{"id":"ab57ccfb-41f0-4727-b85d-6d694e777554","arxiv_id":"2501.12988","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"An LLM-based semantic communication system for image transmission, using LLaVA to encode images as text and Stable Diffusion to decode, claims 4250x compression over OFDM links.","lead":"The authors propose an image transmission system that converts images to text captions using an LLM, sends the captions over an OFDM wireless link, and reconstructs images from the captions using a text-to-image model. They claim a 4250x data size reduction and a higher effective data rate than conventional transmission, but the evaluation lacks a fair baseline and statistical rigor.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central compression/data-rate claims rest on an unvalidated assumption that a text caption is a sufficient semantic representation; the paper's own Fig. 4 example shows semantic content loss, and no task-oriented metric supports the claim.","rationale":"The reader's rejection is justified, and the weakest assumption identified is exactly the one I consider load-bearing: the paper assumes without evidence that a text caption preserves enough semantic content for the reconstructed image to be an acceptable substitute for the original. The qualitative example in Fig. 4 actually demonstrates content loss (the wooden post disappears) rather than semantic preservation, and the reported SSIM and BLEU scores do not measure task-level or human-judged fidelity. Because the 4250x compression ratio and the higher effective data rate both depend on treating this lossy caption as the transmitted representation, the central claims are unsupported. My recommended verdict remains REJECT, so no adjustment to the reader's verdict is needed. The paper does integrate standard open-source components and uses the Sionna simulator, which is a reasonable tool, but the absence of any task-oriented evaluation, statistical support, or reproducible data is fatal to the central claim.","tokens_in":7679,"tokens_out":3974,"duration_ms":42087,"concrete_test":"Run the full pipeline on 500 diverse images from MS COCO validation at SNR=3.5 dB. Compare original versus reconstructed images using (i) CLIP image-text alignment with the transmitted caption, (ii) zero-shot classification accuracy, and (iii) object detection recall for objects named in the caption. If CLIP similarity drops by more than 10% relative to original images, or if classification accuracy or detection recall degrades significantly, the semantic representation is insufficient and the compression/data-rate claims are unsupported. Additionally, recompute Table II at a single stated SNR with all bit counts and success rates reported; if the 23.05 and 19.34 Mbps values cannot be reproduced, the data-rate advantage is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline numbers are only meaningful if LLaVA's text caption is a sufficient semantic representation of the image, so that Stable Diffusion can reconstruct an acceptable substitute. This premise is asserted in Section II-B (Algorithms 1 and 2) but never validated. A caption is a lossy, language-biased summary: it drops object positions, counts, spatial relations, exact colors, text/OCR content, and fine structure. Fig. 4's own example confirms this loss: at SNR=3.3 dB the received prompt loses \"wooden post\" and the reconstructed image omits the post, yet the paper treats this as success. The SSIM metric in Fig. 2 is not a valid semantic fidelity measure here because SSIM rewards pixel-level similarity, which this pipeline is structurally unable to provide; semantic communication should be evaluated by task success or human-judged semantic preservation, neither of which is reported. Consequently, the claimed 4250x reduction is not compression of the image but substitution of a text summary for the image, and Table II's effective data rate inherits a comparison of unlike quantities: bits of a caption versus bits of a raw image, at an unspecified operating SNR. If reconstructed images fail task-oriented metrics on a diverse benchmark, the central claims collapse.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an end-to-end OFDM-based semantic communication system for image transmission. At the transmitter, a LLaVA-based semantic encoder converts an image into a short text prompt; after LDPC coding, QAM modulation, and OFDM transmission over a 3GPP Urban Macro-cell channel model, the receiver decodes the prompt and uses a CLIP text encoder together with Stable Diffusion to regenerate an image. The authors report SSIM and BLEU results versus SNR and claim a 4250-fold reduction in transmitted data size (5.916 Mb to 1.392 Kb) together with a higher effective data rate (23.05 Mbps vs. 19.34 Mbps) than a conventional communication method.","tokens_in":7935,"tokens_out":6168,"duration_ms":62162,"significance":"If the claims were supported, the work would be a potentially interesting demonstration that LLM-generated captions can serve as the semantic source representation for image transmission over realistic channels, achieving compression far beyond conventional source coding. The use of a 3GPP UMa channel model rather than only AWGN, the end-to-end integration of a full physical-layer chain, and the reliance on public tools (Sionna, LLaVA, Stable Diffusion) are all strengths that make the system reproducible in principle. However, the current evidence does not establish the central claims: the evaluation rests on a single illustrative image, no test corpus or error bars are reported, the conventional baseline is not defined, and the adopted metrics (SSIM, and BLEU measured against the encoder's own prompt) do not measure semantic fidelity or task utility. The headline compression ratio conflates source substitution (sending a caption instead of pixels) with compression, so the quantitative claims are not yet credible.","major_comments":[{"comment":"The evaluation reports no dataset, no number of test images, no confidence intervals, and no description of how the SSIM and BLEU curves in Figs. 2 and 3 are averaged. The only concrete example is a single bird image in Fig. 4. Without a defined test corpus and protocol, the central quantitative claims (4250x data reduction and 23.05 vs. 19.34 Mbps) are not reproducible, and the reported curves cannot be meaningfully assessed.","section":"Section IV, Table II"},{"comment":"The 'conventional communication' baseline is never specified. If it transmits the raw uncompressed image, then the 4250x ratio compares a text caption to raw pixels and is not a source-compression ratio; if it uses a compressed format or a learned codec, that must be stated and the comparison must be made at equal end-to-end rate. The manuscript also does not state the SNR at which the effective data rates in Table II are computed, even though the success-rate factor S = Bs/Bt presumably depends on the operating SNR.","section":"Section IV, Table II and Fig. 2"},{"comment":"The BLEU score is computed between the received prompt and the prompt generated by the semantic encoder at the transmitter, which is part of the system under test rather than an external ground truth. A high BLEU score therefore only shows that the wireless channel delivered the transmitted caption with few errors; it says nothing about whether the caption faithfully represents the image content. A task-oriented metric (e.g., retrieval, classification, caption-based QA) or human judgment of semantic preservation is needed to validate the semantic representation.","section":"Section IV, Fig. 3"},{"comment":"SSIM is used as the main reconstruction metric, but the decoder deliberately generates a new image from a text description, so pixel-level similarity to the original is not the appropriate notion of fidelity. The authors acknowledge this variation, yet the claimed 'superior' low-SNR performance is an artifact of comparing a generated image with the original using a metric that the pipeline is structurally unable to maximize. The example in Fig. 4 at SNR=3.3 dB actually shows loss of the wooden post in both the prompt and the reconstructed image, which is a semantic content error; treating this as a successful reconstruction is not justified.","section":"Section IV, Figs. 2 and 4"},{"comment":"The definitions of Bs, Bt, N, and Tsymbol are not tied to the numerical values in Table II. The table reports data sizes of 1.392 Kb and 5.916 Mb and rates of 23.05 and 19.34 Mbps, but the manuscript does not show how the success rates and the total transmission time T were obtained for each system, nor how many OFDM symbols were used for the conventional transmission. Without these details, the comparison in Table II cannot be verified.","section":"Section IV, Effective Data Rate"}],"minor_comments":[{"comment":"The paper repeatedly describes the receiver as an 'LLM-based semantic decoder', but Section III-B shows that the decoder is a CLIP text encoder plus Stable Diffusion; the wording should be corrected for consistency.","section":"Abstract and Section III-B"},{"comment":"The quantity Hse(X̂) is introduced as the entropy of the reconstructed data, but the notation and the surrounding explanation are not sufficient to connect it to the semantic channel capacity formula; moreover, this theoretical framework is not used in the experimental section.","section":"Section II-A, Eq. (4)"},{"comment":"The figure label says 'BLUE score' instead of 'BLEU score'.","section":"Section IV, Fig. 3"},{"comment":"There are numerous typos and grammar errors, including 'recontacted image quality', 'the semantic information prompt was successfully recovered', 'wnoden p st', 'provinding', and 'by by'; the manuscript should be carefully proofread.","section":"Section IV"},{"comment":"The functions EnI, Enp, Align, and Autoreg are not specified; since the implementation is based on LLaVA, the manuscript should state the exact model version, prompt template, and alignment mechanism used.","section":"Algorithm 1"},{"comment":"The table gives subcarrier spacing and the number of OFDM symbols but not the total transmission time T; stating T would make the data-rate calculation in Table II directly verifiable.","section":"Table I"},{"comment":"The paper cites several semantic communication systems but provides no quantitative comparison against any of them (e.g., DeepJSCC or other image-oriented semantic communication baselines); a comparison table would help position the contribution.","section":"Related Work and Section IV"},{"comment":"The reconstructed images referenced as ①, ②, ③, and ④ in the text are not clearly identified in the figure caption, and the caption does not state which panel corresponds to which SNR; please clarify.","section":"Fig. 4"}],"recommendation":"reject","confidential_remarks":"The manuscript reads as a preliminary demonstration rather than a complete study. The lack of a defined test dataset, an unspecified conventional baseline, and the use of metrics that do not measure semantic fidelity make the headline claims unverifiable in their current form. If the authors add a systematic evaluation with task-oriented metrics on a standard dataset, a clearly specified compression baseline, and full details of the data-rate computation, a resubmission could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a straightforward integration of off-the-shelf components—LLaVA for captioning, CLIP plus Stable Diffusion for generation—over a standard OFDM/LDPC/QAM link, and the headline numbers (4250x compression, higher data rate) do not survive contact with the evidence.\n\nWhat is actually useful: the authors build a complete semantic communication chain in Sionna, including a 3GPP UMa channel with mobility, which is more realistic than the AWGN used in many semantic comm papers. The system works in the limited demo sense: a bird image is captioned, the text bits get through, and Stable Diffusion produces a plausible bird. As a demonstration that off-the-shelf vision-language models can be dropped into a physical-layer pipeline, it is fine.\n\nThe soft spots are substantial. The 4250x ratio compares a short caption to an uncompressed raw image; that is substitution, not compression, and a conventional baseline that does not use JPEG or any source coding is a strawman. There is no dataset, no error bars, no diversity of test images—the whole evaluation rests on one example. The BLEU score in Fig. 3 compares the received prompt to the encoder's own output, which is circular and says nothing about preservation of image content. SSIM is the wrong instrument here: a diffusion model is structurally incapable of pixel-level fidelity, so the SSIM curves are expected to cross; they do not establish semantic superiority. The effective data rate in Table II is computed from numbers that are never shown—there is no derivation of R, S, or T, and comparing a 1.392 Kb caption against 5.916 Mb of raw image bits mixes unlike quantities.\n\nThe most damaging evidence is the paper's own Fig. 4. At SNR=3.3 dB the received prompt loses \"wooden post,\" and the reconstructed image lacks the post; the authors call that success. That is an admission that the semantic channel drops information that matters for faithful reconstruction. If the goal is task-oriented communication, they need a task metric; if the goal is reconstruction, they need a baseline that actually compresses.\n\nThe writing is rough in places (\"BLUE\" for BLEU, \"recontacted image\"), and the semantic entropy/capacity theory in Section II is never connected to the design or evaluation.\n\nWho is this for? A workshop demo audience, or a reading group discussion about evaluation pitfalls in semantic communication. It is not yet a paper whose central claims a reviewer can verify. I would desk reject it as a full paper. If the authors add a real baseline (JPEG, deep JSCC), a proper dataset, task-oriented metrics, and a transparent rate calculation, it could become a modest empirical study worth a second look.\n\nRecommendation: not for peer review in current form.","headline":"The 4250x compression claim is an artifact of comparing a caption to raw image bits; the system is an off-the-shelf integration whose evaluation doesn't support its headline numbers.","tokens_in":8451,"tokens_out":3652,"would_cite":false,"duration_ms":38492,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes transmitting images as short text captions generated by a vision-language model and regenerated by a diffusion model at the receiver, cutting transmitted data 4,250-fold while raising effective data rate.","keywords":["semantic communication","large language models","image transmission","OFDM","6G","vision-language models","diffusion models","UMa channel"],"falsifier":"Take a set of images containing small task-critical details such as street signs, license plates, or fine-grained species, run the described LLaVA-to-Stable-Diffusion pipeline over the UMa channel at high SNR, and compare a downstream classifier's accuracy on the original images with its accuracy on the reconstructed ones. If accuracy drops substantially even though SSIM stays high, then the caption bottleneck does not preserve the semantic content that the claimed 4,250-fold compression is supposed to carry.","tokens_in":7485,"feed_emoji":"📡","tokens_out":5664,"duration_ms":57528,"temperature":0.7,"pith_summary":"This paper proposes an end-to-end semantic image-transmission system in which a vision-language model at the transmitter turns an image into a short text caption, and a diffusion model at the receiver turns the caption back into an image. Over a 3GPP urban-macrocell OFDM link, the system sends about 1.392 Kb instead of 5.916 Mb per image, a 4,250-fold reduction, and achieves an effective rate of 23.05 Mbps versus 19.34 Mbps for conventional transmission. The authors report that the reconstructed image remains recognizable even when the channel corrupts parts of the caption, because the receiver regenerates from meaning rather than from exact bits. The pay-off, if true, is a concrete route to the extremely low-overhead image delivery envisioned for 6G, at the price of replacing bit-exact fidelity with semantic fidelity.","feed_headline":"LLM sends images as text prompts: 4,250x less data","feed_subtitle":"A caption-based link beats raw image transmission in effective rate and stays recognizable even when the channel corrupts the text.","key_machinery":"The load-bearing mechanism is the text prompt as a semantic bottleneck: Algorithm 1 compresses an image into caption tokens through an autoregressive vision-language model, and Algorithm 2 embeds those tokens with a CLIP text encoder and regenerates an image through a denoising diffusion process. This pair converts bit errors in the channel into semantic errors, so a corrupted token changes a few letters of the caption while the overall meaning remains recoverable. The channel is a full OFDM chain with LDPC coding, 4-QAM, pilots, channel estimation, and equalization over the UMa model, which is what makes the data-size and rate numbers physically meaningful.","core_discovery":"The paper's central claim is that semantic content, not raw pixels, is what needs to cross a wireless channel, and that modern vision-language and generative models make this practical. The proposed system uses LLaVA as the semantic encoder: its image encoder and aligned text decoder produce a token-by-token prompt such as \"A brown and white bird perched on a wooden post.\" At the receiver, a CLIP text encoder embeds the received prompt and Stable Diffusion generates an image conditioned on that embedding. Evaluated over a 28 GHz UMa channel with 60–120 km/h mobility, the prompt-based system transmits 4,250 times less data and achieves a higher effective data rate than the conventional LDPC-coded QAM-OFDM baseline, because it sends only the text that describes the image. At low SNR the semantic system has higher SSIM than the conventional one; at high SNR the conventional system has higher SSIM, which the authors attribute to the unavoidable variation in generative reconstruction.","pith_inferences":["Beyond the paper's claims, the 4,250-fold ratio compares a lossy semantic description against a bit-exact source; a fair engineering comparison would also count the bits an application actually consumes, so the ratio may be smaller for tasks that need fine-grained details such as text inside the image.","A reader might test the semantic-fidelity claim by replacing SSIM with task-level metrics: if the caption omits a stop sign's text or a vehicle's license plate, a downstream detector on the reconstructed image would fail while SSIM looks acceptable.","The prompt itself is a high-level description, so the scheme implicitly converts image transmission into a text-understanding problem; this suggests a possible extension where the same channel carries editing instructions such as \"same scene at night\" rather than the original caption, turning the link into a generative image-editing channel.","Because the receiver's diffusion model is fixed and pretrained, the system's low-SNR robustness is partly a property of the generative prior rather than of the wireless chain; a different receiver prior would shift the SNR at which the semantic and conventional curves cross."],"forward_implications":["If the caption bottleneck is accepted, image delivery over 6G links can be planned in terms of hundreds of bytes per image rather than megabytes, shifting bandwidth budgets from source coding to semantic coding.","The system's robustness at low SNR means a corrupted caption still yields a recognizable image, so link-budget margins for image traffic could be relaxed compared with bit-exact transmission.","Because the semantic decoder always regenerates something, there is an inherent ceiling on pixel-level fidelity at high SNR; applications that need exact images would keep conventional transmission, while meaning-oriented applications would use the semantic pipeline.","The paper positions the same LLM-based semantic coding as extendable to other data modalities, making language the common representation across future intelligent communicating nodes."],"supporting_citations":[{"why":"Supplies LLaVA, the vision-language model used as the semantic encoder that maps images to caption text.","marker":"[23]"},{"why":"Supplies Stable Diffusion, the semantic decoder that regenerates the image from text conditioning.","marker":"[27]"},{"why":"Supplies CLIP's text encoder, which turns the received prompt into a fixed-dimension semantic latent for generation.","marker":"[26]"},{"why":"Supplies the denoising diffusion probabilistic model mechanism behind the image generation.","marker":"[28]"},{"why":"Provides the OFDM, LDPC, and UMa channel simulation used in the evaluation.","marker":"[21]"},{"why":"Defines the 3GPP urban-macrocell channel model with mobility used as the physical propagation scenario.","marker":"[19]"},{"why":"Supplies the Vision Transformer architecture used inside LLaVA to extract latent image features.","marker":"[24]"},{"why":"Supplies the LLaMA language model fine-tuned inside LLaVA for autoregressive token generation.","marker":"[25]"}],"fun_headline_variants":["LLM sends images as captions, cutting data 4,250x","Semantic link: LLM shrinks image data 4,250x","LLM-based comms: images as text, 4,250x less data","From pixels to prompts: LLM cuts data 4,250x","LLM-powered semantic link: images as captions, 4,250x smaller"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The promised gain rests on the assumption that a short caption is a faithful enough stand-in for the image's meaning, so that an image regenerated from the caption is an acceptable substitute for the original; the paper never validates this against a task, a human reader, or an external semantic ground truth.","fun_headline_variants_meta":{"raw":{"variants":["LLM sends images as captions, cutting data 4,250x","Semantic link: LLM shrinks image data 4,250x","LLM-based comms: images as text, 4,250x less data","From pixels to prompts: LLM cuts data 4,250x","LLM-powered semantic link: images as captions, 4,250x smaller"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00157,"raw_usage":{"total_tokens":6253,"prompt_tokens":915,"completion_tokens":5338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":5234}},"tokens_in":531,"tokens_out":5338,"duration_ms":36393,"temperature":1.0,"reasoning_tokens":5234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:33:24.694340+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of images containing small task-critical details such as street signs, license plates, or fine-grained species, run the described LLaVA-to-Stable-Diffusion pipeline over the UMa channel at high SNR, and compare a downstream classifier's accuracy on the original images with its accuracy on the reconstructed ones. If accuracy drops substantially even though SSIM stays high, then the caption bottleneck does not preserve the semantic content that the claimed 4,250-fold compression is supposed to carry.","supporting_citations":[{"cited_title":"& Lee, Y","cited_arxiv_id":null,"evidence_quote":"Supplies LLaVA, the vision-language model used as the semantic encoder that maps images to caption text."},{"cited_title":"& Ommer, B","cited_arxiv_id":null,"evidence_quote":"Supplies Stable Diffusion, the semantic decoder that regenerates the image from text conditioning."},{"cited_title":"& Others Learning trans- ferable visual models from natural language supervision","cited_arxiv_id":null,"evidence_quote":"Supplies CLIP's text encoder, which turns the received prompt into a fixed-dimension semantic latent for generation."},{"cited_title":"& Abbeel, P","cited_arxiv_id":null,"evidence_quote":"Supplies the denoising diffusion probabilistic model mechanism behind the image generation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the 3GPP urban-macrocell channel model with mobility used as the physical propagation scenario."}],"review_version":1}