{"id":"50244091-9d59-46e2-8aa1-94f6b5788213","arxiv_id":"2411.16187","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Sending nine key points per view and regenerating metaverse scenes with generative AI cuts wireless latency by 92.6% in simulation, but the reported accuracy gains rely on a denoiser that uses the ground truth as input.","lead":"Instead of sending full camera images, this system sends nine key points per view over a noisy wireless link and reconstructs the metaverse scene with stable diffusion and NeRF. In a simulated factory, the authors report a 92.6% latency cut and better reconstruction accuracy than full-image transmission. The accuracy gain is weakened because the denoising step is given the original key points as its target.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OT denoiser is evaluated with oracle access to clean transmitter keypoints (Algorithm 1 lines 5 and 26; §IV.A), making the reported KPE gains and part of the headline accuracy supervised artifacts.","rationale":"The central claim of the paper is that the full GSC pipeline (HgNet encoder, OT denoiser, SD/ControlNet decoder, NeRF) significantly improves metaverse construction accuracy and latency over conventional image transmission. The latency claim is structurally sound: transmitting nine keypoints per frame instead of full images must reduce wireless transfer time, and the reported 92.6% is not the vulnerable part. The accuracy claim is the load-bearing one, and it rests on the OT denoiser. The reader's weakest-assumption analysis correctly identifies that the denoiser is given the clean transmitter keypoint distribution as its target. Section IV.A says the optimization uses 'the original point vectors at the transmitter side', and Algorithm 1 initializes both marginals from K_i and applies the transport plan to K_i. A receiver cannot know these values; they are the very quantities being estimated. Consequently the KPE improvements attributed to OT (5.3% and 10.2% at 0 dB) and part of the headline 45.6% accuracy gain are supervised artifacts, not properties of a deployable system. This is an internal validation flaw, not merely a deviation from consensus. Additional weaknesses (no error bars, underspecified ImageCom KPE comparison, contradictory attribution of the 45.6%/44.7% gains across GSCS-OT and GSCM-OT in §V.C vs §V.D, no code or data) reinforce the case but are secondary. The proposed architecture is plausible and the latency advantage likely real, so a corrected evaluation with a receiver-only denoiser could change the verdict; as written, the central accuracy claim is unsupported.","tokens_in":14826,"tokens_out":5137,"duration_ms":81495,"concrete_test":"Modify Algorithm 1 to remove oracle access: replace every use of the clean keypoint set K_i (lines 5 and 26) with the received noisy keypoint set K'_i (or with a marginal estimated from neighboring views/frames), keep everything else identical, and re-run the Section V.C KPE comparison at 0 dB under AWGN and Rayleigh channels. If GSCS-OT/GSCM-OT no longer beat GSCS/GSCM by the reported 5.3%/10.2%, or if the 45.6% advantage over ImageCom shrinks substantially, the OT contribution and headline accuracy claims are artifacts of access to ground truth.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.A frames the OT problem as moving received key points K' to 'original point vectors at the transmitter side' K, and Eq. (20) uses marginals p_i and q_j of the two distributions. Algorithm 1 then initializes both marginals from K_i (line 5) and updates the denoised output as T*·K_i (line 26). In deployment the receiver does not know K_i, so the OT denoiser is not implementable as described: it is given the clean ground truth it is supposed to recover. The claimed 0 dB KPE gains (5.3% for GSCS-OT over GSCS, 10.2% for GSCM-OT over GSCM, and the 45.6% advantage over ImageCom) therefore measure an oracle-assisted procedure, not a communication system. If K_i in Algorithm 1 is instead intended to mean the received noisy keypoints, then Eq. (20) and Section IV.A's 'original point vectors at the transmitter side' are internally inconsistent, and the OT step would only match the noisy distribution to itself, not to the clean source. Either way, the paper does not validate a deployable denoiser. The latency reduction (92.6%) and the architectural idea remain plausible, but the accuracy/experience claims rest on this supervised simulation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a goal-oriented semantic communication (GSC) framework for wireless metaverse construction. The transmitter uses an hourglass-network encoder to extract a small set of keypoints (nine per camera view) from images, and the receiver reconstructs the metaverse scenery using a Stable Diffusion–based image generator and a NeRF-based 3D constructor, aided by a shared knowledge base. To combat channel noise, the authors add an optimal-transport (OT) denoiser that selectively corrects received keypoints. Experiments in a simulated industrial factory compare the proposed GSCS/GSCM variants (with and without OT) against a conventional full-image transmission baseline (ImageCom). The paper claims a 92.6% latency reduction, a 45.6% improvement in object status accuracy (KPE), and a 44.7% improvement in viewing experience (P2Point).","tokens_in":14992,"tokens_out":5221,"duration_ms":47930,"significance":"If the reported accuracy and latency results were valid, the paper would present a compelling demonstration that a semantic pipeline transmitting only nine keypoints per camera view, together with a generative decoder, can replace full-image wireless transmission for metaverse construction. The latency claim is structurally plausible, because transmitting nine points instead of 1200×600 images must reduce transmission time. The architectural integration of keypoint extraction, Stable Diffusion, NeRF, and OT is also potentially interesting. However, the accuracy and viewing-experience claims are not supported as presented because the OT denoiser is evaluated with oracle access to the clean transmitter keypoints, as detailed below. The paper does provide a concrete algorithm (Algorithm 1), a clear experimental setup, and comparisons across AWGN, Rayleigh, and Rician channels, which are strengths. If the oracle issue can be resolved, the framework could be a meaningful step toward low-latency metaverse construction.","major_comments":[{"comment":"The OT denoiser uses the transmitter's clean keypoint distribution as its optimization target. Eq. (20) defines the transport problem as moving mass from the received distribution Dr to the distribution Dt of the 'original point vectors at the transmitter side'. Algorithm 1 line 5 initializes both marginal distributions p_i and q_j from K_i (the clean keypoint set), and line 26 computes the denoised points as T*·K_i. In a deployed receiver, the clean transmitter keypoints are not available, so this denoiser cannot be implemented as described. Consequently, the reported KPE gains of 5.3% (GSCS-OT over GSCS) and 10.2% (GSCM-OT over GSCM) at 0 dB (Section V.C), as well as the headline 45.6% accuracy improvement over ImageCom, measure a supervised/oracle procedure rather than a communication system. This is a load-bearing issue because the central accuracy and viewing-experience claims rest on these numbers.","section":"Section IV.A and Algorithm 1"},{"comment":"If the authors intended K_i in Algorithm 1 to denote the received noisy keypoints rather than the clean transmitter keypoints, then p and q are both initialized from the same noisy set, and the transport matrix T* would map the noisy distribution onto itself. In that case, Eq. (20) and the accompanying text about 'original point vectors at the transmitter side' are internally inconsistent, and the OT step would provide no denoising relative to the clean source. The manuscript must resolve this ambiguity, either by reformulating the OT problem as a blind denoiser that does not require the clean transmitter distribution or by explicitly stating the oracle assumption and re-evaluating all accuracy claims without access to the transmitter's keypoints.","section":"Algorithm 1 vs. Section IV.A"},{"comment":"The KPE metric in Eq. (9) is defined using 'three dimensional position value' of keypoints, but Section V.C states that the error is measured using the Euclidean distance between the key points' two-dimensional positions. Since the keypoints are extracted from 2D heatmaps (Section III.B), the use of 2D distances is plausible, but the inconsistency between the metric definition and the actual evaluation makes it difficult to interpret the quantitative claims. Please clarify the exact coordinates used in KPE and align the definition with the experimental protocol.","section":"Eq. (9) and Section V.C"}],"minor_comments":[{"comment":"The text says 'Section IV-A provides insights into the configuration of metaverse scenarios and OT denoiser' but the experiment setup is actually in Section V-A. Similarly, Section V.B refers to 'Section IV-B' for experimental results; the correct cross-reference is Section V.B.","section":"Section V.A"},{"comment":"The problem formulation contains typos: 'min_{Pr} C(P_t, P_t)' should presumably be 'C(P_t, P_r)', and 'trθ,δru' is garbled. Please proofread the equation.","section":"Eq. (7)"},{"comment":"The paper claims that traditional OT with Sinkhorn-Knopp has O(n^3) complexity, but Sinkhorn's algorithm is O(n^2) per iteration. The stated O(n^2) complexity of the proposed relaxed algorithm is therefore not a fair comparison as written.","section":"Section IV.A"},{"comment":"The regularization parameter is denoted ε in Algorithm 1 but η in Section IV.A. Please use a consistent symbol.","section":"Algorithm 1"},{"comment":"The sentence 'the GSCS-OT framework also achieves an 92.6% reduction' has a grammatical error; also the units or scale of the latency values in Fig. 12 are not specified.","section":"Section V.F"},{"comment":"Reference [21] cites AnimateDiff rather than the original Stable Diffusion paper; the correct source should be cited. Reference [26] appears unrelated to optimal transport and should be replaced with a standard OT reference.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The oracle-access issue in the OT denoiser is the central concern. If the authors cannot provide a blind denoiser baseline, the paper's main contribution would be reduced to a latency argument plus a generative reconstruction pipeline, with the accuracy claims substantially weakened. I would encourage the editor to require a clear statement of any transmitter-distribution side information that is assumed available at the receiver, and a re-evaluation of the KPE/P2Point results without oracle knowledge, before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: the latency reduction is real, but the accuracy numbers are not. The paper's architecture integrates HgNet keypoint extraction, ControlNet/Stable Diffusion image generation, NeRF reconstruction, and a relaxed OT denoiser for metaverse construction. That integration is new, and the latency claim is structurally sound—transmitting nine keypoints per view instead of 1200x600 images has to cut transmission time. The selective correction filter in Eq. (30) is also a reasonable idea.\n\nThe soft spot is load-bearing. Section IV.A sets the OT problem as moving the received keypoint vectors to the \"original point vectors at the transmitter side,\" and Algorithm 1 (lines 5 and 26) initializes both marginals and the final transport from that clean set Ki. A receiver in deployment does not know the clean keypoints it is trying to recover. So the reported OT gains—5.3% and 10.2% KPE improvements at 0 dB, and part of the headline 45.6% over ImageCom—are measured with oracle access to ground truth. If Ki in the algorithm was meant to be the received keypoints instead, then the OT step matches the noisy distribution to itself, which contradicts Eq. (20) and Section IV.A. Either way, the denoiser is not implementable as described, and the accuracy claims do not hold.\n\nThe evaluation has other weaknesses, though they are minor by comparison: no error bars, the ImageCom KPE comparison is underspecified, and Section V.C first ranks GSCM-OT best at 0 dB but then attributes the 45.6% improvement to GSCS-OT. No code or data is released. The latency measurements are plausible but could not be independently verified from the manuscript alone.\n\nWho this is for: researchers working on semantic communication for metaverse/XR applications. They will get a useful architectural blueprint and a credible latency argument. The accuracy evaluation needs a redesign, but the core idea is promising enough that a serious referee should see it. I'd recommend sending it to peer review with the expectation of heavy revision: fix the OT denoiser with a shared prior or feedback channel, rerun the benchmarks, add error bars, and clarify the ImageCom comparison. I would not cite the accuracy numbers in their current form.","headline":"Latency claim is credible and the system integration is new, but the OT denoiser is given clean transmitter keypoints as its target, so the accuracy improvements are supervised artifacts.","tokens_in":15675,"tokens_out":3231,"would_cite":false,"duration_ms":29232,"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 only nine semantic key points per camera view, with a shared knowledge base and a diffusion-plus-NeRF decoder, reconstructs a metaverse faster and more accurately than sending full images over wireless…","keywords":["semantic communication","metaverse construction","optimal transport denoising","keypoint extraction","stable diffusion","NeRF","wireless transmission","goal-oriented communication"],"falsifier":"Re-run the reported 0 dB experiments with the OT denoiser's target distribution estimated only from the received noisy key points, or from a prior trained without the transmitter's clean points, and compare key point error with the claimed 5.3-10.2% OT gains; if the improvement disappears or reverses, the supervised access to the clean key points is what produced the denoising advantage.","tokens_in":16,"feed_emoji":"🏭","tokens_out":8127,"duration_ms":143885,"temperature":0.7,"pith_summary":"The paper is trying to establish that a metaverse scene can be built over wireless links from a tiny slice of semantic information instead of a full image stream. In its simulated industrial factory, a camera view is reduced to nine key points, encoding the positions of a moving box and a robotic arm, plus an initial shared knowledge base of scenery, camera, and object locations. The receiver then reconstructs multi-view images with a diffusion model and assembles the 3D scenery with NeRF. Against conventional image-based construction, the paper reports a 92.6% reduction in wireless construction latency, a 45.6% improvement in object-status accuracy, and a 44.7% improvement in viewing experience. The interest is in showing that meaning-preserving compression plus generative reconstruction can beat raw image delivery precisely when the channel is noisy.","feed_headline":"Nine key points per view build a metaverse 92.6% faster","feed_subtitle":"Simulated factory tests: 92.6% lower latency, 45.6% better status accuracy than image transmission.","key_machinery":"The mechanism that carries the argument is the pairing of a sparse key-point semantic representation with a shared knowledge base. The semantic encoder, an hourglass network, turns each camera image into nine heatmaps whose maxima define the key points of the moving objects; because the static factory, camera poses, and initial object positions are sent once as a knowledge base, only the small key-point set needs to travel for each update. The receiver's decoder is built from a ControlNet-enabled Stable Diffusion model that synthesizes multiple rotated views conditioned on key points, prompts, and canny edges, followed by NeRF, which lifts those images into a 3D point cloud. The OT-enabled denoiser is the added safeguard against channel noise: it computes a relaxed optimal-transport matrix between received and transmitted key-point sets, combines row- and column-constrained solutions in O($n^{2}$), and applies the correction only to points flagged by a view-consistency filter.","core_discovery":"The paper's central claim, stated on its own terms, is that a goal-oriented semantic communication framework can replace bit-oriented image transmission in metaverse construction without losing the information that matters for object status and viewing experience. The transmitter extracts nine key points per view with an hourglass network, transmits those points over AWGN, Rayleigh, or Rician channels, and a receiver-side OT-enabled denoiser selectively corrects key points that disagree across camera angles. A ControlNet-conditioned Stable Diffusion decoder turns the corrected key points, text prompts, and canny-edge priors into multi-view images, and NeRF builds the metaverse point cloud from them. In the paper's experiments the semantic frameworks rank ahead of image-based construction across the metrics, with GSCS-OT reducing key point error by 45.6% and GSCM-OT reducing point-to-point error by 44.7% relative to ImageCom at 0 dB, and GSCS-OT cutting transmission latency by 92.6%.","pith_inferences":["A practical extension would be a blind OT denoiser that estimates the target distribution from the received key points or a learned prior; without that, the 5.3-10.2% denoising gains are only validated in the supervised simulation setting.","The same transmit-once-shared-prior design should extend to larger metaverse scenes with more movable objects; the number of key points per view and the O(n^2) OT solve would grow, so the latency and accuracy trade-off would need re-testing.","Selective correction across views is a geometric consistency check; in deployments with heterogeneous UAV angles or occlusions, the threshold and neighboring-view window would likely need to adapt to the scene's motion model.","The framework implies that the receiver's compute budget, diffusion sampling and NeRF rendering, may become the next bottleneck for real-time metaverse updates, since the wireless transfer itself is no longer the dominant delay."],"forward_implications":["If the reported gains hold, real-time metaverse construction does not need multi-gigabit image streams: nine key points per view plus a one-time knowledge base is enough to regenerate the scene.","At low SNR, a generative decoder with a shared prior can outperform full-image delivery because the prior supplies structure that the noisy channel has destroyed.","The OT denoiser is a compatible add-on to the semantic pipeline, contributing about 5.3-10.2% key-point-error improvement on top of diffusion-based reconstruction rather than being the whole accuracy story.","Latency shifts from the wireless link to receiver-side computation; semantic extraction, OT correction, and image generation together take less than two seconds per frame in the reported setup."],"supporting_citations":[{"why":"supplies the Point-NeRF industrial factory scene and the NeRF-based metaverse construction approach the framework builds on.","marker":"[24]"},{"why":"defines NeRF, the neural radiance field reconstruction used to build the 3D metaverse from generated images.","marker":"[22]"},{"why":"provides ControlNet, the conditional control mechanism used by the semantic decoder to generate multi-view images from key points and canny edges.","marker":"[31]"},{"why":"is cited as the Stable Diffusion basis for the generative image model at the core of the semantic decoder.","marker":"[21]"},{"why":"supplies the prior use of optimal transport in semantic communications that the OT-enabled denoiser extends.","marker":"[27]"},{"why":"provides the unbalanced optimal transport theory motivating the relaxed transport formulation and its numerical solution.","marker":"[25]"},{"why":"defines the P2Point evaluation metric used to compare transmitter and receiver point clouds.","marker":"[30]"},{"why":"supports the choice of point clouds with motion and color as the evaluation basis for object status accuracy.","marker":"[29]"}],"fun_headline_variants":["Metaverse via 9 key points: 92.6% faster, 45.6% more accurate","Semantic comms: 9 key points build metaverse 92.6% faster","AI + optimal transport slash metaverse latency by 92.6%","Goal-oriented semantics cut metaverse build time 92.6%","9 points per view: metaverse built 92.6% faster, precise"],"cache_read_input_tokens":17664,"weakest_assumption_plain":"The weakest load-bearing assumption is that the OT denoiser's target distribution comes from the transmitter's clean key points: Algorithm 1 initializes the marginals from the known clean key-point set, but a real receiver does not know the clean values it is trying to reconstruct.","fun_headline_variants_meta":{"raw":{"variants":["Metaverse via 9 key points: 92.6% faster, 45.6% more accurate","Semantic comms: 9 key points build metaverse 92.6% faster","AI + optimal transport slash metaverse latency by 92.6%","Goal-oriented semantics cut metaverse build time 92.6%","9 points per view: metaverse built 92.6% faster, precise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000312,"raw_usage":{"total_tokens":1785,"prompt_tokens":964,"completion_tokens":821,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":726}},"tokens_in":580,"tokens_out":821,"duration_ms":6488,"temperature":1.0,"reasoning_tokens":726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:26:13.236794+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the reported 0 dB experiments with the OT denoiser's target distribution estimated only from the received noisy key points, or from a prior trained without the transmitter's clean points, and compare key point error with the claimed 5.3-10.2% OT gains; if the improvement disappears or reverses, the supervised access to the clean key points is what produced the denoising advantage.","supporting_citations":[{"cited_title":"Evalua- tion criteria for PCC (point cloud compression),","cited_arxiv_id":null,"evidence_quote":"defines the P2Point evaluation metric used to compare transmitter and receiver point clouds."},{"cited_title":"Point-NeRF: Point-based neural radiance fields,","cited_arxiv_id":null,"evidence_quote":"supplies the Point-NeRF industrial factory scene and the NeRF-based metaverse construction approach the framework builds on."},{"cited_title":"NeRF: Representing scenes as neural radiance fields for view synthesis,","cited_arxiv_id":null,"evidence_quote":"defines NeRF, the neural radiance field reconstruction used to build the 3D metaverse from generated images."},{"cited_title":"Adding conditional control to text-to-image diffusion models,","cited_arxiv_id":null,"evidence_quote":"provides ControlNet, the conditional control mechanism used by the semantic decoder to generate multi-view images from key points and canny edges."},{"cited_title":"Semantic channel equalizer: Modelling language mismatch in multi-user semantic commu- nications,","cited_arxiv_id":null,"evidence_quote":"supplies the prior use of optimal transport in semantic communications that the OT-enabled denoiser extends."},{"cited_title":"Unbalanced optimal transport: From theory to numerics,","cited_arxiv_id":null,"evidence_quote":"provides the unbalanced optimal transport theory motivating the relaxed transport formulation and its numerical solution."},{"cited_title":"Impact of resolution, colour, and motion on object identification in digital twins from robot sensor data,","cited_arxiv_id":null,"evidence_quote":"supports the choice of point clouds with motion and color as the evaluation basis for object status accuracy."}],"review_version":1}