{"id":"00a9b6db-ed7c-4dc1-a981-2e5eb05c5a7d","arxiv_id":"2505.10405","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper introduces GVIF, a mutual-information-based metric for visual fidelity in generative semantic communication, and uses it to adaptively select critical image features and compression levels for wireless transmission.","lead":"This paper proposes a new visual fidelity metric, GVIF, for generative semantic communication, and uses it to adaptively select which image features to transmit and how much to compress them. It also combines text prompts with semantically critical feature regions to let a diffusion model reconstruct images with higher fidelity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GVIF's core Eq. (27) rests on the scaling assumption in Eq. (21) that is not established and is internally inconsistent with the training procedure.","rationale":"The Reader's weakest_assumption identifies Eq. (21) as the fragile step, and my reading agrees: the scaling assumption is the load-bearing step in the derivation of GVIF, and it is unsupported and likely inconsistent with the independently trained encoders. The paper's Appendix A only validates the diffusion-generation independence assumption (Eqs. 23–24), not the distortion-model scaling. I therefore agree with the Reader's verdict of CONDITIONAL: the empirical system is plausible and the metric's monotonicity properties are definitional, but the theoretical grounding of GVIF is not established. The recommended adjustment is to require either a derivation of Eq. (21) from the hyperprior training or an empirical validation of the shared-latent assumption; absent that, the central claim that GVIF measures the mutual information between distorted and reference features is not supported. I am not recommending REJECT because the system itself appears functional and the metric could be re-interpreted as a heuristic, but the claim as stated is conditional on an unvalidated model.","tokens_in":21327,"tokens_out":2041,"duration_ms":17290,"concrete_test":"Compute the elementwise Pearson correlation between y^c and y^r on a held-out set of images, using the actual encoders Φc and Φr from the experimental setup, after aligning spatial locations and channels. Under Eq. (21), the correlation should be near 1 (up to the shared u_ijc); if the average absolute correlation is well below 0.9, the scaling model fails. A second, independent check: re-derive Eq. (43) without invoking Eq. (21), instead treating y^c and y^r as jointly Gaussian with a general covariance; if the resulting mutual information is not equal to the log-term in Eq. (43), the closed form is invalid outside the shared-latent model.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that GVIF is an accurate per-sample visual fidelity metric for hybrid Gen-SemCom hinges on the distortion model in Eq. (21): y^c = β_ijc · y^r, with the same latent realization u_ijc shared between reference coder Φr and lossy coder Φc. This is the load-bearing step because Proposition 4.1 (Eq. 27) is just a closed-form evaluation of mutual information under that model; if Eq. (21) fails, GVIF is not the mutual information it claims to be. The paper provides no derivation of Eq. (21) — scaling of variances does not imply elementwise scaling of latent realizations — and Appendix A validates only the generation-model assumption in Eqs. (23)–(24), not Eq. (21). The scaling parameter in Eq. (30), β̂_ijc = θ̄^c_ijc / θ̄^r_ijc, is a ratio of estimated variances, which is only consistent with Eq. (21) if the same u_ijc appears in both features. But the hyperprior encoders are trained separately via rate–distortion optimization (Eq. 13), so there is no mechanism forcing the latent draws to coincide; different NN weights give different encoder functions. Moreover, the paper's own denominator uses the reference coder Φr at a 40 dB PSNR while numerators use a lossy coder Φc, and no experiment measures the actual correlation between y^c and y^r. The AGX/cleanup claim is also not independently supported: the only support would be a per-image correlation analysis between y^c and y^r, which is not reported. Additionally, the metric's second stated property — that GVIF captures 'critical information' because high-variance features dominate the log-sum — is essentially definitional, and the experimental validation (Figs. 7–8) shows monotonicity with α and mask PSNR, but that is a property of the formula, not evidence that the model in Eq. (21) holds. The experimental section lacks error bars, code, and a human perceptual study, and the FID-vs-GVIF monotonicity in Fig. 11(b) is also a property of the formula once β and P are fixed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a hybrid generative semantic communication system in which a text prompt and a subset of semantically critical VAE features are transmitted, with the receiver using a diffusion-based inpainting model for reconstruction. The main technical contributions are a CAM-based semantic filtering mechanism controlled by a threshold alpha, and a proposed metric, GVIF, defined in Eq. (27) as the ratio of conditional mutual information between distorted and reference image features under a Gaussian scale mixture model and an HVS Gaussian channel model. The authors also formulate a channel-adaptive optimization problem that jointly selects the filtering threshold and the VAE source coder by maximizing the average GVIF under a latency constraint. Experiments on ImageNet report that GVIF increases with mask PSNR and transmitted critical information volume, and that the optimized system achieves higher mask PSNR and lower FID than the JPEG2000 and VAE-based benchmarks.","tokens_in":21651,"tokens_out":8269,"duration_ms":84388,"significance":"If the proposed metric is valid, the paper would provide a principled, information-theoretic per-sample fidelity measure for hybrid Gen-SemCom, filling a real gap since standard metrics such as PSNR and FID do not directly relate visual quality to the volume of semantically critical transmitted information. The system-level contribution, namely channel-adaptive joint optimization of semantic filtering and source coding, is plausible and well motivated, and the experimental comparison against JPEG2000, VAE-only, and prompt-only generation is useful. The paper also includes derivations in Appendix B and some empirical validation in Appendix A, which are strengths. However, the mutual-information interpretation of GVIF rests on an unvalidated scaling assumption between separately trained VAE encoders, and the computation procedure never actually measures features of the generated image; these issues directly affect the central claim and must be resolved before the metric can be accepted as stated.","major_comments":[{"comment":"The scaling model y^c_ijc = β_ijc y^r_ijc is introduced without derivation, and it is load-bearing for the GVIF claim: Proposition 4.1 evaluates conditional mutual information only because the same latent realization u_ijc is assumed to appear in both y^c and y^r. This assumption does not follow from the GSM model, since the reference and lossy encoders are separately trained neural networks and no mechanism forces their features to share the same underlying random variable u; scaling of variances alone, which motivates Eq. (30), does not imply elementwise scaling of realizations. The paper validates Eqs. (23)–(24) in Appendix A, but provides no validation of Eq. (21). A direct per-element correlation or cosine-similarity analysis between y^c and y^r across the coder family is needed to support the mutual-information interpretation; without it, the closed form in Eq. (27) is not established as the mutual information it claims to be.","section":"§IV-A2, Eq. (21)"},{"comment":"The computation procedure for GVIF never evaluates features of the generated image ~x: θ̄^r and θ̄^c are computed from the reference and lossy source encoders, β̄ is the ratio of their standard deviations, and the metric is assembled from these quantities together with the set P. The actual diffusion output enters only through the assumed independence in Eq. (24). Consequently GVIF is a predicted fidelity based on the source coder and the filter, not a measurement of the image that the receiver actually produced. The Appendix A validation of Eqs. (23)–(24) does not close this gap, because that validation concerns y^g while the computation in Eqs. (28)–(30) bypasses y^g entirely. The paper should either compute GVIF from the generated features or explicitly re-frame GVIF as a model-based prediction and validate that prediction against human or established perceptual scores.","section":"§IV-B2, Eqs. (28)–(30)"},{"comment":"The independence of y^g_ijc outside the transmitted set P is validated only through Pearson correlation coefficients in Fig. 12. Zero correlation does not imply zero mutual information, and Eq. (43) requires I(g^d_ijc; y^r_ijc | θ̄^r, β̄) = 0 for every out-of-P element; the proof collapses if the generated features are merely uncorrelated but dependent. A conditional-Gaussian argument or a direct mutual-information estimate is needed. In addition, the appendix writes y = F(x; Φ_D), which appears to be a typo for the encoder parameter Φ_E, and should be corrected.","section":"Appendix A, Eq. (24)"},{"comment":"The text presents the monotonicity of GVIF with respect to the transmitted set P and the scaling β̄ as a validated property, but this behavior is built into Eq. (27): enlarging P adds positive log terms, and β̄ enters the numerator multiplicatively. The correlation between GVIF and mask PSNR reported in Fig. 7 is therefore partly expected, since both are driven by the same source-coder parameters and transmitted feature count. The independent evidence in Fig. 11(b), namely FID decreasing with GVIF, is valuable and should be emphasized, but the paper should avoid claiming that the monotonicity experiments alone validate the metric's perceptual sensitivity; a comparison against human ratings or established full-reference metrics would strengthen the claim.","section":"§IV-B and §VI-B"}],"minor_comments":[{"comment":"The index in the universal quantifier is written as y∈[Hy], which should presumably be j∈[Hy]; this typo appears in the definition of the GSM model.","section":"Eq. (19)"},{"comment":"The reference coder is defined as Φr = {Φr_E, Φr_D, Ωc_1, Ωr_2}, where the third element appears to be a typo and should be Ωr_1; please check the notation consistency.","section":"§IV-B2"},{"comment":"The symbol B is used both for the total number of encoded bits and for the channel bandwidth in the latency expression; this overloading may confuse readers and should be disambiguated.","section":"Eq. (8)"},{"comment":"The text says 'FID score is a monotonously decreasing function of the GVIF', but the plot appears to show a decreasing relationship; consider using a more precise term such as 'monotonically decreasing' and clarifying whether the claim is empirical rather than theoretical.","section":"Fig. 11"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the system-level results are interesting, but the information-theoretic interpretation of GVIF currently rests on an unvalidated assumption about shared latent realizations between separate VAE encoders, and the metric is computed without reference to the actual generated image. I would ask the authors to supply a direct per-element correlation analysis between y^c and y^r, and to replace the Pearson-correlation validation in Appendix A with a mutual-information-based check. If those analyses do not support Eq. (21), the metric should be re-framed as a heuristic proxy rather than as a conditional mutual information measure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the system half—hybrid Gen-SemCom with CAM-based filtering and diffusion inpainting—is a reasonable, clearly explained contribution, and the experiments support it as a useful way to protect semantically critical regions. Second, the metric half, GVIF, has a load-bearing assumption that is not established. Eq. (21) posits that features from any lossy coder are a scalar multiple of reference features with the same Gaussian scale mixture realization. That is not derived, and it is inconsistent with the training procedure: the source coders are trained separately via rate-distortion optimization, so nothing forces the latent realizations to coincide. The paper computes beta as a ratio of standard deviations; that only matches the elementwise scaling if the same u_ijc appears in both, which the paper never verifies. The correlation measurements in Appendix A validate the generation model (Eqs. 23–24), not Eq. (21). Because Proposition 4.1 is simply the closed-form mutual information under that model, GVIF loses its information-theoretic meaning if the scaling assumption fails.\n\nWhat's good? The paper is honest enough to include Appendix A, the writing is clear, and the experimental setup is plausible. The observed monotonicity of GVIF with mask PSNR and threshold alpha is consistent with a useful heuristic. The CAM-based selection of critical features and the channel-adaptive optimization are the parts I'd want to build on. The references are appropriate; the debt to VIF, the hyperprior, and Repaint is explicit.\n\nThe soft spots beyond Eq. (21): the two stated 'properties' of GVIF are definitional, not empirical discoveries. The FID-vs-GVIF monotonicity in Fig. 11(b) is also largely formulaic given the model. The experiments have no error bars, no code release, and no human perceptual study, which are the standard evidence for a fidelity metric.\n\nWho benefits: readers working on semantic communications and generative image transmission, especially those designing critical-information embedding schemes. They should treat GVIF as a proposed heuristic until Eq. (21) is either justified with real latent-correlation measurements or replaced by a model the paper can defend.\n\nRecommendation: yes, send it to peer review, but with a referee who specifically asks for that evidence. The system contribution is strong enough to warrant revision, and the metric question is worth a careful referee rather than a desk reject.","headline":"Strong systems story, but GVIF's key scaling assumption (Eq. 21) is unproven and likely false as stated; send it to review and make the authors defend or redraw the metric.","tokens_in":22264,"tokens_out":4354,"would_cite":true,"duration_ms":39138,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper defines GVIF, a per-image fidelity measure based on mutual information between received and reference features, and shows that maximizing it adapts semantic filtering and source coding to the channel, beating benchmark schemes.","keywords":["generative semantic communication","visual fidelity metric","mutual information","semantic filtering","class activation mapping","variational autoencoder","diffusion model","channel-adaptive transmission"],"falsifier":"Compute, over many images, the residual $y^c_{ijc}-(\\bar\\theta^c_{ijc}/\\bar\\theta^r_{ijc})y^r_{ijc}$ between the features produced by a lossy VAE coder and those from the reference coder; if the residual variance is comparable to or larger than the visual-noise variance $\\gamma^2$, the scalar-scaling assumption behind Eq. (21) fails and the GVIF numerator in Eq. (43) overstates the shared information. A second check is to regenerate the same transmitted mask and prompt with different diffusion noise samples; if GVIF varies substantially across runs, the single-sample ergodicity assumption used to compute the mutual information is not valid.","tokens_in":21103,"feed_emoji":"📡","tokens_out":11205,"duration_ms":103244,"temperature":0.7,"pith_summary":"The paper proposes a hybrid generative semantic communication system that sends a short text prompt plus only the image features most relevant to the semantic label, then reconstructs the image with a diffusion model at the receiver. To decide which features are worth sending and how much to compress them, it introduces the generative visual information fidelity (GVIF) index, a per-image number between 0 and 1 that measures how much of the original image's feature information survives transmission. GVIF is the ratio of the mutual information between the received features and the reference features to the mutual information of the reference features with itself, both measured under a Gaussian-channel model of human visual processing. The paper shows that maximizing GVIF over the filtering threshold and the source coder yields a channel-adaptive system with higher PSNR on critical regions and lower FID than benchmark schemes. The reason to care is that existing metrics either assess whole distributions, like FID, or ignore semantic content, like PSNR, while GVIF ties visual fidelity directly to the amount of semantically critical information actually transmitted.","feed_headline":"GVIF: one number tracks image fidelity in wireless AI links","feed_subtitle":"The score measures how much semantic-critical detail survives compression, letting each transmission adapt to the channel.","key_machinery":"The load-bearing object is the GVIF identity of Eq. (27), which combines four statistical models into a single ratio. Reference image features follow a Gaussian scale mixture $y^r_{ijc}=\\theta^r_{ijc}u_{ijc}$ with $u_{ijc}\\sim\\mathcal{N}(0,1)$, coming from the VAE hyperprior; any lossy source coder is assumed to scale these features by $\\beta_{ijc}$; semantic filtering sets unselected features to zero; and the diffusion decoder preserves selected features while generating unselected ones independently. The human visual system is modeled as an additive Gaussian channel with noise variance $\\gamma^2$, so the conditional mutual information between distorted and reference features has a closed form. GVIF is then used as the optimization objective: a zero-order gradient method tunes the filtering threshold $\\alpha$ for each fixed coder, and the best pre-trained coder is selected under a latency constraint.","core_discovery":"The central claim is that the visual fidelity of a single generated image can be quantified by $$V(\\bar{\\boldsymbol{\\$\\beta$}},\\mathcal{P};x)=\\frac{\\sum_{(i,j,c)\\in\\mathcal{P}}\\log_2\\!\\left(1+\\frac{(\\bar{\\$\\beta$}_{ijc}\\bar{\\$\\theta$}^r_{ijc})^2}{\\$gamma^{2}$}\\right)}{\\sum_{(i,j,c)\\in\\mathcal{U}}\\log_2\\!\\left(1+\\frac{(\\bar{\\$\\theta$}^r_{ijc})^2}{\\$gamma^{2}$}\\right)},$$ where $\\bar{\\theta}^r_{ijc}$ are the feature scales extracted by a high-quality reference VAE coder, $\\bar{\\beta}_{ijc}$ are per-feature scaling factors from the lossy coder, $\\mathcal{P}$ is the set of transmitted feature positions, $\\mathcal{U}$ is the full feature grid, and $\\gamma^2$ is the variance of the additive Gaussian noise used to model the human visual system. The ratio lies in $[0,1]$, equals one for an unaltered image, and grows when the selected feature set contains high-variance, semantically important positions. The paper further claims that using this ratio as the objective for choosing the filtering threshold $\\alpha$ and the source coder $\\Phi$ produces a channel-adaptive hybrid system that beats a classic image codec and a VAE-only scheme in mask PSNR and FID, and that GVIF correlates with both mask PSNR and the volume of transmitted critical information.","pith_inferences":["A natural extension is to turn GVIF into a differentiable training loss for the source encoder and the filtering mask, since the index is built from encoder-controlled feature statistics; the paper itself only uses it to select among pre-trained coders and a scalar threshold.","The scalar-scaling distortion model suggests a testable generalization: replacing $\\beta_{ijc}$ with a learned linear transform between coder feature spaces would keep the index meaningful for encoders whose latent spaces do not align, without retraining a reference coder.","The single-sample ergodicity assumption implies that GVIF should be stable when the same transmitted features are decoded with different diffusion noise samples; measuring that stability would let GVIF be reported with confidence intervals.","Reading the numerator as preserved 'visual bits' suggests a rate-perceptual-information formulation, connecting GVIF to rate-distortion theory in a way the paper does not spell out."],"forward_implications":["GVIF provides a per-sample fidelity score that increases with both the number of transmitted critical features and the quality of their compression, so it can serve as the objective in other hybrid semantic communication designs.","At low receive SNR the optimized threshold rises to protect only the semantically important features, avoiding the sharp performance cliff that full-feature VAE transmission exhibits below roughly 10 dB.","Because GVIF is computed from encoder-side feature statistics rather than from the final pixels, it can be evaluated before the diffusion decoder runs, allowing online adaptation of transmission parameters.","The experimental trend that FID falls sharply once a small number of critical features are transmitted supports the practical claim that most visual information is concentrated in semantically important regions."],"supporting_citations":[{"why":"supplies the reference-distortion mutual-information framework and the Gaussian-channel model of the human visual system that GVIF adapts to the feature domain.","marker":"[34]"},{"why":"provides the ratio-of-mutual-informations form and the natural-scene-statistics basis for the fidelity index.","marker":"[44]"},{"why":"provides the VAE hyperprior source coder whose Gaussian scale mixture feature model underlies both reference and distortion models.","marker":"[10]"},{"why":"supplies the latent diffusion model used at the receiver for image generation conditioned on prompts and critical pixels.","marker":"[26]"},{"why":"supplies the conditional inpainting procedure that anchors transmitted critical pixels during generation.","marker":"[40]"},{"why":"provides class activation mapping, the mechanism for computing the spatial importance matrix used in semantic filtering.","marker":"[33]"},{"why":"supplies the zero-order gradient estimation method used to optimize the filtering threshold.","marker":"[46]"},{"why":"is the classic image codec baseline against which the proposed system's mask PSNR is compared.","marker":"[50]"}],"fun_headline_variants":["New metric gauges visual fidelity in generative wireless links","GVIF metric drives channel-adaptive compression for sharper images","One score tunes generative semantic comms to channel quality","Fidelity index for generative semantic comms: critical info matters","GVIF: adaptive fidelity metric for AI image transmission"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The index assumes that lower-quality encoding only scales down each feature of the reference image by a fixed factor, leaving the underlying content unchanged; if encoding changes the content that a feature represents, the whole fidelity calculation is measuring the wrong thing.","fun_headline_variants_meta":{"raw":{"variants":["New metric gauges visual fidelity in generative wireless links","GVIF metric drives channel-adaptive compression for sharper images","One score tunes generative semantic comms to channel quality","Fidelity index for generative semantic comms: critical info matters","GVIF: adaptive fidelity metric for AI image transmission"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000773,"raw_usage":{"total_tokens":3511,"prompt_tokens":1122,"completion_tokens":2389,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":738,"completion_tokens_details":{"reasoning_tokens":2310}},"tokens_in":738,"tokens_out":2389,"duration_ms":16490,"temperature":1.0,"reasoning_tokens":2310,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:08:57.846011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, over many images, the residual $y^c_{ijc}-(\\bar\\theta^c_{ijc}/\\bar\\theta^r_{ijc})y^r_{ijc}$ between the features produced by a lossy VAE coder and those from the reference coder; if the residual variance is comparable to or larger than the visual-noise variance $\\gamma^2$, the scalar-scaling assumption behind Eq. (21) fails and the GVIF numerator in Eq. (43) overstates the shared information. A second check is to regenerate the same transmitted mask and prompt with different diffusion noise samples; if GVIF varies substantially across runs, the single-sample ergodicity assumption used to compute the mutual information is not valid.","supporting_citations":[{"cited_title":"Image information and visual quality,","cited_arxiv_id":null,"evidence_quote":"supplies the reference-distortion mutual-information framework and the Gaussian-channel model of the human visual system that GVIF adapts to the feature domain."},{"cited_title":"An information fidelity criterion for image quality assessment using natural scene statistics,","cited_arxiv_id":null,"evidence_quote":"provides the ratio-of-mutual-informations form and the natural-scene-statistics basis for the fidelity index."},{"cited_title":"Variational image compression with a scale hyperprior,","cited_arxiv_id":null,"evidence_quote":"provides the VAE hyperprior source coder whose Gaussian scale mixture feature model underlies both reference and distortion models."},{"cited_title":"High-resolution image synthesis with latent diffusion models,","cited_arxiv_id":null,"evidence_quote":"supplies the latent diffusion model used at the receiver for image generation conditioned on prompts and critical pixels."},{"cited_title":"Repaint: Inpainting using denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"supplies the conditional inpainting procedure that anchors transmitted critical pixels during generation."},{"cited_title":"Learning deep features for discriminative localization,","cited_arxiv_id":null,"evidence_quote":"provides class activation mapping, the mechanism for computing the spatial importance matrix used in semantic filtering."},{"cited_title":"A primer on zeroth-order optimization in signal processing and machine learning: Principals, recent advances, and applications,","cited_arxiv_id":null,"evidence_quote":"supplies the zero-order gradient estimation method used to optimize the filtering threshold."},{"cited_title":"The JPEG2000 still image coding system: an overview,","cited_arxiv_id":null,"evidence_quote":"is the classic image codec baseline against which the proposed system's mask PSNR is compared."}],"review_version":1}