{"id":"6d59d719-c5b0-45a5-96c1-019d51c0c8ab","arxiv_id":"2411.10650","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Progressive transmission of learned-compressed image latents reduces wireless latency and improves low-SNR throughput, but the experimental setup omits actual bit errors and non-progressive baselines.","lead":"This paper tests whether progressive versions of two learned image compression models, hyperprior and VQGAN, can deliver images faster and more reliably over fading wireless channels than an adaptive baseline. It finds that the progressive hyperprior gives the lowest waiting times across all simulated signal-to-noise ratios, but the experimental setup does not simulate actual bit errors or include non-progressive baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reliability and low-SNR latency claims assume an error-free bit pipe: the Rayleigh channel only sets a per-slot bit budget, so VQGAN's 'robust without channel coding' advantage and hyperprior's throughput lead are untested.","rationale":"The paper aims to show that progressive learned image compression improves reliability, throughput, and latency over wireless channels. For that central claim to hold, the actual noisy channel must affect the transmitted compressed bitstream; otherwise the 'wireless reliability' component is reduced to rate-distortion performance under a time-varying bit budget. The reader's weakest assumption identifies exactly this gap: the Rayleigh fading model only sets per-slot bit budgets, and the transmitted arithmetic-coded bytes and codebook indices are never subjected to bit errors. I agree that this is the most load-bearing concern. The VQGAN robustness claim in the abstract and the progressive-hyperprior low-SNR throughput claim both depend on error-free delivery. If bit errors were introduced, the comparison could reverse, because hyperprior entropy coding is fragile while VQGAN index errors are localized. I do not see this as an internal inconsistency in the code or derivations; it is an omitted experimental condition, and it directly undermines the reliability language of the paper. Secondary issues, such as the absence of a non-progressive LIC baseline and the unspecified modulation/coding scheme, are real but less central: even with a non-progressive baseline, the reliability claim would remain untested without bit errors. The paper does have some independent merit: it proposes a concrete progressive pipeline, uses pretrained models, and provides code. Those are useful but do not supply the missing bit-error simulation. The reader's REJECT verdict is therefore appropriate as submitted; no adjustment is needed. If the proposed bit-error test passed, the main reliability concern would be resolved, though the other experimental gaps would still need attention before acceptance.","tokens_in":16540,"tokens_out":5884,"duration_ms":62251,"concrete_test":"Re-run the Section V.F transmission simulation with actual bit errors at the same SNR values: map each model's bitstream to symbols (e.g., BPSK/QPSK or the unspecified modulation), pass them through the Rayleigh fading channel of Eq. (1), demodulate with no channel coding, and decode. To isolate the effect cleanly, also flip random bits at BER = 10^-2, 10^-3, and 10^-4 over 1,000 channel realizations. Count arithmetic-decoder failures and measure PSNR/SSIM for progressive-hyperprior, progressive-VQGAN, and adaptive WebP. If hyperprior PSNR stays near the error-free value at BER 10^-3 while VQGAN degrades gracefully, the simplification is harmless; if hyperprior collapses or requires retransmission, the low-SNR latency/throughput claims in Table II and the abstract must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline results all come from a transmission model in which the Rayleigh channel only determines how many bits fit per slot. In Section V.F, step 4, the bit budget Nbits is 'determined based on the current SNR and selected modulation scheme,' but no step injects bit errors; the hyperprior decoder (Eqs. 16-18) and the VQGAN byte2index block operate on exactly the encoded bytes and indices. Consequently, 'reliability' is operationalized as PSNR/SSIM after lossless delivery of a truncated bitstream, not as resilience to channel noise. This is load-bearing for two central claims: (i) the abstract's assertion that VQGAN is 'robust ... even in the absence of channel coding' is never directly tested, and (ii) the progressive-hyperprior latency/throughput advantage at -10 and -5 dB assumes its arithmetic-coded ybytes and zbytes arrive intact. A single bit error in an arithmetic-coded stream can desynchronize the decoder and collapse PSNR, whereas a flipped VQGAN index substitutes one codeword and degrades only locally. If real uncoded or lightly coded transmission were simulated, the low-SNR ordering in Table II could reverse: hyperprior would need FEC or ARQ, adding latency and reducing effective throughput, while VQGAN could still decode. The conclusion that progressive transmission 'enhances reliability' is therefore not supported by the evidence actually collected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes an adaptive and progressive transmission pipeline for learned image compression (LIC) over a simulated Rayleigh fading channel. Two LIC models are adapted: a hyperprior model whose feature maps are masked and progressively transmitted, and a VQGAN model using residual vector quantization with codebook clustering. The channel simulation determines a per-slot bit budget from the current SNR and modulation scheme, and the receiver progressively decodes as more feature maps or residual codebooks arrive. Experiments on the Kodak dataset compare progressive-hyperprior, progressive-VQGAN, and adaptive WebP in terms of PSNR, SSIM, throughput (Mpps), and latency (Tavg and T99.9%). The paper claims that progressive-hyperprior achieves the lowest latency across all SNRs and higher throughput at low SNR, that progressive-VQGAN provides robustness without channel coding, and that adaptive WebP performs best at high SNR.","tokens_in":16874,"tokens_out":5442,"duration_ms":50610,"significance":"The paper addresses a relevant problem: low-latency and reliable image delivery over dynamic wireless channels using learned compression. The use of pretrained models, a publicly available code repository, and a transparent progressive-masking mechanism are strengths that facilitate reproducibility. However, the current evidence does not support the central reliability claims because the channel model only sets a bit budget and never injects bit errors, and the claimed improvement over non-progressive LIC baselines is not measured. The latency and throughput results are also largely a direct consequence of the chosen metric definitions. If the authors add a realistic bit-error model and direct non-progressive baselines, the framework could still be a valuable contribution to the field.","major_comments":[{"comment":"The transmission simulation never injects bit errors: Nbits is computed from the current SNR and modulation scheme, but the receiver decodes exactly the bytes and indices that were encoded. As a result, the reported PSNR/SSIM values measure loss due to truncation and masking, not resilience to channel noise. The abstract's claim that VQGAN is 'robust ... even in the absence of channel coding' and the conclusion that progressive-VQGAN operates 'without the need for channel coding' are therefore not tested. I request an experiment that actually corrupts bits in the ybytes/zbytes streams and in the VQGAN indices, or an equivalent error-sensitivity analysis, so that the reliability claims are directly evidenced.","section":"V.F and Eqs. (16)-(18), (23)-(24)"},{"comment":"The abstract states that the progressive transmission framework maintains or improves throughput 'compared to non-progressive counterparts,' but Table II contains no non-progressive LIC baselines. It only compares progressive-hyperprior, progressive-VQGAN, and adaptive WebP. Without evaluations of the non-progressive hyperprior and non-progressive VQGAN under identical channel conditions, the claimed improvement over non-progressive LIC models is not established. Please add those baselines or explicitly restrict the claim to the three compared methods.","section":"Abstract and Section VI.A (Table II)"},{"comment":"Because latency is measured in transmission slots and throughput counts each pixel exactly once, any scheme that transmits fewer bits per image will necessarily appear faster and achieve higher pixel throughput. The latency advantage of progressive-hyperprior at low SNR is thus a direct consequence of transmitting fewer feature maps within the bit budget, and the explanation in Section VI.B that 'compact feature maps' cause the advantage is circular. Please report these metrics as functions of the actual compressed size, or reframe the conclusions in terms of a rate-quality-latency trade-off.","section":"V.D and VI.B"},{"comment":"The text says Adaptive WebP 'cannot transmit at -10, -5, and 0 dB due to channel capacity constraints,' but Table II reports a nonzero throughput and PSNR/SSIM values for Adaptive WebP at 0 dB. This contradiction needs to be corrected and the exact failure condition of WebP clarified.","section":"VI.B"}],"minor_comments":[{"comment":"The final sentence 'where Adaptive WebP fails.' is a sentence fragment attached to the preceding text; please merge it into the previous sentence or remove it.","section":"Abstract"},{"comment":"The hyperprior model description contains an incomplete sentence: 'and is trained using rate-distortion parameter λ = 0.1, and for using 0-100 % of the channels. This particular model is' — please complete the description.","section":"V.A.2.a"},{"comment":"The numeric formatting is inconsistent (e.g., '272 .00', '26 .20', '34 .00'); please use uniform spacing and define the dash used for missing values.","section":"Table II"},{"comment":"The caption refers to 'observer-based feature masking' while the text describes importance-based sorting; please align the terminology.","section":"Figure 3"},{"comment":"The word 'Untimately' should be 'Ultimately'.","section":"V.E.1"},{"comment":"The unit 'Mpps' is not defined in the notation table; please specify that it means megapixels per second.","section":"V.D"}],"recommendation":"major_revision","confidential_remarks":"The revision path is feasible: the authors can add bit-error injection and non-progressive baselines, and revise the reliability and latency claims accordingly. My main concern is that the headline conclusions may change once actual bit errors are modeled, because entropy-coded hyperprior bitstreams are fragile and would likely require FEC or ARQ, whereas VQGAN index errors may degrade more gracefully. The authors should be prepared for that outcome when they perform the additional experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper before citing it: it takes two existing progressive learned image compression ideas (ProgDTD-style masked hyperprior and BRIC-style residual VQ) and bolts them onto a wireless transmission simulation, measuring waiting time and throughput. That extension is real but modest; the novelty is the application, not the method. The waiting-time analysis is a genuine addition, and the code is open source, which is good.\n\nWhat the paper does well: it defines a clean adaptive bit-budget procedure, evaluates across SNR values, and gives a practical comparison against adaptive WebP. The progressive-hyperprior and progressive-VQGAN pipelines are clearly described, and the authors are honest that they use pretrained models. The latency metric (average and 99.9th percentile waiting time) is more relevant for delay-sensitive applications than the usual inference-time focus, and that is a useful contribution.\n\nThe soft spots are serious. The main one, which the stress-test note gets right, is that the Rayleigh channel only sets the per-slot bit budget; no bit errors are ever injected into the transmitted arithmetic-coded or codebook-index bitstreams. So the central reliability claim—that VQGAN is robust without channel coding and that hyperprior needs FEC—is never tested. A flipped bit in an arithmetic-coded stream can desynchronize the decoder, while a flipped VQGAN index only swaps a codeword. The paper's low-SNR ordering could plausibly reverse under a realistic BER model. That is load-bearing, not a minor gap.\n\nSecond, the throughput and latency advantages are partly circular: sending fewer bits trivially lowers latency and raises pixels-per-second throughput. The paper never includes a non-progressive LIC baseline, so we cannot see whether the progressive version is anything more than 'transmit less data.' Third, the modulation and coding scheme is never specified, which makes the channel simulation hard to reproduce. The abstract also ends with a dangling 'where Adaptive WebP fails,' which should have been caught.\n\nIn short, this is a well-motivated engineering paper with a promising pipeline but the experimental evidence does not support its headline claims. The underlying ideas are salvageable with a proper end-to-end channel simulation (including FEC/ARQ options) and a non-progressive control. I would send it to review rather than desk-reject, because the problem is real and the framework is worth the referee time—but I would expect major revision. For your own work, cite the progressive masking idea with care, not the reliability results.","headline":"A sensible but thinly validated application of progressive LIC to wireless links; the reliability claims rest on an unmodeled error-free bit pipe.","tokens_in":755,"tokens_out":772,"would_cite":false,"duration_ms":22784,"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":"Progressive-hyperprior posts lowest latency on fading channels, and progressive learned codecs keep weak wireless links usable where Adaptive WebP fails.","keywords":["learned image compression","progressive transmission","hyperprior model","VQGAN","wireless image transmission","Rayleigh fading","reliability throughput latency","Kodak dataset"],"falsifier":"Re-run the same Kodak experiment but inject bit errors into the arithmetic-coded hyperprior stream and the VQGAN codebook indices at the bit error rates implied by the Rayleigh channel's SNR, without adding channel coding; if even modest error rates stop the hyperprior decoder or erase its latency and throughput edge over VQGAN, the central reliability claim is refuted.","tokens_in":16366,"feed_emoji":"📡","tokens_out":8139,"duration_ms":77712,"temperature":0.7,"pith_summary":"When a wireless channel fluctuates, transmitting an image as one fixed packet means waiting for the worst conditions and losing everything if a deep fade hits. This paper argues that learned image compression models can be made progressive: send the most important part of the compressed image first, then refine, so the receiver always has something usable and the waiting time drops. The authors demonstrate this for two learned codecs, a hyperprior autoencoder whose latent channels are ranked by importance, and a VQGAN whose token indices are sent coarse-to-fine through residual quantization. Over simulated Rayleigh fading channels, the progressive-hyperprior model achieves the lowest 99.9th-percentile waiting time at every SNR tested and the highest throughput at low SNR, while Adaptive WebP stops transmitting at the lowest SNRs. The practical stake is that IoT and task-oriented systems, which can act on a coarse image immediately, would get lower latency without giving up the ability to refine quality later.","feed_headline":"Progressive-hyperprior posts lowest latency on fading channels","feed_subtitle":"A learned codec that sends the most important feature maps first keeps images flowing when Adaptive WebP stalls.","key_machinery":"The central mechanism for the hyperprior model is channel-importance ranking of the latent feature maps: each of the 192 bottleneck channels is scored by the average MSE degradation when it is masked, and transmission sends the most important channels first, with the receiver padding the absent channels with zeros and arithmetic-decoding only the received bytes. For VQGAN, the central mechanism is residual vector quantization: a large codebook is clustered into codebooks of increasing bits-per-index, and progressive decoding sums residual codewords stage by stage, so early stages give a coarse image and later stages refine it. Both mechanisms are driven by a bit budget $N_{\\text{bits}}$ computed from the Rayleigh channel's current achievable rate, so the amount of data transmitted adapts to the channel in each slot.","core_discovery":"The paper claims that progressive, channel-adaptive transmission turns learned image compression into a practical wireless scheme. Instead of shipping the entire bitstream and hoping the channel holds, the system computes a per-slot bit budget from the current SNR and sends only what fits, starting with the most informative portion: for the hyperprior model, that means sorting the latent feature maps by how much reconstruction quality is lost when they are masked, then sending the top fraction; for VQGAN, it means sending indices from a coarse codebook first and then residual codebooks that progressively refine the image. On the Kodak benchmark across 1,000 Rayleigh fading realizations, progressive-hyperprior yields the lowest average and 99.9th-percentile waiting times from -10 to 5 dB, and higher throughput than progressive-VQGAN and Adaptive WebP at low and moderate SNR, while progressive-VQGAN gives the best PSNR and SSIM among the learned models under poor channel conditions, and Adaptive WebP produces the best image quality and throughput only at favorable SNR.","pith_inferences":["The paper's latency metric counts transmission slots, not encode and decode wall-clock time; an editorial next step is an end-to-end latency budget that includes those compute times, where VQGAN's avoidance of channel coding might change the ranking.","Because the simulation does not inject bit errors inside the transmitted streams, a direct extension is to re-run the comparison with actual bit errors and no error correction, testing whether the hyperprior's lossless arithmetic decoding stays reliable or whether VQGAN's codebook robustness wins.","The sorting-by-importance principle for hyperprior channels is not limited to the two codecs tested; the same budget-driven, most-important-first policy could be applied to other learned image codecs with structured latent spaces."],"forward_implications":["A receiver using progressive-hyperprior always has a decodable image within a couple of slots, even at -10 dB, because the most important channels arrive first and decode independently of the rest.","At -10 and -5 dB, progressive-hyperprior transmits roughly 18 and 66 megapixels per second respectively, while Adaptive WebP transmits zero, so the progressive learned schemes extend usable coverage to very low SNR.","The 99.9th-percentile waiting time of progressive-hyperprior is lower than both baselines at every SNR in the study, which matters for delay-sensitive applications that must meet a deadline almost always.","Setting the hyperprior parameter $N_{\\max}$ to 96 instead of 192 meets the same PSNR and SSIM targets at higher throughput, giving a practical knob for trading image quality against latency.","Progressive-VQGAN's decoder reconstructs images without relying on lossless compression or channel coding, so it can serve applications where computational simplicity and robustness under poor channels matter more than peak fidelity."],"supporting_citations":[{"why":"Supplies the hyperprior autoencoder architecture with lossless arithmetic coding that the paper makes progressive by channel sorting.","marker":"[9]"},{"why":"Supplies the VQGAN architecture and its vector-quantized latent space, the basis for the codebook-based progressive scheme.","marker":"[11]"},{"why":"Supplies the double-tail-drop progressive training and the feature-map sorting mechanism used to rank and transmit hyperprior channels.","marker":"[24]"},{"why":"Supplies the 100K-codebook VQGAN with a dimension-reducing projector that the paper clusters into bits-per-index codebooks.","marker":"[31]"},{"why":"Supplies residual vector quantization, the mechanism behind the progressive VQGAN refinement stages.","marker":"[25]"},{"why":"Defines the Adaptive WebP baseline that the paper compares against and that fails at low SNR.","marker":"[4]"},{"why":"Supplies the sum-of-sinusoids Rayleigh fading channel model that generates the per-slot bit budgets.","marker":"[35]"},{"why":"Provides the Kodak image benchmark on which all PSNR, SSIM, throughput, and latency results are reported.","marker":"[33]"}],"fun_headline_variants":["Progressive learned codec trims wireless image latency","Progressive hyperprior wins latency race on fading channels","Learned codec goes progressive to boost wireless image delivery","Progressive transmission cuts wait time for images on poor channels","Adaptive progressive codec improves latency and reliability in wireless"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reliability ordering rests on the assumption that, once the channel has set the per-slot bit budget, the transmitted compressed bytes arrive without bit errors; if they were actually corrupted, the hyperprior's lossless entropy decoder could fail and the VQGAN-versus-hyperprior reliability ranking would need to be tested directly.","fun_headline_variants_meta":{"raw":{"variants":["Progressive learned codec trims wireless image latency","Progressive hyperprior wins latency race on fading channels","Learned codec goes progressive to boost wireless image delivery","Progressive transmission cuts wait time for images on poor channels","Adaptive progressive codec improves latency and reliability in wireless"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2751,"prompt_tokens":1034,"completion_tokens":1717,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":1638}},"tokens_in":650,"tokens_out":1717,"duration_ms":13641,"temperature":1.0,"reasoning_tokens":1638,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:28:05.720041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same Kodak experiment but inject bit errors into the arithmetic-coded hyperprior stream and the VQGAN codebook indices at the bit error rates implied by the Rayleigh channel's SNR, without adding channel coding; if even modest error rates stop the hyperprior decoder or erase its latency and throughput edge over VQGAN, the central reliability claim is refuted.","supporting_citations":[{"cited_title":"ProgDTD: Progressive learned image compression with double-tail-drop training,","cited_arxiv_id":null,"evidence_quote":"Supplies the double-tail-drop progressive training and the feature-map sorting mechanism used to rank and transmit hyperprior channels."},{"cited_title":"BRIC: Bottom-up residual vector quantization for learned image compression,","cited_arxiv_id":null,"evidence_quote":"Supplies residual vector quantization, the mechanism behind the progressive VQGAN refinement stages."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Adaptive WebP baseline that the paper compares against and that fails at low SNR."},{"cited_title":"Novel sum-of-sinusoids simulation models for Rayleigh and Rician fading channels,","cited_arxiv_id":null,"evidence_quote":"Supplies the sum-of-sinusoids Rayleigh fading channel model that generates the per-slot bit budgets."},{"cited_title":"Kodak lossless true color image suite,","cited_arxiv_id":null,"evidence_quote":"Provides the Kodak image benchmark on which all PSNR, SSIM, throughput, and latency results are reported."}],"review_version":1}