{"id":"3e9f7984-1902-490a-8fa2-1cf08a2d3371","arxiv_id":"2411.17428","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Using both image features and CLIP text prompts to select among diffusion-generated candidates improves image reconstruction quality and low-SNR robustness in a simulated semantic communication system.","lead":"This paper transmits a short image description and text prompt instead of the image itself, then uses a diffusion model at the receiver to generate many candidate pictures and select the one whose semantic features match best. It is an incremental engineering proposal for semantic communication, relevant to anyone tracking 6G technologies that aim to send meaning rather than raw bits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I's 35.62 dB PSNR is likely a retrieval artifact: the receiver's diffusion model was trained on the complete CIFAR-100 and STL-10 sets, so 5,000 sampled candidates can contain the test image itself, turning selection into near-duplicate lookup.","rationale":"I read the paper as proposing a generate-and-select decoder for image semantic communication. For the headline claim to hold, the receiver's diffusion prior must supply a candidate set from which a genuinely noisy semantic feature vector can pick a faithful reconstruction. The most direct threat is that the candidate set contains the transmission source itself because the diffusion model was trained on the full test distribution. This is a real, testable threat and is distinct from the architectural value of multimodal fusion. The paper's OOD figure does not resolve it because those datasets are not in the training set and the figure lacks quantitative detail. A disjoint train/test rerun would settle the question: if PSNR collapses, the paper's central quantitative claim is unsupported; if it survives, the concern is refuted and the method has genuine merit. I agree with the reader's weakest-assumption analysis and see no reason to alter the REJECT verdict based on this stress test. The idea of using multimodal features to supervise selection is interesting, and the complexity analysis is a useful contribution, but no code or checkpoints are released, so independent verification of the numerical results is not currently possible.","tokens_in":36624,"tokens_out":3308,"duration_ms":32678,"concrete_test":"Train the receiver's diffusion model on a disjoint split, e.g., 80% of the images in each CIFAR-100 and STL-10 class, reserving the remaining 20% for testing, and repeat the Table I protocol at SNR = 10 dB for SeqMMSemCom. If PSNR drops from 35.62 dB to below roughly 22 dB while DeepJSCC and WITT baselines remain essentially unchanged, memorization and retrieval are responsible for the headline. As a supplementary check, under the original protocol, detect how often any of the 5,000 candidates is an exact or near-exact match (e.g., L2 distance below 1e-3) to the transmitted image; frequent matches would confirm that the selection stage degenerates to database lookup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section VI-A states that the receiver's diffusion model is trained on the complete CIFAR-100 and STL-10 training sets, while Section VI-C evaluates on images randomly selected from those same datasets. After the reported 50,000 training epochs on small datasets, diffusion models can memorize individual training images. With p = 5,000 conditionally sampled candidates at inference (Section VI-B), the original image or a near-identical copy is plausibly present in the candidate set. The selection mechanism in Eqs. (14)-(19) then retrieves that near-duplicate using the transmitted image and text features, yielding PSNR up to 35.62 dB and MS-SSIM 0.996 (Table I) without the receiver ever constructing the image from noisy semantics. This is a retrieval result, not a generative reconstruction result. The central claim of enhanced fidelity and robustness at low SNR depends entirely on this candidate-set overlap. The OOD experiments on FGVC-Aircraft and Stanford-Cars use datasets absent from training and cannot validate the in-distribution numbers; moreover, Fig. 11 is garbled and no quantitative table accompanies it. Thus the load-bearing assumption that the DM places a near-identical test image in the candidate set is unexamined and would likely be violated under a properly disjoint evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MMSemCom, a semantic communication framework that transmits multimodal semantic features (ResNet-50 image features and CLIP text prompts) and reconstructs the image at the receiver by generating a large candidate set with a class-conditional diffusion model and then selecting the best candidate using sequential or simultaneous selection mechanisms. The authors also extend the scheme to a multiuser scenario using Walsh codes and provide communication-overhead and computational-complexity analyses. The headline empirical claims are high reconstruction fidelity (e.g., 35.62 dB PSNR and 0.996 MS-SSIM for sequential MMSemCom at 10 dB SNR in Table I) and robustness at low SNR.","tokens_in":36959,"tokens_out":5819,"duration_ms":61560,"significance":"The generate-then-select paradigm is, in principle, an interesting alternative to direct neural decoding in semantic communication, and the manuscript contains a fairly complete architecture description with explicit overhead and complexity formulas. However, the empirical validation is undermined by a train/test overlap: the receiver's diffusion model is trained on the complete CIFAR-100 and STL-10 sets, while evaluation images are drawn from those same sets. With 50,000 training epochs and p=5,000 sampled candidates, the reported near-perfect PSNR/MS-SSIM figures are plausibly explained by retrieval of memorized near-duplicates rather than by generative reconstruction. A second architectural gap is that the class condition required by the diffusion model is not specified as being transmitted or recovered in the simultaneous mode. The OOD generalization claim is supported only by a garbled figure with no quantitative table. The paper does not provide code or model checkpoints, so the experiments are not independently reproducible.","major_comments":[{"comment":"The evaluation protocol allows the diffusion model to have memorized the test images. Section VI-A states that the receiver's diffusion model is trained on the complete CIFAR-100 and STL-10 image sets, and Section VI-C states that test images are randomly selected from those same datasets; no held-out split is described. After 50,000 training epochs on datasets of 60,000 and 13,000 images, and with p=5,000 diffusion samples per transmitted image at inference, the candidate set can contain the original image or a near-identical copy. The selection mechanism in Eqs. (14)-(19) then retrieves that near-duplicate using the transmitted image and text features, which explains the near-perfect PSNR (35.62 dB) and MS-SSIM (0.996) in Table I. This makes the headline result a retrieval artifact rather than a generative reconstruction result. The authors must evaluate on a properly disjoint train/test split and report the overlap rate between the generated candidate set and the target image, or otherwise demonstrate that the reported gains persist for images not in the training set.","section":"VI-A, VI-C, Table I"},{"comment":"The receiver's diffusion model is class-conditional (Eq. (9), Algorithm 1), but the manuscript never specifies how the receiver obtains the class condition c. In the sequential method, the transmitted text prompts may indirectly carry class information, but in the simultaneous method the transmitter sends only the fused feature vector FFused (Section VI-B). No class label or equivalent condition is included in the transmitted overhead (Eq. (25)-(26)), and no mechanism is described for recovering c from the received fused feature. Without a specification of how c is known at the receiver, the conditional generation step is not well defined, and the claimed end-to-end operation cannot be reproduced. Please clarify how c is transmitted or inferred, and update the overhead analysis accordingly.","section":"III-B1, III-B2, Section VI"},{"comment":"The out-of-distribution generalization claim is not quantitatively supported. Section VI-D4 asserts that SeqMMSemCom and SimMMSemCom exhibit only minor degradation on FGVC-Aircraft and Stanford-Cars and maintain a significant advantage over BPG and LaMoSC, but Fig. 11 is rendered as unreadable placeholder text and no numerical PSNR/MS-SSIM results or comparison table are provided. Without numbers, the reader cannot verify the claimed OOD robustness. Provide a table with mean PSNR, MS-SSIM, and baselines on these datasets.","section":"VI-D4, Fig. 11"},{"comment":"The comparison with BPG and WebP is not a fair channel comparison. BPG and WebP are source codecs, yet they are plotted as functions of SNR and transmission distance in Figs. 7-10, without specifying how compressed bits are mapped to channel symbols, protected, and transmitted over the AWGN or Rayleigh channel. To support the claim of outperforming 'existing communication systems,' all schemes should be evaluated under the same channel-use/bit-rate constraints and the same channel model. As presented, the gains over BPG and WebP are not well defined.","section":"VI-C, Figs. 7-10"}],"minor_comments":[{"comment":"The summation indices in the distance formulas are inconsistent with the feature dimensions: in Eq. (14) the sum runs over p (the number of generated images) while the text feature FText(x) has dimension ntext × dtext; in Eq. (17) the sum runs over q, but FImage(x) is a single vector of dimension dimage; in Eq. (19) the sum again runs over p for a fused feature of dimension dfusion. Please correct these expressions or clarify the intended averaging.","section":"Eqs. (14)-(19)"},{"comment":"The training budget is described as '50,000 epochs' in Section VI-A and '50,000 iterations' in the bullet points of Section VI-C; please use consistent terminology, since the number of epochs is highly relevant to the memorization concern.","section":"VI-A, VI-C"},{"comment":"The table header says 'DeePJSCC' while the text and references use 'DeepJSCC'; please fix the capitalization.","section":"Table III"},{"comment":"In the path-loss paragraph, 'in Eq. 2' appears to refer to the channel model in Section II, but that equation is not numbered; please add the equation number or fix the cross-reference.","section":"VI-D1"}],"recommendation":"reject","confidential_remarks":"The train/test overlap in Section VI is a foundational evaluation flaw, and the unspecified class-condition recovery in the simultaneous mode is a structural gap. These issues are not cosmetic; they affect the validity of every reported performance number. If the authors can resubmit with a disjoint train/test evaluation, a specified condition-recovery mechanism, and quantitative OOD results, the generate-then-select idea might be viable as a retrieval-based semantic communication system, but the current manuscript's central claims are unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"At its core, this paper tests a reasonable idea: use both image and CLIP text features to pick the best candidate from a diffusion-generated set, instead of decoding directly from noisy semantics. That's a clean extension of the authors' earlier selection-based system, and they add a fused attention variant and a Walsh-code multiuser scheme. The complexity and overhead analysis is a nice touch, and Table II (prompt count ablation) shows they were thinking about design trade-offs.\n\nThe problem is the load-bearing experiment. The receiver's diffusion model is trained on the complete CIFAR-100 and STL-10 training sets (Section VI-A), and the test images are randomly drawn from those same datasets (Section VI-C). With 50,000 epochs and p=5,000 candidates per image, the candidate set can contain a near-duplicate of the target; the selection mechanism then retrieves it, not reconstructs it. That reading is supported by the paper's own sentence in the MS-SSIM analysis: 'The images produced through diffusion training closely resemble the original images.' The 35.62 dB PSNR and 0.996 MS-SSIM are exactly what you'd expect from near-duplicate lookup. This makes the headline comparison against DeepJSCC, WITT, and LaMoSC uninformative. The OOD tests on FGVC-Aircraft and Stanford-Cars were the right idea, but Fig. 11 is garbled and there is no table, so they don't rescue the paper.\n\nSmaller issues: the simultaneous method's cliff at low SNR is explained as noise sensitivity, but it could just be that the fused feature is less discriminative for retrieval; the paper doesn't separate those. The inference time in Table III says 35ms, which cannot include sampling 5,000 diffusion candidates; that number needs a footnote or clarification. And they never compare against a simple retrieval baseline (e.g., transmitting enough feature bits to look up the image from a shared database), which would sharpen what the generative model actually contributes. It's also not clear how the attention fusion module is trained—jointly with the selection or separately—and that matters for reproducibility.\n\nNet: the idea deserves a test, but this is a pilot study, not a demonstration. To be convincing they'd need a disjoint train/test split, a retrieval baseline, quantitative OOD results, a description of the fusion training, and a clear statement of the true inference cost. I would not accept these numbers as they stand, but I'd send it to a serious referee because the architecture is coherent, the problem is real, and the fix is expensive but well-defined. It could become a solid contribution after major revision.","headline":"Promising selection-based idea, but the headline PSNR numbers are likely retrieval artifacts from training-set overlap.","tokens_in":37430,"tokens_out":3857,"would_cite":false,"duration_ms":32463,"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":"This paper claims that transmitting both image features and CLIP text prompts lets a receiver generate thousands of candidate images and select the best one, keeping reconstruction quality high even under noisy channels.","keywords":["semantic communication","multimodal semantics","diffusion model","image reconstruction","CLIP","selection mechanism","wireless image transmission","generative model"],"falsifier":"Transmit a test image whose class or visual style was excluded from diffusion training and measure PSNR at 10 dB SNR; if the best candidate among the 5,000 generated images is far from the target, the selection mechanism cannot recover the reported 35 dB and performance falls to the quality of the closest class-conditional sample.","tokens_in":36426,"feed_emoji":"📡","tokens_out":5307,"duration_ms":46248,"temperature":0.7,"pith_summary":"The paper proposes a semantic communication system in which the receiver does not reconstruct an image from noisy features directly. Instead, the transmitter sends compact multimodal features, a CNN image descriptor plus CLIP text prompts, and the receiver's diffusion model generates a large set of candidate images from which the closest match is selected. The central claim is that this generation-and-selection decoding keeps image fidelity high at low signal-to-noise ratios, with the sequential variant reporting 35.62 dB PSNR at 10 dB SNR on in-distribution test images. A sympathetic reader would care because it recasts the hard decoding problem as a retrieval problem over clean generated images, which is more tolerant of channel noise.","feed_headline":"Image delivery by selection beats direct decoding at low SNR","feed_subtitle":"A receiver generates 5,000 image candidates and picks the best using image and text semantics.","key_machinery":"The central object is the MultiSem multimodal semantic representation: a ResNet-50 image feature vector and CLIP text prompts, either concatenated or fused through a multi-head cross-attention network with residual connections and layer normalization. The receiver uses a class-conditional diffusion U-Net to generate a candidate image set, and a selection module scores each candidate by a combined cosine and Euclidean distance against the received features. Text prompts act as a coarse filter that narrows the candidate set, and image features then refine the choice, which is what makes sequential MMSemCom robust at low signal-to-noise ratios.","core_discovery":"Adding text-level semantics to image-level semantics for supervising generation improves both fidelity and robustness in wireless image transmission. The system extracts image features with a ResNet-50 backbone and derives text prompts with CLIP, combines them into a multimodal semantic representation called MultiSem, and transmits that representation over AWGN or Rayleigh channels. At the receiver, a class-conditional diffusion U-Net trained on the same dataset samples 5,000 candidate images, and a selection module ranks them by averaged cosine and Euclidean distances to the received multimodal features. Sequential selection, which filters first by text then by image features, is the most noise-robust variant and reaches 35.62 dB PSNR at 10 dB SNR, while simultaneous selection using only 256 bytes of fused features reaches 30.75 dB. The paper argues that because noise affects which candidate is chosen rather than the pixels of the output, the system avoids the performance cliff that end-to-end decoders show at low SNR.","pith_inferences":["If the candidate set is generated offline and fixed, the receiver-side selection is effectively a nearest-neighbor search in a shared embedding space; channel noise then perturbs only the query features, so the scheme is a form of robust retrieval rather than generation.","The best achievable PSNR is capped by the best candidate among the generated images, so increasing the candidate count should raise that ceiling at the price of linearly more sampling; a PSNR-versus-N curve would quantify this trade-off directly.","The simultaneous mode's 256-byte overhead invites a direct comparison with conventional codecs transmitting the same byte budget, which would isolate whether the gain comes from the fused features or from the diffusion prior.","The authors' stated future move to text-to-image synthesis would remove the dependence on the training dataset, but it would also change the claim from pixel-level fidelity of a known image to semantic fidelity of a described scene."],"forward_implications":["Channel noise acts on the selection query rather than on the output pixels, so the decoded image is always a clean sample from the diffusion model's learned distribution.","Sequential MMSemCom sustains above 25 dB PSNR even at 0 dB SNR and reaches 35.62 dB at 10 dB SNR on in-distribution data, avoiding the sharp degradation seen in end-to-end baselines.","Simultaneous MMSemCom transmits only 256 bytes of fused features and still achieves 30.75 dB PSNR at 10 dB SNR, giving the highest PSNR-per-kilobyte efficiency among the compared modes.","Increasing the number of CLIP prompts improves selection accuracy up to about three to five prompts, after which communication overhead grows faster than reconstruction quality.","Walsh-code orthogonal multiplexing extends the generation-and-selection framework to multiple users, with both users maintaining high PSNR and MS-SSIM once the SNR is above roughly 10 dB."],"supporting_citations":[{"why":"Earlier selection-based image generation system that this work extends by adding text supervision.","marker":"[22]"},{"why":"Supplies the CLIP model used to extract text prompts from images.","marker":"[24]"},{"why":"Provides the conditional text-to-image diffusion generation mechanism the receiver adapts for class-conditional sampling.","marker":"[26]"},{"why":"Supplies the ResNet-50 architecture used as the image feature extractor.","marker":"[27]"},{"why":"UniPC scheduler used to accelerate diffusion sampling to 40 inference steps.","marker":"[30]"},{"why":"DeepJSCC is a main end-to-end baseline the proposed system is compared against.","marker":"[31]"},{"why":"LaMoSC is a large-language-model semantic communication baseline the proposed system is compared against.","marker":"[32]"},{"why":"WITT is a transformer-based wireless image transmission baseline the proposed system is compared against.","marker":"[33]"}],"fun_headline_variants":["Multimodal selection boosts image transmission at low SNR","Pick the best of 5,000 generated images for wireless fidelity","Text and image cues jointly select best generated image","Diffusion generation plus semantic selection beats direct decoding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported high PSNR depends on the receiver's diffusion model having been trained on the same dataset from which the test images are drawn, so a near-identical copy of the target image is likely to exist somewhere among the 5,000 generated candidates.","fun_headline_variants_meta":{"raw":{"variants":["Multimodal selection boosts image transmission at low SNR","Pick the best of 5,000 generated images for wireless fidelity","Text and image cues jointly select best generated image","Diffusion generation plus semantic selection beats direct decoding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1378,"prompt_tokens":993,"completion_tokens":385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":321}},"tokens_in":609,"tokens_out":385,"duration_ms":4506,"temperature":1.0,"reasoning_tokens":321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:07:35.539488+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Transmit a test image whose class or visual style was excluded from diffusion training and measure PSNR at 10 dB SNR; if the best candidate among the 5,000 generated images is far from the target, the selection mechanism cannot recover the reported 35 dB and performance falls to the quality of the closest class-conditional sample.","supporting_citations":[{"cited_title":"Selection-based image generation for semantic communication systems,","cited_arxiv_id":null,"evidence_quote":"Earlier selection-based image generation system that this work extends by adding text supervision."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"Supplies the CLIP model used to extract text prompts from images."},{"cited_title":"Conditional text image generation with diffusion models,","cited_arxiv_id":null,"evidence_quote":"Provides the conditional text-to-image diffusion generation mechanism the receiver adapts for class-conditional sampling."},{"cited_title":"Unipc: A unified predictor- corrector framework for fast sampling of diffusion models,","cited_arxiv_id":null,"evidence_quote":"UniPC scheduler used to accelerate diffusion sampling to 40 inference steps."},{"cited_title":"Deep joint source- channel coding for wireless image transmission,","cited_arxiv_id":null,"evidence_quote":"DeepJSCC is a main end-to-end baseline the proposed system is compared against."},{"cited_title":"Lamosc: Large language model-driven semantic communication system for visual trans- mission,","cited_arxiv_id":null,"evidence_quote":"LaMoSC is a large-language-model semantic communication baseline the proposed system is compared against."},{"cited_title":"Witt: A wireless image transmission transformer for semantic communications,","cited_arxiv_id":null,"evidence_quote":"WITT is a transformer-based wireless image transmission baseline the proposed system is compared against."}],"review_version":1}