{"id":"aa9a11be-ad7b-4f23-9cb6-97597c1bc600","arxiv_id":"2412.20066","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MaIR combines a stripe-based S-shaped scanning strategy and a sequence-shuffle attention block to improve Mamba-based image restoration, reporting new best PSNR on 14 benchmarks.","lead":"MaIR is a Mamba-based model for restoring degraded images that scans images in nested S-shaped stripes to keep local structure and spatial continuity, then blends the scan sequences with a lightweight attention block. It reports top results on super-resolution, denoising, deblurring, and dehazing benchmarks, though gains over prior models are small.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation margins in Tables 7–9 are 0.01–0.05 dB, so the central attribution of gains to NSS+SSA is not statistically secured; a control test swapping only the shuffle indexing is needed.","rationale":"The reader identified the absence of statistical validation as the weakest assumption, and the ablation tables confirm that the smallest margins (0.03 dB for NSS, 0.04 dB for shift-stripe) are precisely in the regime where seed-level variance in lightweight SR can exceed the reported difference. The paper's own Table 9 even shows flat PSNR across stripe widths 4–16, which is consistent with either robustness or insensitivity. A second issue is internal tension in the SOTA claim: Table 1 shows MambaIR ahead on several cells and Table 6 shows UVM-Net ahead on SOTS-Indoor, so 'state-of-the-art' is only true under a selective reading. None of this is fatal; the design is clear, the code is released, and the larger denoising/deblurring gains (0.21 dB on Urban100 denoising, 0.35–0.77 dB on deblurring) are more convincing. The appropriate verdict remains CONDITIONAL: require seed-variance reporting and reproducibility checks on the smallest ablation cells before treating the NSS/SSA causal claims as established.","tokens_in":18395,"tokens_out":1671,"duration_ms":15548,"concrete_test":"Run the released MaIR-Tiny lightweight SR (x2, Urban100) training recipe at least three times with different seeds, and also run the 'w/o NSS' (Z-shaped) and 'w/o SSA' (sequence-wise addition) configurations under identical schedules. If the mean gap between MaIR and w/o NSS falls below one standard deviation, the ablation attribution in Tables 7–8 is not statistically supported. As a secondary check, verify whether the reported Table 1 values (e.g., MaIR x4 Set5 32.93, MambaIR 33.03) are reproducible from the released checkpoints.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that the NSS scanning strategy and SSA aggregation drive the reported improvements. The strongest direct evidence is the ablation study in Tables 7–8, but the margins are extremely small: replacing NSS with a plain Z-shaped scan costs only 0.03 dB PSNR on Urban100 (32.97 vs 32.94), removing shift-stripe costs 0.04 dB (32.97 vs 32.93), and replacing SSA with simple sequence-wise addition costs 0.07 dB (32.97 vs 32.90). These differences are close to typical run-to-run training variance for a lightweight SR model, and the paper reports no repeated seeds, error bars, or significance tests anywhere. The shift-stripe mechanism is described as a way to preserve boundary locality, yet Table 9 shows stripe widths 4, 8, and 16 all give identical PSNR 32.97 and SSIM differences of only 0.0004; if the mechanism were load-bearing, one would expect a clearer sensitivity. A second gap is that the claimed 'SOTA' is contradicted on some entries of the paper's own tables (e.g., MaIR at x4 on Set5 is 32.93 vs 33.03 for MambaIR, and UVM-Net beats MaIR on SOTS-Indoor PSNR 40.17 vs 39.45), so the strongest claim rests on selective best entries rather than a systematic margin. The concern is therefore not about the design being implausible, but about whether the empirical evidence as presented can rule out training noise as the cause of the small ablation and cross-method differences. Because the code is released, this is directly testable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MaIR, a Mamba-based image restoration network whose two main contributions are a Nested S-shaped Scanning strategy (NSS) and a Sequence Shuffle Attention block (SSA). NSS divides feature maps into stripes and scans them in an S-shaped path, with a shift-stripe mechanism across successive modules; SSA aggregates the four scanned sequences by computing channel-wise attention weights after a shuffle operation. The method is evaluated on super-resolution, denoising, deblurring, and dehazing, with comparisons to about forty baselines across fourteen datasets, and the paper claims state-of-the-art performance on all four tasks. The architecture is described clearly and the code is released.","tokens_in":18755,"tokens_out":5843,"duration_ms":59500,"significance":"If the claimed gains are robust, NSS and SSA are simple and largely architecture-agnostic modifications that could benefit Mamba-based restoration beyond this specific network. The paper's strengths are the breadth of the evaluation (four tasks, fourteen datasets), the clear description of the scanning and aggregation designs, and the public code release, which makes the results directly checkable. However, the empirical case is weakened by the absence of any uncertainty quantification: many reported improvements over strong baselines are just a few hundredths of a dB, and the ablation margins used to attribute the gains to NSS and SSA are of the same size as typical run-to-run variation. The significance therefore depends on whether the small margins are systematic rather than noise.","major_comments":[{"comment":"The claim that MaIR 'surpasses 40 baselines' and achieves state-of-the-art performance on all four tasks is not supported by the paper's own tables. In Table 1 (×4, Set5), MambaIR reports 33.03 dB PSNR while MaIR reports 32.93 dB; in Table 1 (×2, Set5), MambaIR reports 38.57 dB versus MaIR's 38.56 dB; and in Table 6 (SOTS-Indoor), UVM-Net reports 40.17 dB versus MaIR's 39.45 dB. The abstract and conclusion should be revised to state precisely where MaIR is best and to acknowledge the dataset-specific exceptions, rather than claiming a global state-of-the-art. The same issue appears in the intro's contribution bullet, which says MaIR 'obtains state-of-the-art performance on four tasks across 14 benchmarks.'","section":"Abstract; Tables 1 and 6"},{"comment":"The central attribution of the performance gain to NSS and SSA rests on ablation margins of 0.01–0.05 dB PSNR. Table 7 shows 32.97 vs 32.94 when replacing NSS with a Z-shaped scan, and 32.97 vs 32.93 when removing the shift-stripe mechanism; Table 8 shows 32.97 vs 32.90 when SSA is replaced by simple sequence-wise addition; Table 9 shows identical PSNR values of 32.97 for stripe widths 4, 8, and 16. No repeated seeds, error bars, confidence intervals, or significance tests are reported anywhere in the paper. Because these margins are comparable to typical run-to-run variance for lightweight super-resolution models, the current evidence does not rule out training noise as the source of the reported differences. Please provide multi-seed results with mean±std, paired significance tests, or a control experiment that changes only the shuffle indexing while preserving the permutation class, to demonstrate that the observed margins are systematic.","section":"§4.5, Tables 7–9"},{"comment":"The paper repeatedly describes NSS as 'cost-free' and concludes that MaIR improves image quality 'without introducing extra computations,' but this is contradicted by the reported MACs. Table 2 shows MaIR-Tiny using 207.8G MACs at ×2 versus 167.1G for MambaIR-Tiny, while MaIR-Small uses 542.0G versus 567.5G for MambaIR-Small. The efficiency claim should be clarified: the stripe reordering itself may be rearrangement-only, but SSA adds pooling, group convolution, and weighted summation, and the actual overhead should be reported for each configuration. As written, the 'cost-free' claim is misleading.","section":"§2, §3.2, Conclusion; Table 2"}],"minor_comments":[{"comment":"'shift-stipe mechanism' should be 'shift-stripe mechanism.'","section":"§3.2"},{"comment":"In the sentence listing CNN-based deblurring methods, the citation '[39, 39, 48, 60]' contains a duplicate entry; one occurrence of [39] should be removed.","section":"§2.1"},{"comment":"The experimental settings are deferred entirely to the supplementary materials. To make the main paper self-contained, please include the essential training details (number of RMGs and RMBs, patch size, optimizer, iterations, learning rate schedule, and hardware) in the main text or an appendix.","section":"§4"},{"comment":"The diagrams for NSS and SSA are difficult to read at the printed size; the stripe boundaries, shuffle order, and group-convolution labels should be enlarged and clarified.","section":"Figures 3 and 4"},{"comment":"The Charbonnier loss is typeset as 'L = p ∥y − y′∥2 + ϵ2'; the square-root symbol appears to be missing. Please correct the equation.","section":"§3.1, Eq. (2)"},{"comment":"The phrase 'surpasses 40 baselines across 14 challenging datasets' should specify whether the comparison is per-dataset best or an aggregate, since the paper's own tables contain non-winning entries.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical architecture study, and the code release is a genuine strength. The main risk is that the headline claims (state-of-the-art, cost-free, and the attribution of gains to NSS/SSA) are stated more strongly than the evidence supports. Given the small margins, the paper needs either additional experiments with uncertainty quantification or a substantially more cautious framing. This is fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: MaIR is a genuinely useful architecture paper for Mamba-based restoration. The NSS scanning strategy—stripe partitioning plus S-shaped paths with a shift between blocks—is a sensible way to preserve both locality and continuity, and SSA is a simple, cheap aggregation mechanism. The experiments are extensive: four tasks, 14 benchmarks, code released. If you work in this area, you should know this paper.\n\nWhere it falls short: the \"state-of-the-art\" claim in the abstract is not supported by the paper's own tables. On Set5 x4, MaIR gets 32.93 vs MambaIR's 33.03; on SOTS-Indoor, UVM-Net beats it by 0.72 dB. The paper says it surpasses 40 baselines, but the tables show it is often second-best. That overclaim should be fixed.\n\nThe bigger soft spot is statistical. Ablation margins are tiny: replacing NSS with a Z-scan costs 0.03 dB, removing the shift-stripe costs 0.04 dB, and stripe widths from 4 to 16 give identical PSNR. Without error bars or multiple seeds, these differences sit within typical run-to-run noise for lightweight SR. The stress-test note is right that a control test swapping only the shuffle indexing would help. This is a real weakness, but it is addressable: run several seeds, report variance, and soften the wording.\n\nAlso, MaIR+ appears in the SR tables but is never defined in the paper. That is a missing reference that needs clarification.\n\nAll that said, the core design is plausible, the empirical study is broad, and the qualitative results look consistent with the claimed mechanism. This is a solid paper that deserves a serious referee, but it needs revision before acceptance: define MaIR+, report variance, and align the abstract with the actual numbers.","headline":"A solid Mamba-restoration paper with a plausible scanning design and broad experiments, but the SOTA claim is overstrong and the ablation margins need variance reporting.","tokens_in":19274,"tokens_out":1869,"would_cite":true,"duration_ms":18812,"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":"MaIR claims a Mamba restoration network can preserve both locality and continuity during scanning, and reports the best results across 40 baselines on 14 datasets.","keywords":["Mamba","selective state space models","image restoration","super-resolution","denoising","deblurring","dehazing","scanning strategy"],"falsifier":"Run MaIR and the key ablations (a plain Z-shaped scan in place of NSS, and simple summation or gating in place of SSA) several times with different random seeds on Urban100 ×2 lightweight super-resolution. If the reported 0.03–0.12 dB margins fall inside the seed-to-seed variance, the central claim that NSS and SSA cause the improvement is not supported.","tokens_in":18198,"feed_emoji":"🖼️","tokens_out":5982,"duration_ms":55830,"temperature":0.7,"pith_summary":"This paper argues that Mamba-based image restoration fails not because of the selective scan itself but because of how images are flattened into 1D sequences and how those sequences are recombined. The authors propose MaIR, which scans the image in nested S-shaped stripes to keep nearby pixels close in the sequence while preserving the smooth left-to-right, top-to-bottom order of natural images, and then merges the four scan directions with channel-wise attention instead of simple addition. Trained on standard benchmarks, MaIR reports the best PSNR/SSIM among 40 baselines on super-resolution, denoising, deblurring, and dehazing across 14 datasets. If the result holds, scan-path design and sequence aggregation are cheap, transferable levers for state-space image models.","feed_headline":"Nested S-shaped scan lifts Mamba above 40 restoration baselines","feed_subtitle":"Scanning in nested S-shaped stripes keeps local texture and spatial flow, and channel shuffling merges the passes.","key_machinery":"The two load-bearing components are Nested S-shaped Scanning (NSS) and Sequence Shuffle Attention (SSA). NSS partitions the feature map into non-overlapping stripes and follows an S-shaped path both inside each stripe and across stripes, so a scan can be long-range while keeping neighboring pixels adjacent; the shift-stripe mechanism moves the stripe boundaries between consecutive modules so boundary pixels are not permanently separated. SSA takes the four SSO-processed sequences, spatially pools them, reshuffles the pooled channels so that the k-th channel of all four directions sits together, applies group convolution to produce attention weights per channel, unshuffles the weights, and uses them in a weighted sum of the four sequences. Together they convert the selective scan's 1D processing into a 2D-aware restore-then-fuse operation with no added scanning cost.","core_discovery":"The paper's central claim is that the dominant Mamba-restoration pipeline — flatten a 2D feature map into rows and columns, run selective scans independently, then add the results — throws away exactly the structure that restoration needs. MaIR replaces the flattening with NSS: the feature map is cut into non-overlapping stripes, each stripe is traversed in an S-shaped path, and the paths connect across stripes, so pixels that are near in the image stay near in the 1D sequence even at stripe boundaries; a shift-stripe variant between successive modules moves the boundaries so no region is permanently split. It replaces the additive recombination with SSA, which pools each of the four directional sequences, shuffles the channels so corresponding channels from all directions are adjacent, computes per-channel attention weights with group convolution, and sums the four sequences using those weights. On 14 benchmark datasets covering super-resolution, denoising, deblurring, and dehazing, the paper reports that MaIR exceeds 40 baselines, with the largest margins on dehazing and consistent gains such as 0.03–0.12 dB over MambaIR on classic super-resolution. The message is that scan topology and sequence aggregation, not extra capacity, are what make a Mamba restoration model competitive.","pith_inferences":["Because NSS and SSA add no trainable parameters to the scanning itself, the design should transfer to other Mamba vision tasks, such as low-light enhancement or compression artifact removal, where locality and continuity also matter.","SSA is effectively a learned mixture over scan directions; extending it to more than four directions, or to multi-scale stripes, is a natural next test that the paper does not run.","The paper's own ablations show the PSNR margin over a plain Z-shaped scan is small, so a perceptual study on high-texture regions would be a sharper test of the locality-and-continuity claim than aggregate scores."],"forward_implications":["A Mamba restoration model can capture long-range dependencies without sacrificing 2D structure, so the near-linear complexity of the selective scan is retained.","The shift-stripe mechanism means stripe boundaries are handled by alternating module partitions, so seam artifacts should not accumulate in deep stacks.","SSA gives a learned, per-channel weight to each of the four scan directions, allowing the network to rely more on the direction that carries the most relevant context.","The same architecture, without task-specific changes beyond the loss function, reaches the top reported numbers on super-resolution, denoising, deblurring, and dehazing."],"supporting_citations":[{"why":"Supplies the selective scan operation (SSO) that MaIR's modules wrap around.","marker":"[15]"},{"why":"Introduces the row/column cross-scan flattening of images into 1D sequences that MaIR replaces with NSS.","marker":"[31]"},{"why":"The main Mamba restoration baseline and the architecture whose training setup MaIR follows.","marker":"[18]"},{"why":"Shows windowed scanning that preserves locality; used as a comparison in the NSS ablation.","marker":"[21]"},{"why":"Shows an S-shaped scanning path that preserves continuity; used as a comparison in the NSS ablation.","marker":"[19]"},{"why":"The standard SwinIR restoration baseline and training protocol that anchors the SR, denoising, and deblurring comparisons.","marker":"[27]"},{"why":"Uses an SSM-based aggregation of sequences; the SSA ablation compares against it.","marker":"[67]"},{"why":"The Restormer transformer baseline whose denoising and deblurring numbers MaIR is compared against.","marker":"[53]"}],"fun_headline_variants":["MaIR: nested S-scans + shuffle attention beat 40 baselines","S-shaped stripes keep local flow, shuffle attention merges scans","MaIR's NSS and SSA outdo 40 methods on restoration tasks","Rethink Mamba scans: MaIR's S-paths and shuffled channels win","Nested S-scanning and sequence shuffle lift Mamba to SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results rest on the assumption that differences of roughly 0.03 to 0.12 dB in PSNR against strong baselines, reported without error bars or repeated runs, reflect a genuine improvement of the proposed scanning and aggregation rather than training noise.","fun_headline_variants_meta":{"raw":{"variants":["MaIR: nested S-scans + shuffle attention beat 40 baselines","S-shaped stripes keep local flow, shuffle attention merges scans","MaIR's NSS and SSA outdo 40 methods on restoration tasks","Rethink Mamba scans: MaIR's S-paths and shuffled channels win","Nested S-scanning and sequence shuffle lift Mamba to SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000633,"raw_usage":{"total_tokens":2987,"prompt_tokens":1073,"completion_tokens":1914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":689,"completion_tokens_details":{"reasoning_tokens":1813}},"tokens_in":689,"tokens_out":1914,"duration_ms":13228,"temperature":1.0,"reasoning_tokens":1813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:35:17.091834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MaIR and the key ablations (a plain Z-shaped scan in place of NSS, and simple summation or gating in place of SSA) several times with different random seeds on Urban100 ×2 lightweight super-resolution. If the reported 0.03–0.12 dB margins fall inside the seed-to-seed variance, the central claim that NSS and SSA cause the improvement is not supported.","supporting_citations":[{"cited_title":"SwinIR: Image Restoration Using Swin Transformer","cited_arxiv_id":null,"evidence_quote":"The standard SwinIR restoration baseline and training protocol that anchors the SR, denoising, and deblurring comparisons."},{"cited_title":"Restormer: Efficient Transformer for High-Resolution Im- age Restoration","cited_arxiv_id":null,"evidence_quote":"The Restormer transformer baseline whose denoising and deblurring numbers MaIR is compared against."}],"review_version":1}