{"id":"2685abf8-fb3a-4aa5-908c-b77472cfe84e","arxiv_id":"2412.10338","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An SSM deblurring network that scans image features in interleaved intra-slice and inter-slice routes achieves state-of-the-art perceptual metrics on GoPro, HIDE, RWBI, and RealBlur datasets.","lead":"XYScanNet replaces the flatten-and-scan step in Mamba-based image deblurring with row and column slice scanning, improving perceptual quality at lower compute. It reports large KID, FID, and LPIPS gains over prior deblurring networks while keeping PSNR competitive.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table 1 caption attributes the perceptual gains to the loss function, and no ablation isolates the VSSM; the perceptual state-of-the-art claim is therefore not yet supported.","rationale":"The reader's weakest assumption identifies the most load-bearing concern: the perceptual gains in Table 1 are not attributable to the slice-and-scan VSSM because the training losses are not controlled across methods. The paper's own Table 1 caption makes this attribution explicitly, saying performance 'can be attributed to the loss function discussed in Sec. 4.1.' Section 4.1 also states that additional studies on training strategies are beyond the paper's scope, and the Limitation section repeats that the relationship between loss functions and performance needs further discussion. No ablation removes the VGG feature loss Lp, so the architecture's independent contribution to KID/FID/LPIPS is untested. The matched VSSM comparison in Table 7 is a useful controlled experiment, but its perceptual results are mixed rather than uniformly SOTA, so it cannot rescue the headline perceptual claim. The efficiency and memory claims are not affected by this concern. The appropriate response is to keep the reader's CONDITIONAL verdict, with the missing loss-controlled ablation as the explicit condition; the concern does not by itself warrant rejection because the architecture may still be effective and the efficiency evidence is credible.","tokens_in":16724,"tokens_out":4330,"duration_ms":38189,"concrete_test":"Retrain the full XYScanNet on GoPro with L = Lchar + 0.05Ledge only (dropping the VGG loss Lp) and recompute Table 1's KID/FID/LPIPS on GoPro; simultaneously retrain at least one strong baseline, e.g. FFTformer or LoFormer, with the full loss L. If XYScanNet without Lp retains KID approximately 0.073, FID approximately 0.138, and LPIPS approximately 0.067, the architecture-level claim is supported. If its scores regress toward the previous methods, or if the baselines with Lp jump to comparable scores, the perceptual SOTA claim rests on the loss function, not the VSSM.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central perceptual claim depends on Table 1, where XYScanNet's KID/FID/LPIPS beat methods that were trained with different losses. Section 4.1 defines L = Lchar + 0.05Ledge + 0.0005Lp, and Table 1's caption states: 'This performance can be attributed to the loss function discussed in Sec. 4.1.' No ablation removes Lp from XYScanNet, and no comparison baseline is retrained with the same Lp. The Limitation section concedes that 'further discussions are needed to explore the relationship between loss functions and performance.' Therefore the reported perceptual superiority could be a property of the training objective rather than of the slice-and-scan VSSM. The matched VSSM comparison in Table 7 (same framework and training strategy) partially supports the efficiency claim and shows mixed perceptual changes (KID worsens from 0.121 to 0.129; FID improves from 0.232 to 0.210), so it does not establish the headline perceptual SOTA. Equations (1)-(3) are internally coherent; the defect is empirical attribution, not mathematical inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes XYScanNet, a U-Net for single-image deblurring built on a new 'slice-and-scan' state-space module (VSSM) that alternates Intra-Scanners (pixel-level horizontal/vertical scans within image rows/columns) and Inter-Scanners (global-average-pooled slice-level scans) to avoid the spatial misalignment of flatten-and-scan strategies. It also introduces a dual-gating feature fusion (DGFF) module. Experiments on GoPro, HIDE, RWBI, and RealBlur report competitive PSNR/SSIM, improved perceptual metrics (KID/FID/LPIPS/NIQE/Q-ALIGN), and substantially lower training time and memory than MambaIR's VSSM in a matched framework (Table 7). The central claims are a new state-of-the-art in perceptual deblurring quality and a more efficient Mamba restoration block.","tokens_in":16917,"tokens_out":6668,"duration_ms":54784,"significance":"If the architectural claims hold, the slice-and-scan design is a useful contribution to the Mamba-based restoration literature: it provides a plausible mechanism to preserve local spatial structure while capturing global blur, and the matched VSSM comparison in Table 7 is a strong, internally controlled piece of evidence for the efficiency gains (56.61% training-time reduction, 46.18% inference-memory reduction). The paper also releases code and reports a wider range of perceptual metrics than PSNR alone, which is commendable. However, the headline perceptual superiority is not yet attributable to the architecture because the loss function is not controlled across methods and no ablation isolates the perceptual loss; the paper's own caption and limitation paragraph concede this confound. The efficiency and architectural-validity claims are substantially better supported than the perceptual-SOTA claim.","major_comments":[{"comment":"The headline claim that XYScanNet achieves state-of-the-art perceptual quality is confounded by the training loss. The loss in Eq. (5) includes a VGG perceptual term Lp with λ2=0.0005, and the Table 1 caption itself states that the performance 'can be attributed to the loss function discussed in Sec. 4.1.' None of the compared methods in Tables 1–4 is retrained with this loss, and no ablation removes Lp from XYScanNet. Consequently the KID/FID/LPIPS improvements could be caused by the objective rather than by the slice-and-scan VSSM. The matched VSSM comparison in Table 7, which holds the framework and training strategy fixed, actually shows KID worsening from 0.121 to 0.129 while FID improves from 0.232 to 0.210, so it does not establish perceptual superiority. The Limitation section's admission that 'further discussions are needed to explore the relationship between loss functions and performance' confirms this gap. Please add an ablation that trains XYScanNet without Lp and, ideally, retrains at least the nearest competitor with the same loss.","section":"Sec. 4.1, Eq. (5); Tables 1–4; Limitation paragraph"},{"comment":"The text states that the interleaved intra/inter configuration 'improves PSNR on GoPro by 0.2 dB over the inter-only method,' but Table 5 reports inter-only PSNR 32.18 dB and interleaved 32.11 dB, i.e., a decrease of 0.07 dB. The 0.20 dB improvement is over the intra-only row (31.91 dB), not the inter-only row. The same paragraph's claim of a 20.39% training-time reduction and a 9.52% memory reduction 'compared to the intra-only approach' is also incorrect: those percentages are relative to the inter-only row (103 s to 82 s; 12.82 GB to 11.60 GB), while the intra-only row has lower time and memory (65 s; 10.36 GB). These baseline errors invert the direction of the efficiency/quality trade-off in the ablation and must be corrected.","section":"Sec. 4.3, Table 5"},{"comment":"The supplementary states that each network was trained and tested four times to minimize the effects of randomness, yet no table reports variance or significance tests. Several of the reported advantages are tiny (e.g., Table 2 NIQE 3.383 vs. 3.384 for LoFormer; Table 3 LPIPS 0.091 vs. 0.093) and may fall within run-to-run noise. Please report mean ± std (or confidence intervals) for at least the key perceptual metrics on GoPro, HIDE, and RWBI, and clarify how many seeds underlie the numbers in Tables 1–4. Without this, the 'state-of-the-art perceptual quality' claim is not robustly supported.","section":"Supplementary Sec. 7.1; Tables 1–7"}],"minor_comments":[{"comment":"The sentence 'The horizontal Intra-Scanner is constructed symmetrically' should read 'Inter-Scanner-H' rather than 'Intra-Scanner'; the surrounding text is describing the symmetric horizontal counterpart of the Inter-Scanner.","section":"Sec. 3.2, Inter-Scanner paragraph"},{"comment":"The sentence 'As shown in Tab. 7, our slice-and-scan-based VSSM.' is an incomplete fragment; please complete the sentence or remove it.","section":"Supplementary Sec. 6.2"},{"comment":"The KL-divergence terms in Eq. (4) require non-negative, normalized inputs, but X'_cur and X'_oth are outputs of convolutions that may be negative or unnormalized; please specify how the feature maps are converted to probability distributions (e.g., softmax or abs + normalization) and clarify the intended direction of the asymmetric KL term.","section":"Eq. (4) and surrounding text"},{"comment":"The phrase 'LPIPS st by 0.041 on RealBlur-J' contains a typo ('st') and the reported reduction is 0.041 in LPIPS; please correct the wording so the metric and its improvement are unambiguous.","section":"Sec. 4.3, first paragraph; Table 6"}],"recommendation":"major_revision","confidential_remarks":"This is a solid engineering contribution with a well-designed matched ablation (Table 7) for the efficiency claims, but the headline perceptual-SOTA claim is currently overreach relative to the evidence. The authors' own Table 1 caption attributes the perceptual gains to the loss function, and the Limitation section concedes the loss-performance relationship is unresolved; requiring an Lp ablation and a clearer attribution analysis is essential before acceptance. The Table 5 baseline error and the absence of error bars are fixable but must be addressed. I see no grounds for rejection, but the revision should be substantive rather than cosmetic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper: the slice-and-scan VSSM is a legitimate, useful architecture idea, and the efficiency comparison against MambaIR's VSSM is solid. But the headline claim of state-of-the-art perceptual quality is not supported by the experiments as reported, and the paper itself says so in the Table 1 caption and the Limitation section.\n\nWhat's new: the interleaved Intra- and Inter-Scanners are a sensible fix for the spatial misalignment and local pixel forgetting problems in flatten-and-scan Mamba. The inter-slice GAP-compressed scan is a lightweight way to capture global blur patterns. The equations are coherent, the design is clearly explained, and the DGFF module, while not deeply motivated by the KL gating, shows gains in the ablation. The matched VSSM comparison in Table 7 is the most valuable part: same framework, same training strategy, similar size, and it shows a 56.6% training time reduction and 46.2% inference memory savings over MambaIR's VSSM. That is a real contribution.\n\nThe soft spot is the perceptual claim. Table 1 shows XYScanNet beating LoFormer and others on KID/FID/LPIPS, but the caption attributes the performance to the loss function in Sec. 4.1, which includes a VGG feature loss with weight 0.0005. No ablation removes that loss from XYScanNet, and no baseline is retrained with the same loss. So we cannot tell whether the perceptual gains come from the architecture or simply from the extra Lp term. Table 7 does not rescue the claim: on GoPro, KID actually worsens (0.121 to 0.129) while FID improves. The conclusion's phrase \"state-of-the-art perceptual quality\" is therefore not backed by the evidence. This is a fixable problem—run the loss ablation, retrain at least one strong baseline with the same loss—but as it stands the main quantitative selling point is partly an artifact of training choices.\n\nOther, minor issues: the paper says each network was trained and tested four times, but no variance or significance testing is reported; the RealBlur comparison relies on Q-ALIGN and NIQE, which are no-reference and have their own quirks; and the Limitation section honestly concedes the local-global balance and loss-function questions, which I appreciate. The citation pattern looks fine; no red flags.\n\nWho is this for? Someone working on Mamba-based image restoration. The efficiency analysis and the slice-and-scan idea are worth engaging with. It deserves a serious referee: the architecture is plausible, the efficiency result is concrete, and the main flaw is an empirical gap that can be closed with an ablation. I would not desk-reject it, but I would insist on the loss ablation and variance reporting before acceptance.","headline":"Useful Mamba deblurring architecture with big efficiency gains, but the headline perceptual SOTA claim is undercut by the paper's own admission that the loss drives it.","tokens_in":17505,"tokens_out":1763,"would_cite":true,"duration_ms":18292,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A slice-and-scan state-space model for image deblurring beats prior perceptual scores and cuts training time by more than half.","keywords":["single image deblurring","state space models","Mamba","slice-and-scan","perceptual image quality","vision state space module","feature fusion","image restoration"],"falsifier":"Retrain XYScanNet with the same loss but with its Intra/Inter scanners replaced by MambaIR's flatten-and-scan VSSM, or retrain it without the VGG feature-distance term (λ2Lp); if KID, FID, and LPIPS no longer beat the baselines, the perceptual claim would lose its support. The efficiency claim could also be checked by profiling both VSSMs on identical hardware with identical framework settings.","tokens_in":16474,"feed_emoji":"🖼️","tokens_out":5915,"duration_ms":47546,"temperature":0.7,"pith_summary":"The paper claims that the standard flatten-and-scan strategy used by Mamba-based image restoration models corrupts spatial layout by placing distant pixels next to each other, and that this hurts deblurring sharpness. It proposes a slice-and-scan alternative that scans along image rows and columns, with lightweight inter-slice scans for global blur, and packages this in a new Vision State Space Module. The resulting network, XYScanNet, reports state-of-the-art perceptual quality (KID 0.073, FID 0.138, LPIPS 0.067 on GoPro) while keeping PSNR and SSIM competitive, and it cuts training time by 56.61% and inference memory by 46.18% relative to MambaIR's VSSM in the same framework. A sympathetic reader would care because it offers a concrete architectural reason—spatial misalignment—for why prior Mamba deblurring underperforms, and a cheaper way to fix it.","feed_headline":"Slice-and-scan deblurring tops perceptual metrics, cuts training 56%","feed_subtitle":"XYScanNet scans rows then columns instead of flattening pixels, keeping local order and improving KID, FID, and LPIPS.","key_machinery":"The central object is the slice-and-scan strategy inside a Vision State Space Module (VSSM). Instead of flattening a feature map into one long 1D sequence, Intra-Scanner-H slices along the height dimension so each row is scanned separately, and Intra-Scanner-V does the same along width; Inter-Scanners pool the orthogonal dimension (compression factor δ) and use the scanned output as a sigmoid gate over the original features. The VSSM applies layer norm, pixel-wise and depth-wise convolutions, splits channels into vertical/horizontal branches, and combines the dual-scanner output with a SiLU-activated path via element-wise multiplication and a residual addition. The interleaved placement of Intra- and Inter-VSSMs is what lets the network handle both local detail and large-area blur, and the DGFF module uses KL-divergence-gated feature fusion across levels.","core_discovery":"On its own terms, the paper establishes that replacing flatten-and-scan with slice-and-scan removes the spatial misalignment artifact in Mamba-based deblurring. Intra-Scanners preserve local pixel adjacency within each row or column, while Inter-Scanners compress the orthogonal dimension and use a gating signal to capture cross-slice dependencies, so the network estimates both fine local blur and large-area blur without quadrupling scan routes. The paper reports that XYScanNet achieves the best perceptual scores among compared methods on GoPro (KID 0.073, FID 0.138, LPIPS 0.067, NIQE 4.05) with PSNR 33.91 dB and SSIM 0.968, and that in an identical framework its VSSM outperforms MambaIR's VSSM while reducing training time by 56.61% and inference memory by 46.18%. The authors also claim the interleaved Intra/Inter design is validated by ablations and that their dual gating feature fusion module halves the cost of the asymmetric feature fusion baseline.","pith_inferences":["The perceptual gains are partly attributed by the paper itself to the loss function (Charbonnier plus edge plus VGG feature distance), so an ablation that removes the VGG term from XYScanNet's loss would separate the architecture's contribution from the loss's contribution; no such ablation is reported.","If slice-and-scan is the active ingredient, the same Intra/Inter scanner design should improve other Mamba-based restoration tasks such as denoising and super-resolution, where flatten-and-scan also introduces spatial misalignment.","The inter-slice compression via global average pooling suggests a general recipe for global context in SSMs: scan a cheap pooled projection and use it as a gate, rather than scanning every pixel in four directions.","A controlled comparison that trains MambaIR's VSSM and XYScanNet's VSSM with identical losses (including the VGG term) would tell whether the efficiency and quality advantages persist under equal training conditions; the paper only compares with the same framework and training strategy but not with the same loss."],"forward_implications":["On GoPro, XYScanNet reports the lowest KID (0.073), FID (0.138), and LPIPS (0.067) among the compared methods, with PSNR 33.91 dB and SSIM 0.968.","The GoPro-trained model transfers to HIDE and RWBI, improving no-reference metrics (e.g., NIQE reduction of 0.163 on RWBI) and competitive PSNR on HIDE.","Replacing half of the Intra-Scanners with Inter-Scanners cuts training time by 20.39% and memory by 9.52% with only a 0.02 dB PSNR drop on GoPro, supporting the interleaved design.","In the same baseline framework, the new VSSM beats MambaIR's VSSM on perceptual and many distortion metrics while reducing training time by 56.61% and inference memory by 46.18%.","The DGFF fusion module halves FLOPs (by 49.3%) and parameters (by 53.11%) relative to AFF while improving LPIPS on RealBlur-J."],"supporting_citations":[{"why":"Supplies the S6 selective-scan state-space layer that both Intra- and Inter-Scanners call.","marker":"[13]"},{"why":"Defines the prior Vision State Space Module and flatten-and-scan baseline; the efficiency and quality comparison in Table 7 is against it.","marker":"[17]"},{"why":"Strip-based dual-branch design that motivates the horizontal/vertical intra-slice scanning for motion blur.","marker":"[48]"},{"why":"The training set and primary benchmark whose PSNR, SSIM, KID, FID, and LPIPS results anchor the main comparison.","marker":"[35]"},{"why":"Baseline asymmetric feature fusion that DGFF is compared and measured against for FLOPs and parameters.","marker":"[6]"},{"why":"Strong Transformer deblurring baseline whose perceptual and distortion scores XYScanNet is compared with.","marker":"[22]"},{"why":"Recent perceptual-competitive baseline that XYScanNet beats on KID, FID, and LPIPS.","marker":"[32]"}],"fun_headline_variants":["Slice-and-scan deblurring beats flatten-scan on perceptual metrics","XYScanNet's row-col scans cut training time by 56%","New scanning strategy fixes misalignment in Mamba deblurring","Deblurring with slice-and-scan: 17% better KID, 56% less training","XYScanNet: state-space deblurring with efficient slice-and-scan"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the new slice-and-scan module, not the extra VGG-based perceptual loss added during training, is what produces the reported perceptual improvements over methods trained without that loss.","fun_headline_variants_meta":{"raw":{"variants":["Slice-and-scan deblurring beats flatten-scan on perceptual metrics","XYScanNet's row-col scans cut training time by 56%","New scanning strategy fixes misalignment in Mamba deblurring","Deblurring with slice-and-scan: 17% better KID, 56% less training","XYScanNet: state-space deblurring with efficient slice-and-scan"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000914,"raw_usage":{"total_tokens":3939,"prompt_tokens":970,"completion_tokens":2969,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":2863}},"tokens_in":586,"tokens_out":2969,"duration_ms":20086,"temperature":1.0,"reasoning_tokens":2863,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:56:52.436999+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain XYScanNet with the same loss but with its Intra/Inter scanners replaced by MambaIR's flatten-and-scan VSSM, or retrain it without the VGG feature-distance term (λ2Lp); if KID, FID, and LPIPS no longer beat the baselines, the perceptual claim would lose its support. The efficiency claim could also be checked by profiling both VSSMs on identical hardware with identical framework settings.","supporting_citations":[{"cited_title":"Stripformer: Strip transformer for fast image deblurring","cited_arxiv_id":null,"evidence_quote":"Strip-based dual-branch design that motivates the horizontal/vertical intra-slice scanning for motion blur."},{"cited_title":"Efficient frequency domain-based trans- formers for high-quality image deblurring","cited_arxiv_id":null,"evidence_quote":"Strong Transformer deblurring baseline whose perceptual and distortion scores XYScanNet is compared with."},{"cited_title":"Loformer: Local frequency transformer for im- age deblurring","cited_arxiv_id":null,"evidence_quote":"Recent perceptual-competitive baseline that XYScanNet beats on KID, FID, and LPIPS."}],"review_version":1}