{"id":"a2b48926-1b2f-4ead-a279-e92c2c9618f3","arxiv_id":"2507.09218","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A diffusion-model-based image denoising pipeline applied to STFT spectrograms is proposed to enhance low-SNR bistatic ISAC signals, reportedly cutting estimation error by 63% in simulations.","lead":"This paper proposes turning weak radio signals into pictures, cleaning the pictures with an AI image denoiser, and turning them back into signals to better measure the distance, speed, and angle of drones. It reports a 63% error reduction in simulations, but the method's diffusion-model step is described in a way that would not actually denoise the input.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2's diffusion step is undefined: the reverse process in Algorithm 1 samples from pure Gaussian noise, so feeding the observed RGB image as x_T cannot denoise that image; the 63% claim is unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: Algorithm 2 uses x_T for both the constructed RGB image and the pure-noise starting point of Algorithm 1, creating a fundamental inconsistency. We agree with this assessment. The central claim is that image-space diffusion denoising yields a 63% RMSE improvement; if the denoising step is not defined, the claim cannot be validated. The paper does not describe any conditional sampling, guidance, or noise-remapping mechanism to connect the observed noisy image to the reverse diffusion process. Additional issues, such as the undefined x_hat_t term in the loss function (Eq. 13), the lack of code/data release, and the absence of error bars, compound the uncertainty but are secondary to the core algorithmic gap. To be fair, the authors could have intended a conditional diffusion scheme like guided diffusion or RePaint, but none is described, so the method as presented is internally inconsistent and unreproducible. This is not a disagreement with external consensus; it is a correctness risk in the algorithm specification itself.","tokens_in":18820,"tokens_out":3704,"duration_ms":45507,"concrete_test":"Implement the exact Algorithm 1 with T=500 and the reported improved UNet architecture. Feed the observed RGB image from Algorithm 2 Step 2 as the initial x_T and run the reverse iterations. Then apply ISTFT to the output and compare the reconstructed signal to both the ground truth and the input signal. If the output does not match the input and does not improve over the input, or if drawing x_T from N(0,I) yields the same output distribution, the method as described does not perform the claimed denoising. Also re-run the experiment with x_T drawn from N(0,I) and measure RMSE; if the reported improvement persists, the input image is not being used, confirming the inference procedure is mis-specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central enhancement step is not reproducibly specified. Algorithm 2 Step 3 (line 8) instructs: 'Feed x_T into pretrained diffusion model G, and perform the image super-resolution operation according to Algorithm 1.' Algorithm 1 (lines 1–6) begins with 'x_T ∈ N(0,I)' and iterates the reverse update (line 4) using the learned noise predictor. This sampling procedure generates an image from the model's prior; it has no term that conditions on the observed RGB image constructed in Step 2. If the observed image is literally used as the initial x_T, the input is not a sample from the Gaussian prior that the reverse process is trained to invert, and the update rule (Eq. 12, line 4) has no denoising interpretation because the time index t of the observed image is unknown. The paper provides no posterior-sampling or conditioning mechanism (e.g., guidance, replacement, or inpainting) to map a specific noisy image to its enhanced counterpart. Consequently, as written, the output x_0 is either independent of the input (if Algorithm 1 is followed exactly) or an arbitrary prior sample, so the reported 63% RMSE improvement cannot be reproduced from the described method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an image super-resolution-based signal enhancement (ISR-SE) framework for bistatic ISAC systems. Received low-SNR signals are converted via STFT into RGB images whose channels encode magnitude, frequency index, and phase; an improved UNet-diffusion model is then applied to denoise or enhance these images, after which ISTFT reconstructs the time-domain signal for MUSIC and 2D-DFT based angle, range, and velocity estimation. The central claim is that the proposed method improves estimation accuracy by 63% compared to traditional signal processing.","tokens_in":19114,"tokens_out":5180,"duration_ms":64399,"significance":"If the claimed result held, the paper would illustrate a useful direction: applying generative image-space priors to low-SNR bistatic ISAC signal recovery. The manuscript provides a complete system model, a concrete pipeline, and comparisons against TSP, LMS, and CNN baselines. However, the key enhancement step is not reproducibly specified, the training and evaluation do not meet the standards needed to support the empirical claim, and no code or data are provided. The 63% improvement is thus a fitted simulation outcome rather than a demonstrated, reproducible result. The idea may have merit, but the current manuscript does not establish it.","major_comments":[{"comment":"The reverse diffusion process in Algorithm 1 is defined only for x_T drawn from N(0,I), with no conditioning term on an observed image. Algorithm 2 instructs the reader to feed the observed RGB image x_T into the pretrained model and perform the operation according to Algorithm 1. If the observed image is used as x_T, it is not a sample from the Gaussian prior and its time index t is unknown, so the update in Eq. (12) has no denoising interpretation; if pure noise is used, the output is unrelated to the input. No guidance, replacement, inpainting, or other conditioning mechanism is described. As written, the enhancement step does not map a specific noisy image to a denoised version, making the reported results unreproducible.","section":"Algorithm 2 (Section III-C) and Algorithm 1 (Section III-B)"},{"comment":"The composite loss in Eq. (13) includes SSIM(x_t, \\hat{x}_t), but \\hat{x}_t is never defined, and the paper does not specify how this loss is optimized relative to the standard diffusion noise-prediction objective. The training dataset is described only as 'collected through our own measurement and acquisition process', with no size, SNR distribution, or train/test split; Figs. 6 and 7 show only training loss curves, not validation or test performance. This prevents independent assessment of generalization and makes the claimed improvement unverifiable.","section":"Section III-B, Eq. (13) and training description"},{"comment":"The RGB construction does not preserve the claimed information. The green channel in Eq. (17) is only the normalized frequency index and is deterministic for a fixed STFT configuration; it carries no signal-dependent information. More importantly, the paper gives no explicit inverse mapping from enhanced RGB values to a complex STFT for the ISTFT: it states that RGB channels are 'converted' back to amplitude, frequency, and phase, but no equations or procedures are provided. Without a defined inverse, the claim that the pipeline reconstructs a high-SNR time-domain signal is unsupported.","section":"Section III-A, Eqs. (16)-(18) and Section III-C, Step 4"},{"comment":"The 63% improvement claim is not well-defined. The abstract says the method 'improves the estimation accuracy by 63%', while Section V says 'over 63% compared to the TSP-based method'; no SNR value, specific metric, or statistical significance is attached to this number. The evaluation is entirely based on the authors' own simulations and a custom dataset, with no independent test set or external benchmark. The baselines (e.g., LMS step size, CNN architecture) are also insufficiently specified, so the comparison cannot be reproduced.","section":"Abstract and Section V"}],"minor_comments":[{"comment":"The variance matrix in Eq. (10) is written with a summation symbol (\\sum_t) and is not defined; use a consistent notation such as \\Sigma_t.","section":"Section III-B, Eq. (10)"},{"comment":"Figure 2 contains Chinese text ('改进的UNet去噪网络'); all figure labels should be in English.","section":"Section III-B, Fig. 2"},{"comment":"Specify the LMS step size and filter length, the CNN architecture and training hyperparameters, and the exact evaluation protocol for all baselines.","section":"Section V, baselines"},{"comment":"The variables f_e_l and tau_e_l in Eq. (19) are not defined; clarify their meaning or remove the superscript 'e'.","section":"Section IV, Eq. (19)"},{"comment":"Reference [36] is mis-formatted: Denoising Diffusion Probabilistic Models is an arXiv paper (arXiv:2006.11239) and is not published in IEEE Wireless Communications Letters; please correct the citation.","section":"References"},{"comment":"State the image dimensions used for the RGB representation and clarify whether the diffusion model operates on 8-bit quantized values or on continuous normalized values.","section":"Section III-A, Step 4"}],"recommendation":"reject","confidential_remarks":"The central methodological step is underspecified, and the empirical claim cannot be verified from the manuscript. A corrected submission would need to specify a valid conditioning mechanism for the diffusion model, define the inverse RGB-to-spectrogram mapping, and provide reproducible training and evaluation details. I recommend rejection of the current version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take on arXiv:2507.09218. The idea: map a low-SNR bistatic ISAC signal to an RGB image via STFT (magnitude, frequency index, phase in separate channels), denoise with a UNet-diffusion hybrid, then ISTFT back and feed into MUSIC + 2D-DFT for angle/range/velocity. That combination, as far as I can tell, is new for bistatic ISAC. The paper is clearly written, the system model is standard, and the simulation setup—Rician channel, OFDM, UPA arrays—is reasonable. Comparing against TSP, LMS, and CNN is the right kind of baseline set.\n\nThe problem is the core enhancement step. The paper defines the reverse diffusion process in Algorithm 1 starting from x_T ~ N(0,I), with no conditioning on any observation. Then Algorithm 2 takes the constructed RGB image, calls it x_T, and asks the model to \"perform image super-resolution according to Algorithm 1.\" That does not work. The observed RGB image is not a sample from the Gaussian prior, and the reverse update has no term that ties the output to that specific input. There is no guidance, no replacement, no posterior sampling. So as written, the output would be either a random sample from the model's prior or, if you literally start from the observed image, something that has no denoising interpretation because the noise level at that time index is undefined. The loss function in Eq. (13) also mixes noise prediction with SSIM on the corrupted image, which is not a standard DDPM objective. These are not minor presentation issues; they make the central claim—63% RMSE improvement—not reproducible from the described method.\n\nSecondary concerns: the green channel encodes a deterministic frequency index, but the diffusion model will alter all three channels. The paper does not explain how the modified green channel maps back to a valid frequency axis. That could corrupt the ISTFT. Also, the training data is described as \"our own measurement\" with no details, and no code or data is released. For a simulation paper, that is a lot to take on faith.\n\nWhere does that leave us? The application idea has some merit; the rest of the signal processing is standard and fine. But the main contribution, as specified, is broken. I would not cite it in its current form. Still, I think it deserves a serious referee rather than a desk reject: the flaw is fixable (use a conditional diffusion model or a proper posterior sampling scheme), and the problem—low-SNR enhancement for bistatic ISAC—is relevant enough that a careful revision could be worth publishing. If you send it out, have the reviewer focus on Algorithm 1 vs Algorithm 2 and the RGB-to-STFT mapping.\n\nBest.","headline":"Plausible application idea for low-SNR bistatic ISAC enhancement, but the diffusion denoising step as written cannot denoise the input, so the headline 63% result is not reproducible.","tokens_in":19606,"tokens_out":2953,"would_cite":false,"duration_ms":33076,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that encoding bistatic ISAC echoes as RGB spectrograms and denoising them with a UNet-augmented diffusion model reduces estimation error by 63 percent relative to adaptive filtering.","keywords":["bistatic ISAC","signal enhancement","diffusion models","image super-resolution","short-time Fourier transform","low-SNR sensing","UAV detection","generative AI"],"falsifier":"Run the described pipeline exactly: feed a noisy RGB spectrogram into the reverse diffusion process as written and check whether the output is a denoised version of that same spectrogram. If the reverse process instead produces an image unrelated to the input, the described enhancement mechanism does not operate as stated, and the reported 63 percent accuracy gain would require another explanation.","tokens_in":18620,"feed_emoji":"📡","tokens_out":8781,"duration_ms":93399,"temperature":0.7,"pith_summary":"The paper proposes treating the enhancement of weak bistatic ISAC echoes as an image-restoration problem. Received signals are converted by a Short-Time Fourier Transform into spectrograms, whose magnitude, frequency index, and phase are encoded into the red, green, and blue channels of an RGB image; a diffusion model with an improved UNet noise predictor is then used to denoise the image, and an inverse STFT returns an enhanced time-domain signal. The paper's central claim is that this image-space generative route recovers clean spectral structure under low SNR far better than traditional adaptive filtering or CNN baselines, and it reports a 63 percent improvement in estimation accuracy. The reason a reader should care is that the framework opens a path for generative AI to support joint sensing and communication in future networks without changing the OFDM-based infrastructure.","feed_headline":"Image-space diffusion sharpens weak ISAC signals by 63 percent","feed_subtitle":"STFT spectrograms become RGB images; a UNet-diffusion denoiser lifts low-SNR sensing accuracy for bistatic ISAC.","key_machinery":"The load-bearing mechanism is the combination of the RGB spectrogram encoding and the reverse diffusion process. The encoding is given by explicit formulas: $R = \\log(|Y(f,t)|+\\varepsilon)/\\log(|M_{\\max}|+\\varepsilon)\\times 255$, $G = (f-f_{\\min})/(f_{\\max}-f_{\\min})\\times 255$, and $B = (\\angle Y(f,t)+\\pi)/(2\\pi)\\times 255$. The denoiser is a diffusion model run for $T=500$ steps, with a UNet-based noise predictor enhanced by residual blocks, channel attention, and residual concatenation, trained on a composite loss of MSE and SSIM. The reverse process iteratively removes predicted noise and reconstructs the clean image, which is the step that converts image-domain denoising into signal-domain SNR gain.","core_discovery":"On the paper's own terms, the discovery is that a deterministic RGB encoding of an STFT spectrogram is a sufficient and effective representation for signal enhancement: the red channel carries log-scaled magnitude, the green channel carries normalized frequency index, and the blue channel carries phase. Training a diffusion denoiser on such images lets the model learn the spectral texture of clean OFDM-based ISAC signals, so that reversing the diffusion process on a noisy image suppresses noise while preserving the time-frequency structure needed for sensing. The enhanced image is converted back through ISTFT to a time-domain signal, after which MUSIC-based angle estimation and 2D-DFT range/Doppler estimation run on the restored channel. The authors report that this pipeline reduces estimation RMSE by over 63 percent relative to traditional signal processing and improves communication BER by orders of magnitude at low SNR.","pith_inferences":["Editorial inference: the paper does not describe how the observed noisy image enters the reverse diffusion process; as written, Algorithm 1 starts from pure Gaussian noise. A conditional diffusion formulation, where the observed RGB image steers each reverse step, is the natural way to make the pipeline consistent with standard denoising diffusion.","Editorial inference: because the RGB encoding and ISTFT are invertible, the same denoiser could be retrained for other OFDM-based bistatic sensing setups or other weak-signal radar problems, provided the training data span the relevant SNR range.","Editorial inference: the 63 percent figure is tied to the paper's custom training dataset; making that dataset or the SNR-dependent gain curves public would let others test how well image-space denoising generalizes beyond the simulated scenario."],"forward_implications":["At SNR levels down to -20 dB, the ISR-SE method keeps noise-estimation MSE below all three baselines and retains a lower RMSE for angle, range, and velocity estimates.","Because the enhanced output is a time-domain signal, both communication demodulation and passive sensing benefit from the same processing chain, which is why the reported BER also improves by 1-2 orders of magnitude over CNN at low SNR.","The power-allocation sweep shows a quantifiable tradeoff: increasing the sensing power factor $\\beta_R$ improves range RMSE with diminishing returns while pushing BER toward $10^{-1}$, so systems must pick an operating point between communication and sensing quality.","The framework replaces hand-designed filters and statistical noise assumptions with learned generative priors, so the same architecture can be retrained for different environments without changing the RGB construction or the sensing algorithms."],"supporting_citations":[{"why":"Supplies the Denoising Diffusion Probabilistic Models equations for the forward and reverse diffusion processes used in the enhancement network.","marker":"[36]"},{"why":"Demonstrates conditional denoising diffusion for data reconstruction in wireless communications, the closest prior application the method builds on.","marker":"[34]"},{"why":"Applies diffusion-model-aided reconstruction in a wireless downlink, supporting the use of diffusion for signal recovery.","marker":"[35]"},{"why":"Provides the CNN-based image super-resolution baseline used in the comparative experiments.","marker":"[22]"},{"why":"Provides the CNN-based denoising of weak signals that serves as a baseline in the comparison.","marker":"[23]"},{"why":"Supplies the LMS adaptive filtering algorithm used as a traditional signal enhancement baseline.","marker":"[18]"},{"why":"Supports the LMS-family baseline with noise-variance estimation for adaptive filter coefficient updates.","marker":"[19]"},{"why":"Supplies the channel model parameters used in the simulations of the bistatic ISAC scenario.","marker":"[37]"}],"fun_headline_variants":["Diffusion-based RGB image denoising cuts ISAC sensing error by 63%","UNet-diffusion on STFT spectrograms boosts bistatic ISAC by 63%","Image super-resolution pipeline enhances weak ISAC signals","RGB-mapped spectrograms help diffusion model restore ISAC signals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that feeding the observed noisy RGB image into the reverse diffusion process produces a denoised version of that same image, even though the reverse process is defined only for pure Gaussian noise and no conditioning on the observed image is described.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-based RGB image denoising cuts ISAC sensing error by 63%","UNet-diffusion on STFT spectrograms boosts bistatic ISAC by 63%","Image super-resolution pipeline enhances weak ISAC signals","RGB-mapped spectrograms help diffusion model restore ISAC signals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000277,"raw_usage":{"total_tokens":1686,"prompt_tokens":1014,"completion_tokens":672,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":630,"tokens_out":672,"duration_ms":8159,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:01:31.890573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the described pipeline exactly: feed a noisy RGB spectrogram into the reverse diffusion process as written and check whether the output is a denoised version of that same spectrogram. If the reverse process instead produces an image unrelated to the input, the described enhancement mechanism does not operate as stated, and the reported 63 percent accuracy gain would require another explanation.","supporting_citations":[{"cited_title":"Conditional Denoising Diffusion Probabilistic Models for Data Reconstruction Enhancement in Wireless Communications,","cited_arxiv_id":null,"evidence_quote":"Demonstrates conditional denoising diffusion for data reconstruction in wireless communications, the closest prior application the method builds on."},{"cited_title":"Diffusion Model-Aided Data Reconstruction in Cell-Free Massive MIMO Downlink: A Computation- Aware Approach,","cited_arxiv_id":null,"evidence_quote":"Applies diffusion-model-aided reconstruction in a wireless downlink, supporting the use of diffusion for signal recovery."},{"cited_title":"Deep CNN-Based Super-Resolution Using External and Internal Examples,","cited_arxiv_id":null,"evidence_quote":"Provides the CNN-based image super-resolution baseline used in the comparative experiments."},{"cited_title":"A Method for Denoising Seismic Signals With a CNN Based on an Attention Mechanism,","cited_arxiv_id":null,"evidence_quote":"Provides the CNN-based denoising of weak signals that serves as a baseline in the comparison."},{"cited_title":"Adaptive Noise Suppression of Pediatric Lung Auscultations With Real Applications to Noisy Clinical Settings in Developing Countries,","cited_arxiv_id":null,"evidence_quote":"Supplies the LMS adaptive filtering algorithm used as a traditional signal enhancement baseline."},{"cited_title":"Estimation of Measurement-Noise Variance for Variable-Step- Size NLMS Filters,","cited_arxiv_id":null,"evidence_quote":"Supports the LMS-family baseline with noise-variance estimation for adaptive filter coefficient updates."},{"cited_title":"Study on Channel Model for Frequencies From 0.5 to 100 GHz, V14.3.0,","cited_arxiv_id":null,"evidence_quote":"Supplies the channel model parameters used in the simulations of the bistatic ISAC scenario."}],"review_version":1}