{"id":"a7738cf0-7d57-40da-8a00-422e608783f5","arxiv_id":"2507.17416","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Stable Cascade based semantic image communication system transmits a compact image embedding (0.29% of original size) over noisy channels and reconstructs images faster and with better perceptual quality than GESCO, Img2Img-SC, and JPEG2000+LDPC.","lead":"This paper builds an image communication system that sends a very small image summary and lets a diffusion model recreate the picture at the receiver, even on noisy wireless channels. It reports better reconstruction quality and much faster decoding than three existing baselines, while transmitting only 0.29% of the original image data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Same-SNR comparisons do not enforce equal transmit power or rate, so the claimed superiority over baselines is not yet established.","rationale":"The reader's weakest assumption and my independent read converge on the same issue: the central claims are comparative, and the comparison is not resource-matched. The paper provides no power-normalization rule connecting Eq. (5) to the different numbers of transmitted coefficients, and no bit-rate or channel-use accounting for JPEG2000+LDPC. The 0.29% figure is presented as compression but is only a dimensionality ratio; in a communication system, what matters is bits or channel uses under a power constraint. The arbitrary handling of failed JPEG2000+LDPC decodes is a secondary symptom of the same comparability problem rather than a separate flaw. The paper does have independent support worth crediting: the ablation in Fig. 8a shows fine-tuning improves noise robustness, Fig. 6 documents generalization behavior on DIV2K, and the inference-time measurements are concrete and reproducible in principle. These support the internal mechanics of the proposed system, but they do not establish the headline superiority over the baselines unless the benchmark is placed on equal resources. Because the reader already assigned a CONDITIONAL verdict with medium confidence, my analysis does not move the verdict; it reinforces the condition. The specific test above—recomputing Fig. 4 under equal total power and equal rate—would settle whether the concern actually lands.","tokens_in":9279,"tokens_out":5070,"duration_ms":60187,"concrete_test":"Recompute the Fig. 4 comparison under an equal-resource protocol: fix total transmit power P_total; for each analog method with D transmitted coefficients, add Gaussian noise with per-coefficient variance σ^2 = P_total / (D · 10^{SNR/10}); for JPEG2000+LDPC, sweep JPEG2000 bitrate and QAM/LDPC parameters so the total transmitted bits (including parity) equals the proposed method's channel uses, or report a full rate-distortion curve at each SNR. If the proposed model's curves do not dominate at equal power/rate, the abstract's superiority and 0.29%-overhead claims are unsupported. Also rerun the low-SNR averages without the arbitrary failed-decode score assignments, or present them as a separate sensitivity analysis.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the reported comparisons put all methods on equal communication terms. Eq. (5) defines SNR = 10 log(P/σ^2) with P called 'received signal power', but the paper never specifies whether P is total power across all transmitted symbols or per-symbol power, nor how the digital JPEG2000+LDPC signal is power-normalized. The proposed model sends 2304 continuously valued coefficients while Img2Img-SC sends 16384 and JPEG2000+LDPC sends an unreported number of coded QAM symbols; if P is total received power, per-symbol power differs by an order of magnitude, and if P is per-symbol power, total power differs. Thus 'SNR = 10 dB' does not represent equal resources. The headline compression number, 0.29%, is a ratio of tensor element counts (Table 1), not a bit rate or channel-use count; no bits-per-pixel, LDPC parity overhead, modulation order, or channel-use count is reported for JPEG2000+LDPC. Failed JPEG2000+LDPC decodes are assigned arbitrary PSNR/SSIM = 0 and LPIPS = 1 / FID = 500 values, which can dominate low-SNR averages. Since every headline claim is relative to these baselines, unequal power/rate normalization could change the ranking.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SC-SIC, a Stable Cascade-based semantic image communication system. A compact image embedding extracted by EfficientNet-V2 (e.g., [16,12,12] for 512x512 and [16,24,24] for 1024x1024, corresponding to 0.29% of the original tensor size) is transmitted over an AWGN channel, and a fine-tuned Stage B LDM denoises/reconstructs the latent given the corrupted embedding as conditioning, followed by the VQGAN decoder. Experiments on Cityscapes and DIV2K compare the method against GESCO, Img2Img-SC, and JPEG2000+LDPC using LPIPS, FID, SSIM, and PSNR, and include inference-time, predictability, and ablation analyses. The paper's central claims are that this system reconstructs images more faithfully under noisy channels, is computationally faster (3x at 512x512, 16x at 1024x1024 vs Img2Img-SC), and transmits a much smaller conditioning signal.","tokens_in":9538,"tokens_out":5622,"duration_ms":50201,"significance":"If the central claims hold, the paper makes a useful contribution to DM-based semantic image communication: it demonstrates that a complex multi-stage generative architecture can be repurposed for SIC with a very compact learned embedding as conditioning, and it provides ablation evidence that fine-tuning on noisy conditioning is what confers channel robustness. The generalization test on DIV2K is a particularly honest stress test, and the explicit reporting of generation randomness via repeated-sampling LPIPS statistics is a strength. The reliance on external pretrained components (VQGAN, EfficientNet-V2, stage B checkpoint) is clearly acknowledged, and the promise of released code and fine-tuned weights supports reproducibility if honored. The main reservations concern the fairness and interpretability of the benchmark comparisons, not the soundness of the core training recipe.","major_comments":[{"comment":"The SNR definition in Eq. (5) uses \"received signal power\" P, but the paper does not state whether P is total power across all transmitted symbols or per-symbol power, nor how the JPEG2000+LDPC signal is power-normalized relative to the analog embeddings. Since the proposed method transmits 2304 analog coefficients for 512x512 images, Img2Img-SC transmits 16384, and JPEG2000+LDPC transmits an unreported number of coded QAM symbols, equal SNR values do not necessarily mean equal transmit power or equal bandwidth; because all headline claims are relative to these baselines, the reported comparisons do not yet establish the claimed superiority.","section":"§3.2, Eq. (5), §4.2"},{"comment":"The number of diffusion sampling steps used by the proposed model is never reported, while the text states that GESCO and Img2Img-SC use 1000 and 30 steps, respectively. The inference-time comparison (0.78 vs 2.53 vs 29.19 s) is only meaningful if step counts are comparable or explicitly held constant; as written, the claimed 3x/16x speed advantage could be an artifact of a smaller number of denoising steps.","section":"§4.2, Figure 5"},{"comment":"For failed JPEG2000+LDPC decodes, the paper assigns PSNR=0, SSIM=0, LPIPS=1, and FID=500 before averaging. These arbitrary values can dominate the low-SNR aggregates and inflate the apparent gap between the proposed method and the digital baseline; the authors should exclude failed transmissions and report decode success rates, or otherwise justify the imputation.","section":"§4.3.1, Figure 4"},{"comment":"The \"0.29% of original image size\" figure is a ratio of tensor element counts, not a bit-rate or channel-use comparison. For the JPEG2000+LDPC baseline, the number of channel uses depends on quantization, LDPC parity overhead, and QAM modulation order, none of which are reported, so the bandwidth-efficiency claim is not a rate-matched comparison and cannot be directly compared with the analog transmission of continuous embeddings.","section":"Table 1, §4.3.2"},{"comment":"GESCO is evaluated at 256x512 while all other methods are evaluated at 512x512, and the paper does not discuss the effect of this resolution mismatch on PSNR, SSIM, LPIPS, or FID. Since these metrics are not resolution-invariant, the quality comparison against GESCO in Figure 4 and Table 2 is not strictly apples-to-apples.","section":"§4.2"}],"minor_comments":[{"comment":"Eq. (3) uses inconsistent notation: the left-hand side and the right-hand side both use XVG,t, but the forward process should read XVG,t = sqrt(αbar_t) * XVG + sqrt(1-αbar_t) * epsilon; please fix.","section":"Eq. (3)"},{"comment":"Table 1 lists DIFFSC and CASC, but these two systems are not compared in the experiments; either add the comparisons or clarify why they are listed only for dimensionality.","section":"Table 1"},{"comment":"The horizontal axis in Figure 4 starts at 5 dB, so the results at SNR=1 dB, which are described in the text, are not visible in the plots.","section":"Figure 4"},{"comment":"The statement that increasing the embedding size improves LPIPS, FID, and SSIM by greater than 10% should also report the PSNR change or explain why it is omitted.","section":"§4.3.5"},{"comment":"The manuscript says code and weights \"will be accessible\" at a GitHub URL; please make the link live and include a short reproducibility statement describing the exact sampling configuration (including step count) used for the proposed model.","section":"§4.1, §4.2"},{"comment":"Minor typographical issues include the duplicated \"IEEE\" in the DiffSC reference and the use of \"V AEs\" in the introduction; please proofread.","section":"References, Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a workshop paper, but the fairness of the benchmark is the main obstacle; the same-SNR issue is common in this literature, but here it is load-bearing. The missing step count for the proposed model is easy to fix and should be included in revision. I would be inclined to accept once the comparisons are normalized or explicitly qualified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a believable engineering result that transplants Stable Cascade into SIC, fine-tuning stage B on noisy image embeddings. The ablation showing the original SC fails without fine-tuning is the strongest piece of evidence. But the paper never states its own sampling steps, and the SNR comparison is not clearly power/rate-normalized, so the headline 3x/16x speedups and quality gains over Img2Img-SC and JPEG2000+LDPC are not yet pinned down.\n\nWhat's genuinely new: applying the Würstchen/Stable Cascade two-stage setup to SIC and conditioning stage B on a compact EfficientNet-V2 embedding rather than text. That changes the game relative to text-conditioned diffusion SIC systems like Img2Img-SC because it directly attacks generation randomness. The predictability table supports that: variance 0.003 at 20 dB is real. The paper also does the right thing with ablations: fine-tuning vs. original and embedding size [16,24,24] vs [16,32,32]. And it discloses the GESCO 256x512 limitation and the arbitrary assignment of failed JP2+LDPC decodes, which is more honest than most.\n\nSoft spots, in order: (1) sampling steps for the proposed model are never given. They only report wall-clock times, so the 3x/16x speedup could just be from using 4 steps versus 30. This is a one-line fix and should be in the paper. (2) The SNR definition in Eq. (5) is ambiguous between total and per-symbol power, and no power normalization is specified for JPEG2000+LDPC. With 2304 vs 16384 coefficients on the other side, \"SNR = 10 dB\" may not mean equal transmit power or equal bandwidth. The compression ratio in Table 1 is also tensor-element count, not bits or channel uses. For a fair comparison, they need at least a rate-matched or power-matched protocol. (3) Failed JP2 decodes assigned PSNR=0/SSIM=0/LPIPS=1/FID=500 will dominate low-SNR averages; better to exclude failures and report success rates separately. (4) GESCO at 256x512 is disclosed but still makes the visual comparison hard to interpret. None of these are fatal to the core idea, but they all matter for the relative claims.\n\nBottom line: this is a workshop-length empirical paper with a plausible core and honest disclosure. It deserves a serious referee, not a desk reject, and I'd send it to a venue that will hold the authors to a fairer benchmark. If the code and sampling steps materialize, the result is likely to hold qualitatively. For my own work, I wouldn't cite it until the evaluation is cleaned up.","headline":"Plausible Stable-Cascade-based SIC with an honest ablation, but the missing sampling-step count and unnormalized SNR comparison make the headline speedups and quality gains unproven.","tokens_in":10072,"tokens_out":4939,"would_cite":false,"duration_ms":52404,"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":"An image communication system sending just 0.29% of the image as a latent embedding reconstructs it with a fine-tuned diffusion stage, beating three baselines under channel noise.","keywords":["semantic image communication","diffusion models","Stable Cascade","latent diffusion","AWGN channel","image compression","generative communication","noise robustness"],"falsifier":"Re-run the comparison with a fixed total bit budget: encode the source image with JPEG2000 at the same number of bits as the [16,12,12] embedding plus its required channel coding, transmit both over the same AWGN channel with equal per-symbol energy, and measure LPIPS and FID; if JPEG2000+LDPC or a coded version of the embedding matches or beats the proposed model on perceptual metrics, the claimed superiority under fair rate matching is refuted.","tokens_in":9097,"feed_emoji":"🖼️","tokens_out":12111,"duration_ms":104339,"temperature":0.7,"pith_summary":"This paper argues that semantic image communication can be simultaneously tiny, fast, and noise-robust: it transmits a latent image embedding of only 0.29% of the original image size and reconstructs the image with a diffusion model conditioned on that embedding. The system, built from Stable Cascade's VQGAN stage and a fine-tuned latent diffusion stage, is claimed to beat three benchmarks — a segmentation-conditioned diffusion communicator, a Stable Diffusion-based communicator, and JPEG2000 with LDPC coding — in perceptual reconstruction quality over an AWGN channel, with the largest margins at low SNR. It also claims large speedups (over 3x at 512x512 and over 16x at 1024x1024 versus the Stable Diffusion baseline) and low generation randomness, with the LPIPS standard deviation of repeated reconstructions at 0.003 for SNR of 10 dB and above. If these results hold, diffusion-based semantic image communication would shed its two practical blockers: slow inference and unreliable generation.","feed_headline":"0.29% image embedding beats diffusion and JPEG baselines in noise","feed_subtitle":"Stable Cascade conditioning latent rebuilds images in 0.78 seconds, even at 1 dB SNR.","key_machinery":"The machinery is a compact image embedding $Z$ from the EfficientNet-V2 encoder, used as conditioning for a fine-tuned latent diffusion stage, with a VQGAN (a vector-quantized generative adversarial network) decoder reconstructing the image. Stage A is the VQGAN that maps the $1024\\times1024$ image into latent space; stage B is a latent diffusion model trained with the objective $L = \\mathbb{E}[\\|\\epsilon - \\bar\\epsilon(X_{\\mathrm{VG},t}, t, \\hat Z)\\|_2^2]$, where $\\hat Z = Z + \\epsilon$ is the AWGN-corrupted embedding. Transmitting $\\hat Z$ instead of starting diffusion from a noisy image or from text makes channel noise enter the system as a degraded conditioning signal, and fine-tuning on SNR values between 1 and 20 dB teaches the denoiser to compensate. This mechanism is what lets a [16,12,12] tensor (0.29% of a 512x512 RGB image) produce reconstructions that the paper reports as perceptually superior to the baselines.","core_discovery":"The paper's central claim is that a highly compressed image embedding, extracted by an EfficientNet-V2 encoder and transmitted alone over an AWGN channel, carries enough semantic information for a fine-tuned latent diffusion model to reconstruct the original image faithfully. At the receiver, the noisy embedding is used purely as conditioning for stage B of Stable Cascade; the diffusion process predicts noise and produces a VQGAN latent that the stage A decoder turns into pixels. The fine-tuning steps make the denoiser robust to channel impairments, and the paper reports that this design outperforms GESCO, Img2Img-SC, and JPEG2000+LDPC on LPIPS, FID, SSIM, and PSNR across SNR from 1 to 20 dB. On Cityscapes, the average gains over the next-best diffusion baseline are 43% in FID and 55% in LPIPS, with SSIM up 56% and PSNR up 23%; on unseen DIV2K data the model degrades but still produces semantically coherent reconstructions. The paper also reports reconstruction from a [16,12,12] embedding for 512x512 images (compression ratio 341) in 0.78 seconds and from a [16,24,24] embedding for 1024x1024 images in 1.72 seconds, and its ablation shows that the fine-tuning, not the pretrained Stable Cascade model, is what provides the noise robustness.","pith_inferences":["A natural test is to move beyond AWGN to fading or quantized channels; the conditioning-denoiser mechanism would need retraining or adaptation, since only additive white Gaussian noise is evaluated.","The advertised 0.29% figure counts raw tensor dimensionality; a fully rate-matched comparison that includes the embedding's quantization, power normalization, and any digital protection bits could shrink or widen the gap with JPEG2000+LDPC.","The receiver never sees the clean embedding, so the scheme is effectively a learned joint source-channel code; adding a lightweight confidence or feedback signal to detect overly corrupted conditioning is an unexplored direction.","The same conditioning idea could transfer to video by transmitting compact per-frame embeddings, though the paper only mentions video as future work and does not test temporal consistency."],"forward_implications":["If the reported gains hold, diffusion-based semantic communication can send only a tiny conditioning embedding, making the channel-use cost of image delivery negligible compared with pixel, latent-map, or text-plus-image schemes.","The 0.78-second reconstruction for 512x512 images and 1.72 seconds for 1024x1024 on a single GPU bring diffusion-based image delivery close to interactive use, because the 3x and 16x speedups come from working in Stable Cascade's smaller latent space rather than Stable Diffusion's.","The low generation randomness (LPIPS standard deviation 0.003 at SNR 10 dB and above) means repeated transmission of the same image yields nearly identical reconstructions, which is what a communication system needs for predictable behavior.","Fine-tuning on Cityscapes alone transfers to unseen DIV2K images with semantic coherence if not exact color fidelity, suggesting that one diverse training set could yield a general-purpose semantic image communicator.","The embedding-size ablation shows a tunable bandwidth-fidelity tradeoff: increasing the embedding from [16,24,24] to [16,32,32] improves perceptual scores by more than 10% but lowers the compression ratio from 341 to 192."],"supporting_citations":[{"why":"Supplies the Stable Cascade architecture, including the pretrained VQGAN stage A, the EfficientNet-V2 encoder, and the stage B latent diffusion model that the paper fine-tunes.","marker":"(Pernias et al., 2023)"},{"why":"Defines the Img2Img-SC baseline whose 30-step Stable Diffusion sampling and text-plus-image conditioning is the main speed and quality comparison.","marker":"(Cicchetti et al., 2024)"},{"why":"Defines the GESCO baseline, a segmentation-map-conditioned diffusion communicator, used for low-SNR quality comparison.","marker":"(Grassucci et al., 2023)"},{"why":"Provides the conventional deep joint source-channel coding setting, including the JPEG2000+LDPC benchmark and the QAM/LDPC transmission setup.","marker":"(Bourtsoulatze et al., 2019)"},{"why":"Provides the Cityscapes dataset used for fine-tuning the diffusion stage and for all trained evaluations.","marker":"(Cordts et al., 2016)"},{"why":"Provides the DIV2K dataset used to test zero-shot generalization on unseen image types.","marker":"(Agustsson & Timofte, 2017)"},{"why":"Introduces VQGAN, the stage A compressor that maps the image into the latent space the diffusion stage denoises.","marker":"(Esser et al., 2021)"},{"why":"Introduces EfficientNet-V2, the encoder that extracts the compact transmitted embedding.","marker":"(Tan & Le, 2019)"}],"fun_headline_variants":["0.29% latent beats diffusion baselines in noisy channels","Stable Cascade sends 0.29% of image, beats JPEG and diffusion","Tiny embedding, robust reconstruction: 0.29% size wins in noise","3x faster decode: 0.29% image prior outperforms diffusion baselines","Noise-robust SIC: 0.29% embedding, 16x faster at 1024px"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline comparison assumes that ranking all methods at the same received-signal-to-noise ratio, without equalizing transmit power or counting the extra bits used by JPEG2000 and LDPC coding, is a fair way to compare reconstruction quality at equal bandwidth; if the comparison gave every method the same data budget or transmit power, the reported margins over the baselines could change.","fun_headline_variants_meta":{"raw":{"variants":["0.29% latent beats diffusion baselines in noisy channels","Stable Cascade sends 0.29% of image, beats JPEG and diffusion","Tiny embedding, robust reconstruction: 0.29% size wins in noise","3x faster decode: 0.29% image prior outperforms diffusion baselines","Noise-robust SIC: 0.29% embedding, 16x faster at 1024px"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000269,"raw_usage":{"total_tokens":1660,"prompt_tokens":1021,"completion_tokens":639,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":528}},"tokens_in":637,"tokens_out":639,"duration_ms":5672,"temperature":1.0,"reasoning_tokens":528,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:48:26.025526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison with a fixed total bit budget: encode the source image with JPEG2000 at the same number of bits as the [16,12,12] embedding plus its required channel coding, transmit both over the same AWGN channel with equal per-symbol energy, and measure LPIPS and FID; if JPEG2000+LDPC or a coded version of the embedding matches or beats the proposed model on perceptual metrics, the claimed superiority under fair rate matching is refuted.","supporting_citations":[{"cited_title":"Language-oriented semantic latent representation for image transmission","cited_arxiv_id":null,"evidence_quote":"Defines the Img2Img-SC baseline whose 30-step Stable Diffusion sampling and text-plus-image conditioning is the main speed and quality comparison."},{"cited_title":"B., and G \\\"u nd \\\"u z, D","cited_arxiv_id":null,"evidence_quote":"Provides the conventional deep joint source-channel coding setting, including the JPEG2000+LDPC benchmark and the QAM/LDPC transmission setup."},{"cited_title":"The cityscapes dataset for semantic urban scene understanding","cited_arxiv_id":null,"evidence_quote":"Provides the Cityscapes dataset used for fine-tuning the diffusion stage and for all trained evaluations."},{"cited_title":"and Timofte, R","cited_arxiv_id":null,"evidence_quote":"Provides the DIV2K dataset used to test zero-shot generalization on unseen image types."}],"review_version":1}