{"id":"109dfe2b-b0d2-4e14-b73a-cb77d4a457c9","arxiv_id":"2501.01460","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A dual-branch RWKV-CNN super-resolution network with a multi-scale wavelet loss reports up to 0.11 dB PSNR gain over HAT on remote sensing benchmarks at a fraction of the compute.","lead":"This paper introduces GDSR, a dual-branch neural network for remote sensing image super-resolution that pairs an RWKV global attention branch with a convolutional detail branch, plus a multi-scale wavelet loss. The authors report slightly higher PSNR than the Transformer-based HAT model while using fewer parameters and running several times faster, which matters for large satellite and drone imagery.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"0.09 dB PSNR advantage over HAT is the load-bearing claim, but it rests on single-run comparisons with no variance estimates or released checkpoints; the gap is comparable to internal ablation differences.","rationale":"I read the paper as proposing an efficient dual-branch RWKV-CNN architecture with a wavelet auxiliary loss, and claiming a quantitative SOTA result over HAT. The efficiency claim is arithmetically consistent and supported by the architecture description, so I credit it. The accuracy claim is the load-bearing part: the abstract's headline is a 0.09 dB average over 18 comparisons, and each comparison is a single run with no error bars or code. This is the same weak point the reader identified. My internal-magnitude observation strengthens it: the architecture ablations in Table IV show that swapping the global branch module changes PSNR by 0.14 and 0.06 dB, meaning the claimed 0.09 dB margin is within the range of module-level variation under this protocol. Without variance estimates or checkpoints, the superiority claim is not falsifiable. I therefore keep the reader's CONDITIONAL verdict unchanged; the correct response is to request seeds/checkpoints or a re-run with confidence intervals before accepting the quantitative claim.","tokens_in":26262,"tokens_out":4573,"duration_ms":41785,"concrete_test":"Retrain GDSR and HAT on AID x3 bicubic (and ideally RSSRD-QH CDM x2) with the same data splits and the paper's 200-epoch schedule, using 5 random seeds per method; report mean +/- std PSNR/SSIM/LPIPS and a paired 95% CI for the GDSR - HAT difference. If the authors' released checkpoints are available, recompute Table VII entries from them. The central claim stands only if the CI excludes 0 and reproduced numbers match within 0.02 dB.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The efficiency half of the central claim is internally consistent: 13.17/20.81 M params = 63%, 338.73/652.17 G FLOPs = 52%, and 14.8/4.6 FPS = 3.22x (Table X). The accuracy half is not. The \"average 0.09 dB over HAT\" is computed across 18 settings (Tables VII-VIII) with per-setting differences of only 0.00-0.23 dB, no standard deviations, no seeds, and no released checkpoints; the paper does not state whether HAT was retrained under the same 200-epoch schedule or evaluated with its original recipe. The fragility is visible internally: Table IV's ablation replacing RWKV with Transformer (GDSR_TC) or Mamba (GDSR_MC) changes PSNR by 0.14 and 0.06 dB on AID x3, the same order as the headline 0.09 dB gap. On perceptual metrics the \"outperforms\" claim is already ambiguous: on RSSRD-QH CDM x2 GDSR ties HAT in PSNR (23.75) but has worse LPIPS (0.6197 vs 0.6120). Thus the SOTA-superiority claim depends entirely on the unverified assumption that the single-run gap exceeds training variance and that HAT was not disadvantaged by the comparison protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GDSR, a dual-branch network for remote sensing image super-resolution that combines an RWKV-based global branch with a convolutional detail branch, fused by a Global-Detail Reconstruction Module, and trained with an additional dual-group multi-scale wavelet loss. Experiments on AID, UCMerced, and a custom drone-based RSSRD-QH dataset under bicubic and comprehensive degradation report PSNR/SSIM/LPIPS comparisons against CNN, GAN, Transformer, Mamba, and hybrid baselines, plus ablations of branches, RWKV alternatives, depth, and wavelet-loss variants. The abstract's headline claim is that GDSR outperforms HAT by on average 0.09 dB PSNR while using 63% of its parameters, 51% of its FLOPs, and running 3.2 times faster.","tokens_in":26509,"tokens_out":6938,"duration_ms":64887,"significance":"If the headline accuracy claim is accepted, the contribution is a useful efficiency-oriented architecture: the RWKV-CNN dual-branch design is a plausible way to obtain linear-complexity global modeling with local detail refinement, and the efficiency arithmetic in Table X is internally consistent (13.17M/20.81M parameters, 338.73G/652.17G FLOPs, 14.8/4.6 FPS). The 30-category AID breakdown in Table IX is a genuinely informative generalization analysis, and the authors are candid about LPIPS trade-offs and about their model's limitations. However, the paper does not ship code, checkpoints, or the RSSRD-QH dataset, and all headline numbers are single runs, so the quantitative superiority over HAT is not yet verified; the significance of the empirical claim is therefore conditional on added statistical and reproducibility evidence.","major_comments":[{"comment":"The headline 'average 0.09 dB over HAT' is not supported as stated. The 18 per-setting PSNR differences range from 0.00 dB (RSSRD-QH CDM x2, where GDSR is also worse in SSIM and LPIPS) to 0.23 dB, with no error bars, no repeated seeds, and no released checkpoints. These gaps are the same order as ablation differences in Table IV, where replacing RWKV with Transformer or Mamba shifts PSNR by 0.14 dB and 0.06 dB on AID x3. Without run-to-run variance estimates or released trained models, the reported 0.09 dB advantage cannot be distinguished from training noise. Please report mean and standard deviation over at least three seeds for GDSR and the closest baselines, or release checkpoints so the comparison can be reproduced.","section":"Section IV-D, Tables VII and VIII"},{"comment":"The comparison protocol is underspecified for baselines. The text states that training used 64x64 patches, batch size 16, Adam, and 200 epochs, but it does not state whether HAT, SwinIR, MambaIR, FreMamba, and the other baselines were retrained under exactly the same patch sampling, degradation generation, learning-rate schedule, and hyperparameter budgets, or whether published checkpoints/final models were used. If the baselines were not trained under the identical protocol, the accuracy and efficiency comparison is not controlled. Please document per-baseline training budgets, data splits, and degradation code, and preferably release training scripts.","section":"Section IV-B"},{"comment":"The wavelet-loss contribution is too small to be considered established. The proposed Dual-Group Multi-Scale loss adds 0.0185 dB PSNR and 0.0008 SSIM over the L1 baseline while increasing LPIPS from 0.3476 to 0.3519; the text calls this a 'significant' improvement. With a single run, this delta is indistinguishable from training noise, and the LPIPS regression runs against the abstract's 'enhanced reconstruction fidelity' claim if perceptual quality is included. Please provide seeded trials or temper the claim to a small PSNR/SSIM improvement with a known perceptual cost.","section":"Section IV-C, Table VI"},{"comment":"The custom RSSRD-QH dataset is central to the evaluation, including six of the eighteen settings averaged in the headline claim, but no dataset URL, download link, or release plan is provided; only the geographic distribution of the counties is listed in Table I. Independent verification of the SOTA claim is therefore impossible. Please release the dataset under a clear availability statement or, if it cannot be released, re-state the headline claim based only on public benchmarks.","section":"Section IV-A, Table VIII"}],"minor_comments":[{"comment":"The abstract contains the typo 'GSDR' instead of 'GDSR'; please correct it throughout.","section":"Abstract"},{"comment":"Several figure captions are duplicated and mismatched: the ERF figure is labeled both 'Figure 1' and 'Fig. 2', the wavelet-loss overview appears as both 'Figure 2' and 'Fig. 6', and the feature-visualization figure appears as both 'Figure 5' and 'Fig. 7'. Please renumber the figures consistently.","section":"Figures 1, 6, and 7"},{"comment":"The text refers to experiments on 'F2B and F2C', but the ablation study in Section IV-C only defines F2B and F2D; F2C is never introduced. Please correct the reference or add the missing definition.","section":"Section IV-D, paragraph after Table IX"},{"comment":"There are small formatting errors: the Table IV header reads 'Effect of RWKV .' with an extra space, and Table IX has 'Categoties' instead of 'Categories'.","section":"Table IV and Table IX"}],"recommendation":"major_revision","confidential_remarks":"I agree with the stress-test assessment: the efficiency half of the central claim is internally consistent, but the accuracy half rests on single-run comparisons with no variance estimates, no released checkpoints, and an underspecified baseline protocol. The wavelet-loss contribution is also numerically marginal. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The paper is within the journal's scope; the main gate should be reproducible, multi-seed evidence and a clear availability statement for the custom dataset."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a real architectural idea and an honest set of ablations; the efficiency arithmetic checks out. But the headline accuracy claim over HAT is not yet supported.\n\nWhat's actually new: first use of RWKV with Re-WKV and Omni-Shift for remote sensing SR; the Global-Detail dual-branch design (RWKV branch + CNN branch with GDRM fusion) and the Dual-Group Multi-Scale Wavelet Loss are concrete, specific additions. The paper describes the modules well enough to reimplement, and the ablations (branch counts, depth, replacement of RWKV with Transformer/Mamba, wavelet loss variants) are more thorough than typical. Credit where due: the limitation paragraph in the conclusion is straight about the parameter-heavy detail branch and the lack of hyperspectral evaluation, and Table VI openly shows that wavelet loss hurts LPIPS. The efficiency numbers are internally consistent: 13.17/20.81M params, 338.73/652.17 GFLOPs, 14.8 vs 4.6 FPS.\n\nSoft spots, in proportion. The quantitative SOTA claim is the load-bearing one. The 0.09 dB average over HAT is the mean of 18 settings, many of which are 0.02-0.05 dB apart, and no standard deviations, seeds, or checkpoints are provided. The internal ablated model variants GDSR_TC and GDSR_MC differ from GDSR by 0.06 and 0.14 dB on the same AID x3 benchmark, which is the same order as the headline advantage. That means the claimed edge over HAT is plausibly within training variance. LPIPS is worse than FreMamba on several settings, so 'state-of-the-art' depends on which metric you look at. The missing direct comparison to Restore-RWKV, which supplies the RWKV blocks, makes it hard to attribute the gain. The private RSSRD-QH dataset is not released, and the paper doesn't say whether HAT was retrained under the identical 200-epoch recipe or evaluated with official weights.\n\nNone of this kills the architectural claim; the dual-branch idea is plausible and the efficiency story is compelling. But the specific 0.09 dB SOTA statement needs repeated runs, released code/data, and matched training budgets before I'd trust it.\n\nWho it's for: people working on efficient SR backbones or RWKV for vision will find the design worth reading. As a rigorous benchmark claim, it needs more evidence.\n\nRecommendation: send to peer review. It deserves a serious referee, with the understanding that the quantitative section needs substantial strengthening before acceptance.","headline":"GDSR is a genuinely new RWKV-CNN dual-branch SR design with an internally consistent efficiency story, but the 0.09 dB SOTA claim over HAT rests on single-run numbers with no variance estimates or released artifacts.","tokens_in":27160,"tokens_out":3683,"would_cite":true,"duration_ms":32862,"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":"GDSR beats HAT with 63% of its parameters","keywords":["remote sensing image super-resolution","RWKV","dual-branch network","global-detail reconstruction","wavelet loss","linear-complexity attention","image restoration","efficient super-resolution"],"falsifier":"Train GDSR, HAT, SwinIR, MambaIR, and FreMamba from the same data splits with three or more seeds and identical degradations, then compare the mean and standard deviation of PSNR and SSIM; the central claim fails if the mean GDSR-HAT gap falls within one standard deviation of the repeated runs, or if HAT wins on any dataset and scale, and the 3.2x speed claim fails if both models are timed on the same GPU with identical input sizes and batch sizes.","tokens_in":25996,"feed_emoji":"🛰️","tokens_out":6928,"duration_ms":61905,"temperature":0.7,"pith_summary":"Remote sensing images are large and contain both broad scene structure and fine local texture, and most super-resolution models concentrate on just one of these. This paper argues that the two should be handled by separate branches working in parallel and then fused. Its central claim is that a network pairing an RWKV-based global branch with a convolutional detail branch, joined by a Global-Detail Reconstruction Module and trained with a multi-scale wavelet loss, beats the leading Transformer-based method HAT by an average of 0.09 dB PSNR while using only 63% of its parameters, 51% of its FLOPs, and running 3.2 times faster. If correct, this makes efficient, high-quality super-resolution practical for large-scale remote sensing imagery.","feed_headline":"GDSR beats HAT with 63% of its parameters","feed_subtitle":"Dual-branch RWKV-CNN super-resolution runs 3.2x faster at similar or better quality.","key_machinery":"The mechanism that carries the argument is the Global-Detail dual-branch architecture. One branch, the Residual Global Extraction Group, applies RWKV (Receptance Weighted Key Value, a linear-complexity attention-style sequence model adapted to images) to model long-range dependencies, while the other, the Residual Details Extraction Group, applies residual convolutions for local texture. Between them sits the Global-Detail Reconstruction Module, whose Permuted Spatial Attention aligns both feature maps and an adaptive gating layer fuses them, with the fused output fed only into the detail branch to avoid frequency interference. The training objective adds a Dual-Group Multi-Scale Wavelet Loss that compares stationary-wavelet subbands of the SR and HR images separately for low-frequency structure and aggregated high-frequency detail at three scales, so the network is pushed to recover high-frequency information rather than only minimizing pixel error.","core_discovery":"On its own terms, the paper's discovery is that global and local feature extraction are complementary low-pass and high-pass operations, and that explicitly decoupling them into a dual-branch architecture yields better reconstruction than any single-paradigm network. The global branch uses RWKV's recurrent weighted key-value attention with omnidirectional token shifting to capture long-range dependencies, while the detail branch stacks residual convolutional blocks to recover fine textures. The Global-Detail Reconstruction Module aligns and adaptively fuses the two feature streams, and the Dual-Group Multi-Scale Wavelet Loss constrains low-frequency structure and aggregated high-frequency subbands across three resolutions. Across the AID, UCMerced, and RSSRD-QH benchmarks under both bicubic and comprehensive degradation, the authors report that GDSR beats the state-of-the-art Transformer HAT by 0.04 to 0.11 dB in PSNR, with the average gain at 0.09 dB, while consuming 63% of the parameters and 51% of the FLOPs and running 3.2 times faster.","pith_inferences":["The dual-branch low-pass/high-pass decoupling is not specific to super-resolution; the same RWKV-global/CNN-detail split with attention fusion could be tested on denoising, deblurring, or pan-sharpening of remote sensing images, but the paper does not make that claim.","Because RWKV has linear complexity in sequence length while windowed Transformers have growing costs, the 3.2x speed advantage at 160x160 inputs may widen at larger operational satellite-image patches; this is extrapolation, since the paper only reports one input size.","The reported margins over HAT are small enough that repeated-seed training with released checkpoints is needed to know whether the average 0.09 dB advantage is systematic; the paper reports single runs, so the stability of the ranking is an open question.","If the wavelet loss's LPIPS regression is considered a cost, combining it with a perceptual or adversarial term might preserve fidelity gains while closing the perceptual gap; the paper only evaluates it with L1 reconstruction."],"forward_implications":["RWKV becomes a viable backbone for large-scale image restoration, not just language modeling, because its linear complexity sidesteps the quadratic cost of self-attention on high-resolution remote sensing inputs.","The reported 0.04 to 0.11 dB PSNR gains over HAT, SwinIR, and MambaIR suggest that decoupling global and local processing helps across scales (x2, x3, x4) and across two degradation types, including a real-world UAV dataset.","Because the model has 13.17M parameters, 338.73G FLOPs, and 14.8 FPS at 160x160 input, it fits deployment settings where Transformer and Mamba baselines are too slow or memory-heavy.","The wavelet loss improves PSNR and SSIM over pixel-loss training while slightly worsening LPIPS, and the ablation shows the multi-scale dual-group design mitigates that tradeoff among wavelet-based losses.","Retraining with stripe noise added to the degradation model removes sensor stripes, indicating the framework can absorb remote-sensing-specific degradations beyond the standard CDM."],"supporting_citations":[{"why":"The Transformer-based HAT is the state-of-the-art baseline GDSR claims to outperform by an average of 0.09 dB PSNR while using fewer parameters and FLOPs.","marker":"[13]"},{"why":"Supplies the Re-WKV attention and Omni-Shift token shifting used in the global extraction branch of GDSR.","marker":"[20]"},{"why":"Establishes RWKV as an efficient 2D vision architecture whose global modeling motivates the RGEG branch.","marker":"[33]"},{"why":"SwinIR is a window-attention Transformer baseline and also provides the RTSB used in the GDSR-TC ablation.","marker":"[12]"},{"why":"MambaIR is the state-space-model baseline and provides the ResidualGroup used in the GDSR-MC ablation.","marker":"[15]"},{"why":"Provides the wavelet-loss construction and the uniform/asymmetric coefficient baselines that the proposed Dual-Group Multi-Scale Wavelet Loss is compared against.","marker":"[41]"},{"why":"Establishes the wavelet-domain subband loss paradigm that the proposed loss extends to multi-scale dual-group weighting.","marker":"[40]"},{"why":"The AID benchmark supplies the primary training and test images for the main quantitative comparisons and per-category generalization results.","marker":"[44]"},{"why":"The UCMerced benchmark provides the second public dataset used for the cross-dataset accuracy comparisons.","marker":"[45]"},{"why":"The Local Attribution Map analysis supports the claim that the global branch activates more pixels and the detail branch uses them correctly, which is the mechanism behind GDSR's gains.","marker":"[53]"}],"fun_headline_variants":["GDSR beats HAT with 63% params and 3.2x speed","Wavelet-loss dual branch: GDSR beats HAT, 3.2x faster","GDSR: RWKV+CNN beats HAT with 51% FLOPs","Remote sensing SR: GDSR cuts params by 37%, runs 3.2x faster","Dual-branch GDSR: wavelet losses, linear complexity, beats HAT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy advantage over HAT and the other baselines assumes that all compared models were trained with the same data splits, the same degradation pipeline, and comparable tuning effort, and that differences of 0.04 to 0.11 dB are larger than run-to-run variation; the paper reports no repeated seeds, error bars, or released checkpoints to confirm that equivalence.","fun_headline_variants_meta":{"raw":{"variants":["GDSR beats HAT with 63% params and 3.2x speed","Wavelet-loss dual branch: GDSR beats HAT, 3.2x faster","GDSR: RWKV+CNN beats HAT with 51% FLOPs","Remote sensing SR: GDSR cuts params by 37%, runs 3.2x faster","Dual-branch GDSR: wavelet losses, linear complexity, beats HAT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000451,"raw_usage":{"total_tokens":2341,"prompt_tokens":1081,"completion_tokens":1260,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":1142}},"tokens_in":697,"tokens_out":1260,"duration_ms":8617,"temperature":1.0,"reasoning_tokens":1142,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:51:58.298139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GDSR, HAT, SwinIR, MambaIR, and FreMamba from the same data splits with three or more seeds and identical degradations, then compare the mean and standard deviation of PSNR and SSIM; the central claim fails if the mean GDSR-HAT gap falls within one standard deviation of the repeated runs, or if HAT wins on any dataset and scale, and the 3.2x speed claim fails if both models are timed on the same GPU with identical input sizes and batch sizes.","supporting_citations":[{"cited_title":"Vision-RWKV: Efficient and scalable visual per- ception with RWKV-like architectures,","cited_arxiv_id":null,"evidence_quote":"Establishes RWKV as an efficient 2D vision architecture whose global modeling motivates the RGEG branch."},{"cited_title":"Activating more pixels in image super-resolution transformer,","cited_arxiv_id":null,"evidence_quote":"The Transformer-based HAT is the state-of-the-art baseline GDSR claims to outperform by an average of 0.09 dB PSNR while using fewer parameters and FLOPs."},{"cited_title":"Mambair: A simple baseline for image restoration with state-space model,","cited_arxiv_id":null,"evidence_quote":"MambaIR is the state-space-model baseline and provides the ResidualGroup used in the GDSR-MC ablation."},{"cited_title":"Training transformer models by wavelet losses improves quantitative and visual performance in single image super-resolution,","cited_arxiv_id":null,"evidence_quote":"Provides the wavelet-loss construction and the uniform/asymmetric coefficient baselines that the proposed Dual-Group Multi-Scale Wavelet Loss is compared against."},{"cited_title":"Training generative image super-resolution models by wavelet-domain losses enables better control of artifacts,","cited_arxiv_id":null,"evidence_quote":"Establishes the wavelet-domain subband loss paradigm that the proposed loss extends to multi-scale dual-group weighting."},{"cited_title":"Bag-of-visual-words and spatial extensions for land-use classification,","cited_arxiv_id":null,"evidence_quote":"The UCMerced benchmark provides the second public dataset used for the cross-dataset accuracy comparisons."}],"review_version":1}