{"id":"69c8788e-b76c-4718-afdc-46ba816341f3","arxiv_id":"2412.14812","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A conditional decoupled diffusion model trained on CKMImageNet reconstructs complete channel knowledge maps from partial observations and localizes base stations more accurately than interpolation.","lead":"Using a diffusion model, this paper reconstructs complete radio channel maps from sparse measurements and finds base stations more accurately than interpolation methods. It is a step toward cheaper, environment-aware wireless networks in 6G.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The BS localization result in Table III is not reproducible: the paper never specifies how the BS position is inferred from the reconstructed CKM, so the 3.30 m error is unverifiable and may depend on an arbitrary post-processing rule.","rationale":"The central claim is an empirical performance improvement on CKMImageNet. The reader's weakest assumption was transferability of the learned prior to other environments. While that is a real limitation, the paper's stated claim is specifically about the dataset, so transferability concerns future applicability, not the validity of the reported experiments. A more immediate threat is the unreported BS localization protocol: Table III presents a quantitative metric that is central to the abstract's 'sensing performance' aspect, yet the transformation from a CKM image to a BS position is never defined. If the localization rule is arbitrary or tuned per method, the 3.30 m result is not meaningful. This is concrete and testable. The paper also lacks error bars and comparisons to learning-based baselines, but those are secondary. Therefore, the appropriate verdict remains CONDITIONAL, with the condition being full specification of the localization algorithm and robustness checks.","tokens_in":7490,"tokens_out":9483,"duration_ms":64376,"concrete_test":"Seek the localization algorithm from the authors and re-run the Table III experiment with at least two alternative localization rules (e.g., global argmax of channel gain, weighted centroid above a threshold, or a matched-filter peak) for all methods. If the relative ordering of BS loc. err. changes, the claim 'superior performance' for BS localization is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B and the conclusion claim superior BS localization performance (Table III: BS loc. err. 3.30 m for proposed vs. 8.81 m for KNN). However, the paper never states the algorithm that converts a reconstructed channel gain image into a predicted BS position. The scenario masks a 64m by 64m area around the BS, so the high-gain region itself is unobserved; the localization must rely on context (e.g., diffraction shadowing, distance-dependent path loss). Without knowing whether the method uses argmax, centroid, template matching, or a custom network, the comparison across methods is not well-defined. The reported improvement could be driven by the choice of localization rule rather than by the diffusion reconstruction quality. This is load-bearing because the BS localization claim is a headline result and is currently not independently checkable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript addresses CKM construction from partially observed channel gain images. It treats the problem as an inverse problem and applies a conditional decoupled diffusion model (DDM) from [14]: a VAE maps images into a latent space, a Swin-B encoder extracts features from the masked observation, and a dual-decoder UNet predicts both the signal-decay component and the noise component. Training and sampling are given in Algorithms 1 and 2. Experiments on CKMImageNet compare the method with KNN, Kriging, bilinear, and RBF interpolation under three masking settings (buildings masked, buildings not masked, and BS vicinity masked), reporting lower MSE/NMSE/RMSE/MAE in all settings and a BS localization error of 3.30 m.","tokens_in":7639,"tokens_out":6174,"duration_ms":43614,"significance":"If the results are taken at face value, the paper demonstrates that a conditional diffusion model trained only on partially observed CKM images can reconstruct radio maps more accurately than classical interpolation and can support BS localization. The use of a public benchmark dataset and the explicit training and sampling algorithms are strengths. However, the current validation has several load-bearing gaps: the building-masking evaluation appears to exclude the missing pixels, the BS localization rule is not specified, and no generative baselines or statistical uncertainties are reported. These issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"The evaluation for the building-masking scenario appears circular. Section IV-B states that error metrics are computed for all locations without building occupancy after masking buildings. Since the mask removes exactly the building pixels, the non-building pixels are the observed part of the input, and the missing building regions are excluded from the reported errors. Table I therefore does not demonstrate accurate reconstruction of the missing data. Please report metrics computed only on the masked (unobserved) pixels, or clarify the evaluation protocol if I have misread it.","section":"IV-B, Table I"},{"comment":"The BS localization result is not reproducible as reported. The paper states that a 64 m by 64 m region around the BS is masked and then reports a BS localization error of 3.30 m for the proposed method, but it never specifies how the BS position is estimated from the reconstructed CKM (e.g., global argmax, centroid, template matching, or a learned regressor). Without this rule, and without confirmation that the same rule is applied to all baselines, the comparison in Table III is not well defined. Please specify the localization algorithm and the test protocol.","section":"IV-B, Table III"},{"comment":"There is an inconsistency between the latent-space description and the algorithms. The text and Fig. 2 state that a VAE maps the pixel-space image x to a latent z and that diffusion is performed in latent space, but Algorithms 1 and 2 use x0, xt, and c = -x0 throughout and return xt as the reconstructed image. It is unclear whether the diffusion operates in pixel space or latent space, where the VAE encoder and decoder enter, and how the conditioning encoder e(y) is combined with the latent code. Please rewrite the algorithms to match the actual pipeline; this is essential for reproducibility.","section":"III-B, Algorithms 1-2 and Fig. 2"},{"comment":"The reported gains are point estimates without error bars, standard deviations, or significance tests. Tables I-III aggregate over 500-1000 images, but no variability is reported, so the reader cannot judge whether the improvements over the best baseline (e.g., RBF in Table II) are stable. Please report per-image statistics or confidence intervals, and avoid the word 'significantly' unless a statistical test is performed.","section":"IV-B, Tables I-III"},{"comment":"The paper is positioned against generative CKM methods, citing RadioUNet [10], RME-GAN [11], and RadioDiff [12], but none of these is used as a baseline. The experimental comparison is limited to interpolation methods, so the abstract's claim of significantly improving CKM construction compared with benchmarking schemes is only supported relative to interpolation. Please add at least one generative baseline (e.g., RadioDiff or RME-GAN) on the same dataset, or explicitly narrow the claim.","section":"I and IV-B"}],"minor_comments":[{"comment":"The sentence 'our proposed diffusion based model significantly outperforms significantly the benchmark approaches' contains a duplicated adverb; please revise.","section":"IV-B"},{"comment":"The phrase 'Table I plots the average error metrics' should be 'lists'; consider also adding units to the table headers.","section":"IV-B, Table I"},{"comment":"The phrase 'the smallest time step is 1×104' should be '1×10^-4' if that is the intended value.","section":"IV-A"},{"comment":"Algorithm 1 samples a random mask matrix H but does not specify its distribution; please state how the masks are generated during training.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript is a straightforward application of an existing diffusion model to a public dataset, and the novelty is primarily empirical. The main concerns are the circular evaluation in the building-masking experiment, the underspecified BS localization protocol, and the absence of generative baselines. If these are resolved, the paper could be a useful application study; in its current form I would not recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it does something genuinely useful: it takes a published decoupled diffusion model, adapts it with a conditional encoder and dual-decoder UNet, and shows that it can inpaint a channel knowledge map (CKM) from sparse observations without any auxiliary environment or BS location data. On the CKMImageNet benchmark the gains over KNN, Kriging, Bilinear, and RBF are consistent and large, especially when buildings are masked (RMSE drops from about 66 dB to 20.7 dB). The training and sampling algorithms are clearly specified, and the dataset is public, so the non-BS inpainting results are at least reproducible in principle. That is real empirical work. Second, the BS localization result in Table III has a load-bearing gap: the paper reports a 3.30 m localization error but never says how the BS position is inferred from the reconstructed CKM. The masked region around the BS is 64m by 64m, so the high-gain area itself is unobserved; the method must be relying on context. Without knowing whether the localization uses argmax, centroid, template matching, or something else, the comparison against KNN et al. is not well-defined and the claimed improvement could come from the post-processing rule rather than the diffusion reconstruction. That is not a reason to dismiss the paper, but it is a reason to require the authors to specify the algorithm before taking the headline claim seriously.\n\nThe soft spots are otherwise proportionate. The paper omits comparisons with the generative baselines it cites—RadioUNet, RME-GAN, RadioDiff—which are the natural competitors. There are no error bars or significance tests, and the learned prior is trained only on CKMImageNet ray-traced data (28 GHz, fixed heights, specific cities); transferability to other environments is untested. No code or data are released, which makes the missing details harder to fill in. The mathematical framework is sound and the paper does not overclaim beyond its experiments, aside from the under-specified BS localization part.\n\nWho is this for? Researchers working on CKM construction, radio map estimation, or inverse problems with diffusion models. It is a competent engineering contribution that advances a practical problem, but it is not a methodological breakthrough. It deserves a serious referee, and I would recommend acceptance only after the BS localization algorithm is specified, comparisons with generative baselines are added, and at least some variance estimates are reported.","headline":"A competent empirical application of decoupled diffusion models to CKM inpainting, with real gains over interpolation baselines but one headline result (BS localization) that is underspecified to the point of being unverifiable.","tokens_in":8160,"tokens_out":1551,"would_cite":true,"duration_ms":11746,"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":"A conditional diffusion model reconstructs complete channel knowledge maps from partial observations, outperforming interpolation baselines.","keywords":["channel knowledge map","CKM construction","diffusion model","decoupled diffusion model","inverse problem","image inpainting","base station localization","radio environment map"],"falsifier":"Take the trained model and run it on a held-out set of ray-traced 128 by 128 CKMs from an environment not represented in the training cities, or generated at a different carrier frequency such as 3.5 GHz or with a different receiver height. If the resulting RMSE and MAE on unobserved locations no longer beat Kriging or RBF interpolation, the claimed advantage is an artifact of the training distribution rather than a general CKM prior.","tokens_in":7255,"feed_emoji":"📡","tokens_out":6520,"duration_ms":42335,"temperature":0.7,"pith_summary":"Channel knowledge maps (CKMs) are site-specific databases of channel properties at every location, and building a complete map from sparse measurements is an ill-posed inverse problem. The paper proposes to solve this problem with a conditional decoupled diffusion model that learns the prior distribution of CKMs from partially observed channel gain images alone, without auxiliary environment maps or transmitter positions. On the CKMImageNet dataset, the model reconstructs missing regions with lower MSE, NMSE, RMSE, and MAE than KNN, Kriging, bilinear, and RBF interpolation, in both building-masked and non-building-masked scenarios. It also predicts channel knowledge around a base station accurately enough to locate the base station to 3.30 m in a complex urban environment. If the result holds, generative inpainting is a viable route to environment-aware wireless communication when dense channel measurements are unavailable.","feed_headline":"Diffusion inpainting beats interpolation for wireless channel maps","feed_subtitle":"Trained only on partial 28 GHz maps, it fills gaps and locates base stations to 3.3 m.","key_machinery":"The mechanism is the conditional decoupled diffusion model (DDM). Instead of DDPM's discrete noise schedule, DDM uses a continuous forward process $x_t = x_0 + \\int_0^t f_s\\,ds + \\int_0^t dw_s$ with the image-to-zero drift chosen as $f_t = -x_0$, so $x_t = (1-t)x_0 + \\sqrt{t}\\,\\epsilon$ and $x_1$ is pure Gaussian noise. The reverse process predicts both the drift $c_\\theta$ and the noise $\\epsilon_\\theta$ with a UNet that has two decoder branches, conditioned on multi-level features $e(y)$ extracted by a Swin-B encoder from the masked CKM image; training minimizes the sum of squared errors on $c$ and $\\epsilon$. A VAE maps pixel space to latent space to reduce computation. This decoupling lets sampling use arbitrary step sizes with few steps, while the condition encoder injects the observed locations at every reverse step.","core_discovery":"The central claim is that a conditional decoupled diffusion model, trained only on masked and unmasked CKM images, can serve as a generative prior that turns partial observations $y = Hx$ into a complete CKM $\\hat{x}$ by sampling from the posterior, circumventing the rank-deficiency of the mask $H$. The authors report that on 500 test images without building masking the proposed method reaches RMSE 10.7758 dB and MAE 5.1412 dB, the best among the four interpolation baselines; with building masking, RMSE is 20.6712 dB versus 43.2206 dB for Kriging, the best baseline. In a base-station vicinity masking test over 1000 images, the method's BS localization error is 3.30 m, compared with 8.81 m for KNN. The same architecture also produces visually plausible inpainted maps, supporting the claim that learned diffusion priors capture global propagation structure that local interpolation misses.","pith_inferences":["The paper does not simulate angle-of-arrival or angle-of-departure maps even though the system model includes them; applying the same conditional DDM to those modalities is a direct test of whether the learned prior transfers beyond channel gain.","The model is a conditional generator rather than a measurement-consistent sampler; explicitly enforcing $y = Hx$ at the end of sampling could reduce residual error at the cost of extra steps.","The training distribution is narrow (28 GHz, fixed antenna heights, urban ray-tracing); one stress test is to train on multi-frequency or multi-city data and measure whether the RMSE advantage over Kriging persists out of distribution."],"forward_implications":["CKM construction no longer requires auxiliary environmental data or transmitter location information, only a set of partially observed channel knowledge images.","Diffusion-based inpainting can outperform classical interpolation (KNN, Kriging, bilinear, RBF) on both reconstruction error and downstream localization.","The same trained model can locate a base station to within 3.30 m from a masked 64 m by 64 m vicinity, even when most links are non-line-of-sight.","The decoupled formulation allows flexible sampling steps, so the method can trade computation against reconstruction quality for real-time deployment.","Because the model inpaints the full map, it can provide channel knowledge at unmeasured locations for environment-aware beamforming and resource allocation."],"supporting_citations":[{"why":"Defines the CKM concept and motivates why complete channel knowledge at unmeasured locations is valuable for environment-aware communications.","marker":"[3]"},{"why":"Supplies the KNN interpolation baseline that the proposed method is compared against in Tables I, II, and III.","marker":"[7]"},{"why":"Supplies the Kriging interpolation baseline that the proposed method is compared against in Tables I, II, and III.","marker":"[8]"},{"why":"Supplies the RBF interpolation baseline that the proposed method is compared against in Tables I, II, and III.","marker":"[9]"},{"why":"Supplies the denoising diffusion probabilistic model framework and the simplified training objective that the CKM prior is built on.","marker":"[13]"},{"why":"Supplies the decoupled diffusion model architecture with image-to-zero and zero-to-noise stages that the conditional CKM inpainting method adapts.","marker":"[14]"},{"why":"Provides the CKMImageNet dataset of ray-traced channel gain images used for training and testing the model.","marker":"[15]"}],"fun_headline_variants":["Diffusion model fills channel map gaps from sparse data","Generative diffusion beats interpolation for wireless maps","Partial observations? Diffusion rebuilds the full channel map","Diffusion outshines Kriging in channel knowledge mapping","Sparse data? Diffusion generates accurate wireless channel maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the diffusion prior learned from CKMImageNet's 20,000 ray-traced 128 by 128 channel gain images (28 GHz, fixed 10 m transmitter and 1 m receiver heights, specific cities) is the true prior for every test CKM; if a target area differs in layout, frequency, building materials, or antenna heights, the learned prior will not match and the reported gains may disappear.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model fills channel map gaps from sparse data","Generative diffusion beats interpolation for wireless maps","Partial observations? Diffusion rebuilds the full channel map","Diffusion outshines Kriging in channel knowledge mapping","Sparse data? Diffusion generates accurate wireless channel maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000957,"raw_usage":{"total_tokens":4040,"prompt_tokens":866,"completion_tokens":3174,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":3098}},"tokens_in":482,"tokens_out":3174,"duration_ms":15494,"temperature":1.0,"reasoning_tokens":3098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:53:21.725871+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained model and run it on a held-out set of ray-traced 128 by 128 CKMs from an environment not represented in the training cities, or generated at a different carrier frequency such as 3.5 GHz or with a different receiver height. If the resulting RMSE and MAE on unobserved locations no longer beat Kriging or RBF interpolation, the claimed advantage is an artifact of the training distribution rather than a general CKM prior.","supporting_citations":[{"cited_title":"An adaptive k-nearest neighbor algorithm,","cited_arxiv_id":null,"evidence_quote":"Supplies the KNN interpolation baseline that the proposed method is compared against in Tables I, II, and III."},{"cited_title":"Fixed rank kriging for cellular coverage analysis,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kriging interpolation baseline that the proposed method is compared against in Tables I, II, and III."},{"cited_title":"Reducing the calibration effort for probabilistic indoor location estimation,","cited_arxiv_id":null,"evidence_quote":"Supplies the RBF interpolation baseline that the proposed method is compared against in Tables I, II, and III."},{"cited_title":"Denoising diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"Supplies the denoising diffusion probabilistic model framework and the simplified training objective that the CKM prior is built on."},{"cited_title":"CKMImageNet: A comprehensive dataset to enable channel knowledge map construction via computer vision,","cited_arxiv_id":null,"evidence_quote":"Provides the CKMImageNet dataset of ray-traced channel gain images used for training and testing the model."}],"review_version":1}