{"id":"3dfe713a-66a5-4a5c-b0e0-5d4c120a3418","arxiv_id":"2504.16637","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RouteWinFormer selects a few relevant nearby windows per image patch and reports state-of-the-art PSNR on eight named restoration benchmarks across defocus deblurring, desnowing, dehazing, and deraining.","lead":"This paper presents RouteWinFormer, a Transformer for cleaning corrupted images that gathers information from a small set of nearby image windows chosen dynamically, rather than from the whole image. It reports top benchmark scores on defocus blur, snow, haze, and rain removal, while claiming that long-range attention is unnecessary for most restoration.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Own tables contradict the abstract's SOTA claim: RWF-S is below FSNet on Rain100H (PSNR 31.49 vs 31.77) and loses SSIM on NH-HAZE (0.69 vs 0.81); the paper also says 9 datasets but lists 8.","rationale":"After reading the paper in good faith, the most load-bearing point is not the Figure 1 attention-distance measurement. Even if that measurement is model-dependent, the route-window mechanism could still be shown to work by clean, reproducible benchmarks. The decisive problem is that the benchmarks printed in the paper are not clean: the abstract's claim of outperforming SOTA on all 9 datasets is falsified by the paper's own numbers on Rain100H (Table 5) and, on the SSIM axis, on NH-HAZE (Table 4). These are not interpretive issues or missing artifacts; they are direct contradictions of the strongest claim. I therefore disagree with the reader's choice of weakest assumption, while agreeing that the paper is not ready as-is. The correct disposition is REJECT for the claim as currently stated; a resubmission that narrows the claim to specific datasets/baselines and provides code, data splits, and the missing MSR/lambda details could be reconsidered as a conditional accept.","tokens_in":13026,"tokens_out":9205,"duration_ms":78013,"concrete_test":"Re-tabulate every dataset-metric cell in Tables 1, 3, 4, and 5 as printed, counting wins/losses versus FSNet, ConvIR, OKNet, and GRL. If any loss remains, rewrite the abstract and Section 4 to state per-dataset and per-baseline comparisons explicitly (e.g., 'beats all compared methods on DPDD, CSD, Haze4K, Rain100L, and Test1200; is competitive on Rain100H and NH-HAZE'). Independently, request the authors' evaluation code and per-image metric files; recomputing the two disputed rows (Rain100H, NH-HAZE) would settle whether the broad claim can be restored.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that RouteWinFormer 'outperforms state-of-the-art methods across 9 datasets' (abstract; Section 4). This claim is contradicted by the paper's own tables. In Table 5 (Section 4.2, Image DeRaining), RWF-S on Rain100H is PSNR 31.49 / SSIM 0.904, while FSNet, immediately above it, is 31.77 / 0.906. The text concedes only improvement over ConvIR (31.09 / 0.903), not over FSNet. In Table 4 (Image Dehazing), on NH-HAZE RWF-S reaches PSNR 21.49 but SSIM 0.69, while FSNet and ConvIR reach 0.81 and 0.80 SSIM respectively; a 0.9 dB PSNR gain combined with a roughly 0.1 SSIM loss is not 'consistent' outperformance. Additionally, the abstract and the start of Section 4 state 9 datasets, but the tables enumerate 8 (DPDD, CSD, SRRS, Haze4K, NH-HAZE, Rain100H, Rain100L, Test1200). The architecture may be salvageable, but the headline empirical claim as written is not supported by the evidence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RouteWinFormer, a U-shaped window-based Transformer for image restoration. It introduces the Route-Windows Attention Module (RW AM), which dynamically selects the top-k most relevant nearby windows on the basis of cross-shaped or rectangle-shaped regional similarity, and a Multi-Scale Structure Regularization (MSR) term that supervises sub-scale residual predictions against down/upsampled clean images. The paper reports PSNR/SSIM results on defocus deblurring (DPDD), desnowing (CSD, SRRS), dehazing (Haze4K, NH-HAZE), and deraining (Rain100H, Rain100L, Test1200), claiming state-of-the-art performance across nine datasets.","tokens_in":13325,"tokens_out":4987,"duration_ms":44594,"significance":"If the reported results are reproducible, the paper makes a plausible contribution: it offers a dynamic middle-range attention mechanism with a reduced complexity compared with global attention, and a model-agnostic structural regularization that is shown in the ablations to give consistent PSNR gains. The ablation studies in Tables 6 and 7 support the value of the two main components. However, the manuscript currently lacks code, model weights, dataset splits, and error bars, and several key definitions and hyperparameters are missing. More importantly, the headline claim that RouteWinFormer outperforms state-of-the-art methods across all datasets is contradicted by the paper's own tables. The core ideas are interesting, but the empirical claims need substantial correction and additional documentation before the contribution can be assessed reliably.","major_comments":[{"comment":"The central claim that RouteWinFormer \"outperforms state-of-the-art methods across 9 datasets\" is contradicted by the paper's own tables. In Table 5, RWF-S on Rain100H reaches PSNR 31.49 / SSIM 0.904, below FSNet's 31.77 / 0.906; the text only claims superiority over ConvIR on that dataset. In Table 4, on NH-HAZE RWF-S has SSIM 0.69 versus FSNet's 0.81 and ConvIR's 0.80, a substantial structural-similarity regression despite a PSNR gain. The abstract and the start of Section 4 state \"9 datasets,\" but the tables enumerate eight (DPDD, CSD, SRRS, Haze4K, NH-HAZE, Rain100H, Rain100L, Test1200). These inconsistencies must be resolved and the state-of-the-art claim must be made consistent with the reported numbers.","section":"Abstract and §4.2, Tables 4–5"},{"comment":"The motivating measurement in Figure 1 is not reproducible as presented: the paper does not state which model produced the attention maps, which datasets and how many images were used, how the attention distance was normalized, or whether the same pretrained model was applied to all degradation types. Since average attention distance is architecture- and training-dependent, Figure 1 does not establish that long-range attention is unnecessary for image restoration. Please provide the complete measurement protocol, or soften the motivation to a qualitative observation that does not rely on an unreported experiment.","section":"§1, Figure 1"},{"comment":"The Multi-Scale Structure Regularization is not fully specified. Equation (9) uses R_i, described as \"the predicted residual image at i-th scale,\" but the manuscript does not define the network head or projection that produces R_i from the sub-scale features. The loss weight λ in Eq. (10) is introduced but never given a value. In addition, Eq. (11) states that α is \"empirically default set to 0.1, respectively,\" without an ablation or reference. Please define R_i and λ explicitly and provide evidence for the choice of α.","section":"§3.3, Eqs. (9)–(10)"},{"comment":"Several architectural hyperparameters are undefined or under-specified: the window size k, the number of candidate windows rn, and the top-k count ri do not appear in Table 2. The implementation details only say that the default ri is [1, 1, 1, 1]. Without these values, the complexity claim O(ri k^2 h w c) and the reported FLOPs cannot be verified. Please report k, rn, and ri for each model variant, and clarify how the top-k selection is performed relative to these values.","section":"§3.2, Eqs. (4)–(8) and Table 2"},{"comment":"The experimental evaluation is not independently checkable as reported: no code, trained models, or random seeds are provided; no error bars or confidence intervals are given; and dataset split/partition details are absent (e.g., which DPDD indoor/outdoor images are used and how NH-HAZE is evaluated). For a paper whose central claim is empirical superiority, this level of reporting is insufficient. Please release code and checkpoints, and include error bars or at least standard deviations over multiple runs.","section":"§4.1, Tables 1–7"}],"minor_comments":[{"comment":"The phrase \"Route-Windows Attnetion\" contains a typo; it should be \"Attention.\"","section":"Abstract"},{"comment":"The text says RouteWinFormer achieves a 0.9 dB gain over OKNet on NH-HAZE, but OKNet is not listed in Table 4; please add the comparison or correct the sentence.","section":"§4.2, Image Dehazing"},{"comment":"The PSNR labels in the qualitative figures are confusing because \"Input PSNR Target\" appears as a single line and the per-method values are not clearly aligned with their images; please reformat and verify the values, for example the ordering in Figure 3 is ambiguous.","section":"Figures 3–6"},{"comment":"The notation in Eq. (2) reuses X for both the intermediate and output of the FFN, which makes the expression GELU(DWConv(Xt)) * Xb ambiguous; please introduce distinct variable names.","section":"§3.1, Eq. (2)"},{"comment":"References [He et al., 2010a] and [He et al., 2010b] are duplicates of the same paper; merge them.","section":"References"},{"comment":"The abbreviations \"w MSR\" and \"wo MSR\" are inconsistent with the \"with/without MSR\" phrasing used in the text; please make them uniform.","section":"Table 7"}],"recommendation":"major_revision","confidential_remarks":"The core ideas are not unsound, but as written the empirical claims are overstated and several load-bearing details are missing. I would not recommend acceptance without corrected tables, a detailed measurement protocol for Figure 1, full definitions of R_i and λ, and clearly reported hyperparameters. No concerns about citation behavior or novelty disclosure beyond what is stated in the report."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the architecture is reasonable and the experiments are broad, but the paper's central claim—that RouteWinFormer beats state-of-the-art on 9 datasets—is not supported by its own tables. On Rain100H, RWF-S gets 31.49/0.904 vs FSNet's 31.77/0.906; on NH-HAZE, RWF-S PSNR is best but SSIM at 0.69 trails FSNet (0.81) and ConvIR (0.80). That's not consistent SOTA across datasets, and the abstract and Section 4 say '9 datasets' while the tables enumerate 8 (DPDD, CSD, SRRS, Haze4K, NH-HAZE, Rain100H, Rain100L, Test1200).\n\nWhat's actually new: top-k routed window attention with cross- and rectangle-shaped candidate regions, applied to restoration, plus the MSR auxiliary loss. Dynamic routing is known from BiFormer and deformable attention, but the specific combination and the shape variety are a reasonable engineering contribution. The paper does well on breadth: four tasks, recent strong baselines (FSNet, ConvIR, OKNet, GRL), and ablations showing both RW AM and MSR give reproducible gains (e.g., +0.31 dB on CSD for RW AM, +0.45 dB on DPDD indoor for MSR). Efficiency numbers are reported for three model sizes.\n\nSoft spots:\n\n1. The overclaim above. The paper only says it beats ConvIR on Rain100H/L, not FSNet, and on NH-HAZE the SSIM regression is ignored in the text. That needs to be fixed in the abstract and conclusion.\n\n2. Figure 1's \"normalized average attention distance\" is presented without saying which model, which images, or how normalization was done. The middle-range premise is therefore unsupported. If the maps come from a pretrained long-range model, the inference that long-range attention is unnecessary doesn't follow.\n\n3. Missing details: the MSR residual heads R_i are never defined (Fig 2 shows R2–R4 but the text doesn't describe them), the loss weight lambda in Eq 10 is not given, and neither the candidate window count rn nor the window size k appear in Table 2. All fixable, but as written the method isn't fully reproducible.\n\n4. Minor: Table 4 doesn't include OKNet although the text credits it; some visual captions are confusing.\n\nThe architecture is plausible and the empirical wins on DPDD, CSD, Test1200, and Haze4K are substantial enough to take seriously. A good referee could push this into shape. I'd send it to review (not desk reject) with a request for a major revision: tighten claims to match tables, fill the missing definitions, and ideally release code.\n\nFor your reading group: it's a decent example of how quickly \"SOTA across N datasets\" collapses under one table check.","headline":"A plausible restoration architecture with a wide experimental net, but the 'SOTA on 9 datasets' headline is contradicted by its own tables on Rain100H and NH-HAZE.","tokens_in":13831,"tokens_out":5730,"would_cite":false,"duration_ms":47566,"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":"Use mid-range attention, not global, for image restoration","keywords":["image restoration","window attention","transformer","middle-range attention","top-k routing","multi-scale structure regularization","defocus deblurring","deraining"],"falsifier":"Take a trained global-attention restoration model, compute its average attention distance on each of the nine datasets with an explicit normalization rule, and compare with the 0.3 threshold; if any dataset exceeds 0.3, or if increasing RouteWinFormer's top-$k$ to include farther windows improves PSNR on images with large-scale degradation, the central claim is contradicted.","tokens_in":12814,"feed_emoji":"🖼️","tokens_out":6674,"duration_ms":58315,"temperature":0.7,"pith_summary":"This paper argues that image restoration transformers do not need global attention: a normalized average attention distance measured across several degradation datasets stays below 0.3, so the pixels a restoration model needs to consult are usually nearby, not across the whole image. Building on that, the paper proposes RouteWinFormer, a U-shaped window transformer whose Route-Windows Attention Module lets each window attend to a small set of relevant neighboring windows selected dynamically by a router. A second contribution, Multi-Scale Structure Regularization, forces the network's sub-scales to reconstruct structure from down-up-sampled clean images, giving the full scale a generalized prior for degradation patterns. The paper reports consistent PSNR and SSIM gains over leading methods such as FSNet, ConvIR, OKNet, and GRL on nine datasets spanning defocus deblurring, desnowing, dehazing, and deraining.","feed_headline":"Use mid-range attention, not global, for image restoration","feed_subtitle":"A route-window transformer picks the closest relevant windows and tops nine restoration benchmarks.","key_machinery":"The load-bearing component is the Route-Windows Attention Module (RW AM). It averages query and key tensors within each window to form window description vectors, computes a regional similarity matrix between the central window's description and the descriptions of nearby candidate windows, applies a Softmax, and then uses TopK to keep only the $r_i$ most relevant windows. Absolute positions of those windows are recovered with a gather from an index table, relative position biases are gathered from a bias table, and the selected windows' keys and values are concatenated with the central window's before softmax attention. Two regional similarity shapes, cross-shaped and rectangle-shaped, are alternated across blocks to match stripe-like and block-like degradation patterns. A second component, Multi-Scale Structure Regularization (MSR), builds a loss term that gives each sub-scale a down-up-sampled clean target and asks the sub-scale to predict the structure residual, steering the sub-scales toward texture and structure learning.","core_discovery":"The central claim is that middle-range attention is sufficient for image restoration, and that a transformer built around dynamically selected nearby windows can outperform both fixed-window and global-attention models. The paper computes normalized average attention distance over degradation datasets and finds all values below 0.3, concluding that long-range modeling is not a practical necessity. RouteWinFormer realizes this by partitioning the feature map into non-overlapping windows, having a router score nearby candidate windows by regional similarity (cross-shaped and rectangle-shaped regions, alternated across blocks), selecting the top-$k$ most relevant windows, gathering their keys and values, and running multi-head attention over the merged set plus the central window. This cuts the attention cost from $O(h^2w^2c)$ to $O(r_i k^2 h w c)$. Training adds Multi-Scale Structure Regularization, which compares each sub-scale output with a down-up-sampled clean image and asks the sub-scale to predict the structure residual, so sub-scales learn textures and structure while the original scale learns degradation patterns. The paper reports that this design achieves the best published numbers on nine restoration benchmarks.","pith_inferences":["If the attention-distance measurement is repeated with a fixed protocol (one trained model, per-dataset statistics, disclosed normalization), the 0.3 threshold could become a design rule: restore with a receptive field covering roughly the top third of the image rather than the whole image.","The router's top-$k$ selection could be made input-adaptive in a stronger sense, for example by choosing more windows for images with large-scale degradation, which would test whether the middle-range finding holds across degradation scales.","The structure residual predicted by MSR is close to a high-frequency detail map; a natural extension is to condition the sub-scales explicitly on edge or gradient maps of the clean image rather than only through the loss.","The same routing idea could transfer to video restoration, where relevant context is likely found in nearby frames and neighboring spatial windows rather than across an entire frame."],"forward_implications":["Window-based restoration transformers can drop global attention without sacrificing accuracy, reducing memory and compute on high-resolution images.","Because window selection is content-adaptive, the effective receptive field can shrink or grow per region, unlike fixed shifted-window schemes.","The Multi-Scale Structure Regularization term is model-agnostic and can be added to any U-shaped restoration network to improve structure fidelity.","A single architecture with one set of hyperparameters transfers across defocus deblurring, desnowing, dehazing, and deraining, suggesting middle-range context is a shared property of these degradations.","The complexity reduction from $O(h^2w^2c)$ to $O(r_i k^2 h w c)$ makes mid-range attention practical for high-resolution restoration inputs."],"supporting_citations":[{"why":"Supplies the DPDD dataset and evaluation protocol for defocus deblurring, the primary benchmark in Table 1.","marker":"[Abuolaim and Brown, 2020]"},{"why":"Introduces the average attention distance measure that motivates the middle-range claim in Figure 1.","marker":"[Dosovitskiy et al., 2021]"},{"why":"Provides the Swin Transformer's shifted-window attention and relative position bias table that the blocks are built on.","marker":"[Liu et al., 2021b]"},{"why":"Establishes window-based SwinIR as a fixed-window attention baseline for restoration.","marker":"[Liang et al., 2021]"},{"why":"Offers the Restormer baseline with transposed global attention that RouteWinFormer is compared against.","marker":"[Zamir et al., 2022]"},{"why":"Supplies the global-attention GRL baseline and the anchored stripe attention approach that RouteWinFormer contrasts with.","marker":"[Li et al., 2023]"},{"why":"Provides NAFBlock, whose simplified channel attention and depthwise convolution form the feed-forward network.","marker":"[Chen et al., 2022a]"},{"why":"Provides FSNet, a leading baseline, and the training and evaluation conventions followed across the nine datasets.","marker":"[Cui et al., 2024b]"}],"fun_headline_variants":["Mid-range attention outperforms global on 9 restoration tasks","RouteWinFormer: pick relevant windows, beat SOTA on 9 benchmarks","For image restoration, middle-range attention is all you need","Dynamic route-window transformer leads 9 restoration benchmarks","Global attention overkill: mid-range transformer tops restoration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that middle-range attention is enough rests entirely on Figure 1's normalized average attention distance, which is reported without stating which network produced the attention maps, which images were used, or how normalization was applied; if that measurement is not representative of trained restoration models, the motivation for restricting attention to nearby windows is not established.","fun_headline_variants_meta":{"raw":{"variants":["Mid-range attention outperforms global on 9 restoration tasks","RouteWinFormer: pick relevant windows, beat SOTA on 9 benchmarks","For image restoration, middle-range attention is all you need","Dynamic route-window transformer leads 9 restoration benchmarks","Global attention overkill: mid-range transformer tops restoration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001311,"raw_usage":{"total_tokens":5335,"prompt_tokens":931,"completion_tokens":4404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":4321}},"tokens_in":547,"tokens_out":4404,"duration_ms":32281,"temperature":1.0,"reasoning_tokens":4321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:58:41.992177+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained global-attention restoration model, compute its average attention distance on each of the nine datasets with an explicit normalization rule, and compare with the 0.3 threshold; if any dataset exceeds 0.3, or if increasing RouteWinFormer's top-$k$ to include farther windows improves PSNR on images with large-scale degradation, the central claim is contradicted.","supporting_citations":[{"cited_title":"Defocus deblurring using dual-pixel data","cited_arxiv_id":null,"evidence_quote":"Supplies the DPDD dataset and evaluation protocol for defocus deblurring, the primary benchmark in Table 1."},{"cited_title":"An image is worth 16x16 words: Trans- formers for image recognition at scale","cited_arxiv_id":null,"evidence_quote":"Introduces the average attention distance measure that motivates the middle-range claim in Figure 1."},{"cited_title":"SwinIR: Image restoration using swin transformer","cited_arxiv_id":null,"evidence_quote":"Establishes window-based SwinIR as a fixed-window attention baseline for restoration."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Offers the Restormer baseline with transposed global attention that RouteWinFormer is compared against."}],"review_version":1}