robustness and the PAGNet's SNR-conditioned behavior are central. If training used only one or a narrow range of SNRs, the results across the rest of the sweep are extrapolation, and the comparison is asymmetric because the DynaJSCC baseline was explicitly trained over 2-12 dB as stated in Section V.A.4. The authors must report the training SNR(s) and, if the broad-range claim is retained, train over the evaluated range or provide a matched training condition.","section":"Section V.A.3 vs. Section IV.A and Figs. 6, 8, 9"},{"comment":"The derivation of the distortion term contains an unproven and internally problematic substitution: the text states that 'xr determines a unique x although such mapping is non-injective,' which is contradictory because a non-injective mapping cannot uniquely determine its pre-image. This assertion is then used to replace p(x|s,xr) with p(x|s,xc) and later with p(x|s,hat_xc), which is load-bearing for identifying Eq[T2] with the MSE distortion term in Eq. (6) and for the final loss in Eq. (8). The authors need to either prove that xr plus the decoder constraints indeed identify x uniquely, or replace this step with a correct factorization that treats xr as a deterministic function of x and xc rather than as a sufficient statistic that identifies x.","section":"Section IV.A, paragraph after Eq. (4)"},{"comment":"The PAGNet decoder takes SNR as an input to determine aggregation weights, and the paper advertises robustness to SNR fluctuations, yet the training procedure does not state whether the SNR is varied or sampled during training. If the channel is fixed during training as stated in Section IV.A, the SNR embedding in PAGNet has not observed different SNRs during optimization, so its behavior at off-training SNRs is unverified. The authors should specify the SNR sampling strategy during the three training stages or demonstrate that a fixed-SNR training regime still yields the claimed generalization, for example by including a training-SNR ablation.","section":"Section III.B and Section V.A.3"},{"comment":"The central empirical claim that ParaSC 'surpasses the performance of both traditional and DL-based competitors in a large range of scenarios' is supported only by single curves without error bars, confidence intervals, or multiple-seed results. The test datasets are small (76 CVRG-Pano images and 24 Kodak24 images), so differences near curve crossings may not be statistically meaningful, and the paper should include variance or significance measures. In addition, Table I compares only the semantic JSCC transmitter's parameters, FLOPs, and throughput, while ParaSC's overall system also includes the conventional JPEG/LDPC/QPSK branch; the 'lightweight' claim should be scoped to the semantic stream or the table should report end-to-end complexity, otherwise the comparison is incomplete.","section":"Section V.B and Table I"}],"minor_comments":[{"comment":"The word 'thrid' should be 'third' in the sentence introducing Eq[T3].","section":"Section IV.A"},{"comment":"The training details are incomplete for reproduction: no learning rate, batch size, number of epochs, or optimizer settings beyond the default beta values are given, and the poly learning-rate schedule is described only qualitatively.","section":"Section V.A.3"},{"comment":"The predefined rate set W and its cardinality M are not defined in the text; Section V.A.2 mentions '32 available equidistant output dimensions ranging from 4 to 128,' but the connection between this and W, and how the FC layers select among these dimensions, should be made explicit.","section":"Section IV.B"},{"comment":"The discussion of DynaJSCC's CBR points says the optimal learned CBR is 0.167 and other points are obtained by changing its binary mask; the paper should clarify whether those mask-based points were all seen in training and whether the comparison is on a fair rate-adaptation basis.","section":"Section V.B"},{"comment":"The throughput measurement should report the inference framework, batch size, and whether the number is for the semantic encoder alone or for the full ParaSC transmitter, since the table is used to support the lightweight claim.","section":"Section V.E, Table I"},{"comment":"The notation mapping x -> z_w via F_theta and then hat_z_w -> hat_x_c via F_theta^{-1} is confusing because F_theta is also used for the compressor; a separate symbol for the decoder would improve readability.","section":"Section II.C.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's architectural idea is timely and the experimental setup is extensive, but the missing training-SNR disclosure and the questionable uniqueness assumption in the loss derivation need to be fixed before the claims can be accepted. The rate-adaptation mechanism is a standard conditional compression formulation and should be positioned as such; this does not by itself reduce the paper's value. I would also urge the authors to provide error bars or multi-seed results, since the reported curves are otherwise difficult to judge."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a real architectural proposal, not a repackaging. The idea is simple and pragmatic: keep your deployed JPEG/LDPC chain and add a lightweight learned semantic stream that is conditioned on the compression residual; at the receiver, merge the two streams with pixel-level weights that depend on SNR feedback. The PAGNet mechanism is the most interesting part—it visibly masks corrupted regions of the conventional decode under low SNR. The rate-adaptation via a residual-conditioned entropy model is a sensible extension of [21] and seems to work in the ablation (semantic rate drops as JPEG quality factor increases). Compared to DJSCC and DynaJSCC, the reported gains on PSNR/MS-SSIM/LPIPS across two datasets and both AWGN and Rayleigh channels are consistent with the claim of \"large range\", with the caveat that low-SNR performance is sometimes below DynaJSCC—the paper says this too.\n\nWhat I think are the real soft spots:\n\n1. Training SNR is never reported. Section IV.A explicitly says \"the channel state is fixed during the training stage,\" and Section V.A.3 gives no SNR range or sampling schedule. The evaluation sweeps 2–12 dB, and the whole point of PAGNet is robustness to SNR fluctuations. If the model was trained at a single SNR, the curves outside that point are extrapolation and the comparison against DynaJSCC (trained over 2–12 dB) is asymmetric. This is a load-bearing omission, not a minor one.\n\n2. The derivation in IV.A has a hand-wavy step: claiming x_r determines a unique x while simultaneously saying the mapping is non-injective. T4 is dropped as a constant without formal justification. This is the kind of thing that makes a careful reader stop trusting the rest of the math. It is probably fixable—the variational bound is standard—but it needs a clean rewrite.\n\n3. No code, no data, no error bars. Single curves make it impossible to tell whether the gains are within run-to-run variation. The complexity table is interesting, but WITT appears only there, not in the performance comparison, so the \"DL-based competitors\" claim is a bit narrower than the abstract implies.\n\nIf I were the editor, I would send this to reviewers. The architecture is concrete, the problem (retrofitting practical SSCC with semantics) is real, and the results, if reproducible, are useful. But I would ask the authors to release code/data, report training SNR and seeds, add error bars, and clarify the low-SNR region where they don't beat DynaJSCC. This is not a desk-reject, it is a revise-and-resubmit candidate.","headline":"A practical parallel-stream SSCC+JSCC image transmission framework with a promising architecture, but one load-bearing training detail (the training SNR) is missing and the simulations are not reproducible as written.","tokens_in":25111,"tokens_out":3414,"would_cite":true,"duration_ms":34758,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that adding a lightweight learned semantic stream to a conventional JPEG/LDPC/QPSK image pipeline yields better reconstruction than either approach alone, while keeping the extra bandwidth small.","keywords":["semantic communication","joint source-channel coding","separate source-channel coding","rate adaptation","image transmission","residual coding","parallel-stream transmission","wireless image reconstruction"],"falsifier":"Run a controlled experiment in which the model is trained at a single SNR (say 6 dB) and tested at 2 and 12 dB on the same datasets and metrics; if PSNR/MS-SSIM/LPIPS gains over the best single-stream baseline disappear or reverse at the off-training SNRs, the claimed large-range superiority would fail. Alternatively, compare against an oracle that picks the best baseline per SNR; if the oracle matches or beats ParaSC everywhere, the parallel stream is not strictly necessary.","tokens_in":24119,"feed_emoji":"🖼️","tokens_out":5305,"duration_ms":54599,"temperature":0.7,"pith_summary":"This paper proposes ParaSC, a wireless image-transmission framework that runs a conventional JPEG/LDPC/QPSK pipeline in parallel with a lightweight learned semantic stream. The central claim is that the combination outperforms both pure conventional separate source-channel coding and pure learned joint source-channel coding on PSNR, MS-SSIM, and LPIPS across most tested signal-to-noise ratios and bandwidth ratios. The semantic encoder is driven by the residual between the original image and the JPEG-decoded image, so it spends bits on what the conventional codec threw away, and a conditional entropy model shrinks the semantic rate accordingly. The decoder's parallel aggregation network uses received SNR to down-weight corrupted image-stream regions, which softens the cliff effect that breaks pure SSCC at low SNR. A sympathetic reader would care because the design promises a practical upgrade path: improve an already-deployed SSCC system by adding a small, tunable neural stream rather than replacing the codec.","feed_headline":"Hybrid JPEG-plus-semantics beats either codec alone","feed_subtitle":"A small learned stream fixes what JPEG discards and softens the cliff effect.","key_machinery":"The carrying mechanism is the residual-conditioned semantic path. The encoder computes the compression residual $x_r = x - x_c$ and treats it as an attention signal that highlights high-frequency content JPEG discarded; the same residual is quantized into a hyperprior that conditions a Gaussian entropy model, so each semantic patch is assigned the minimum number of bits implied by its conditional entropy. At the receiver, PAGNet—a per-pixel, per-scale weighting network—takes the received SNR as an embedding and learns to suppress corrupted regions of the JPEG-decoded image while letting the semantic stream dominate at low SNR. The rate-distortion objective is derived as a conditional variational-autoencoder ELBO whose two optimizable terms are reconstruction distortion and semantic rate.","core_discovery":"On the paper's own terms, the discovery is that classical separation-based coding and deep JSCC are complementary rather than competing: the conventional stream supplies the bulk of the image at moderate SNR, while a learned semantic stream that is conditioned on the residual $x - x_c$ compensates for compression loss and carries the reconstruction when the channel is poor. The paper develops ParaSC to make this concrete: a residual-enhanced attention encoder produces semantic features and a residual hyperprior; a conditional entropy model allocates a variable code length to each semantic patch; and the PAGNet decoder fuses the two streams with pixel-level, SNR-conditioned weights. Numerical results across CVRG-Pano and Kodak24, on AWGN and Rayleigh fading channels, are reported to show that ParaSC surpasses the JPEG-based SSCC baselines and the learned JSCC baselines over most of the tested SNR and CBR ranges, with the semantic stream consuming only a small fraction of the total rate (about 10% or less for the $\\lambda_1=0.01$ setting when the image-stream rate is above 0.07).","pith_inferences":["If the framework's robustness to channel fluctuations comes from PAGNet's SNR conditioning rather than from having seen many SNRs at training, then a natural extension is to train once and adapt only the PAGNet weights at deployment, which could be tested by fine-tuning on a single new SNR.","The same residual-conditioning idea could carry to video or multi-view images, where the 'residual' is temporal prediction error rather than compression loss, making the learned stream a general error-compensation channel.","A stronger claim the paper does not make is that the hybrid is optimal; one could test whether an oracle that switches between pure SSCC, pure JSCC, and the hybrid at each SNR would beat ParaSC, which would reveal whether the parallel stream is always the best use of bandwidth.","The rate-adaptation module currently conditions on compression quality $q$; a testable extension is to also condition on predicted channel state, allowing the semantic rate to shrink when the channel is clean and grow when it is noisy."],"forward_implications":["An existing JPEG-plus-LDPC system can be upgraded by appending a semantic encoder and decoder, without retraining or replacing the conventional codec.","The cliff effect of SSCC is softened: at low SNR the semantic stream dominates, and at high SNR the conventional stream dominates, so reconstruction quality degrades gracefully.","Semantic rate can be traded against image-stream rate through the loss weight, letting the framework hit varied bandwidth budgets with a single trained model.","At the tested settings the semantic stream is light: about 8.3% extra rate yields the largest PSNR gain at the worst channel condition, and the model has far lower parameter count and FLOPs than transformer-based JSCC options."],"supporting_citations":[{"why":"Supplies the autoencoder-based learned JSCC baseline that ParaSC must outperform.","marker":"[43]"},{"why":"Supplies the rate-adaptive JSCC baseline with an SNR-based policy that ParaSC is compared against.","marker":"[20]"},{"why":"Supplies the transformer-based wireless image transmission baseline used in the complexity comparison.","marker":"[22]"},{"why":"Provides the nonlinear transform source-channel coding approach whose entropy-estimation idea the rate-adaptation module builds on.","marker":"[21]"},{"why":"Provides the scale-hyperprior variational compression framework that the residual-conditioned entropy model extends.","marker":"[34]"},{"why":"Defines the JPEG compression standard used as the conventional source codec in the image stream.","marker":"[1]"},{"why":"Supplies the 5G LDPC channel code used for the conventional image stream.","marker":"[4]"},{"why":"Supplies the MS-SSIM perceptual quality metric used in the evaluation.","marker":"[44]"},{"why":"Supplies the LPIPS perceptual quality metric used in the evaluation.","marker":"[45]"}],"fun_headline_variants":["Parallel semantic stream boosts classic codec, robust to fading","Add a tiny learned stream to JPEG for cliff-free image delivery","Classic plus semantic coding outperforms either approach alone","Residual-conditioned semantic codec bests conventional and deep baselines","Lightweight semantic path strengthens traditional image transmission"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper trains with the channel state fixed, but evaluates across a wide SNR range without reporting the training SNR(s); the claimed robustness to SNR fluctuations rests on the trained network generalizing to channel conditions outside its training distribution.","fun_headline_variants_meta":{"raw":{"variants":["Parallel semantic stream boosts classic codec, robust to fading","Add a tiny learned stream to JPEG for cliff-free image delivery","Classic plus semantic coding outperforms either approach alone","Residual-conditioned semantic codec bests conventional and deep baselines","Lightweight semantic path strengthens traditional image transmission"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000679,"raw_usage":{"total_tokens":3099,"prompt_tokens":973,"completion_tokens":2126,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":2045}},"tokens_in":589,"tokens_out":2126,"duration_ms":16968,"temperature":1.0,"reasoning_tokens":2045,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:03:13.511947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled experiment in which the model is trained at a single SNR (say 6 dB) and tested at 2 and 12 dB on the same datasets and metrics; if PSNR/MS-SSIM/LPIPS gains over the best single-stream baseline disappear or reverse at the off-training SNRs, the claimed large-range superiority would fail. Alternatively, compare against an oracle that picks the best baseline per SNR; if the oracle matches or beats ParaSC everywhere, the parallel stream is not strictly necessary.","supporting_citations":[{"cited_title":"Deep joint source-channel coding for wireless image transmission with adaptive rate control,","cited_arxiv_id":null,"evidence_quote":"Supplies the rate-adaptive JSCC baseline with an SNR-based policy that ParaSC is compared against."},{"cited_title":"WITT: A wireless image transmission transformer for semantic communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the transformer-based wireless image transmission baseline used in the complexity comparison."},{"cited_title":"Nonlinear transform source-channel coding for semantic communications,","cited_arxiv_id":null,"evidence_quote":"Provides the nonlinear transform source-channel coding approach whose entropy-estimation idea the rate-adaptation module builds on."},{"cited_title":"Variational image compression with a scale hyperprior,","cited_arxiv_id":null,"evidence_quote":"Provides the scale-hyperprior variational compression framework that the residual-conditioned entropy model extends."},{"cited_title":"The JPEG still picture compression standard,","cited_arxiv_id":null,"evidence_quote":"Defines the JPEG compression standard used as the conventional source codec in the image stream."},{"cited_title":"Multiscale structural similarity for image quality assessment,","cited_arxiv_id":null,"evidence_quote":"Supplies the MS-SSIM perceptual quality metric used in the evaluation."},{"cited_title":"The unreasonable effectiveness of deep features as a perceptual metric,","cited_arxiv_id":null,"evidence_quote":"Supplies the LPIPS perceptual quality metric used in the evaluation."}],"review_version":1}