{"id":"b0fbe2e9-df82-4776-a64f-9eabc5b6a8ab","arxiv_id":"2411.17214","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A transformer super-resolution model using dilated multi-range attention and a multi-scale star module reaches state-of-the-art accuracy with lower computation than prior efficient SR models.","lead":"The paper introduces MAT, an image super-resolution model that mixes attention over small, medium, and large pixel neighborhoods using a cheap dilation trick. It reports higher sharpness scores and roughly 3.3 times faster upscaling than the popular SRFormer-light on standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table VII contradicts the SOTA claim: MambaIR reports 33.32 dB on Manga109 ×4 vs MAT+ 32.49 dB, though Section D claims superiority on all five datasets.","rationale":"The reader's weakest assumption concerned the sparse sampling pattern of SMA. That assumption is actually supported by the paper's own evidence: Table II shows a monotonic improvement as dilation increases from {1,1,1} to {5,7,9} (32.94 to 33.22 dB), and LAM/DI analyses (Figs. 6, 14) show a larger diffusion index at δ=Maximum. Thus, while the generalization of the dilation schedule to other patch sizes is untested, it is not the weakest link. The more damaging issue is the internal contradiction in the main classical-SR comparison. Table VII reports MambaIR at 33.32 dB on Manga109 ×4, 0.83 dB above MAT+ (32.49 dB), while Section D claims superiority across all datasets and scales. Because the paper's stated contribution is state-of-the-art performance with efficiency, this direct counterexample in the authors' own table is a correctness risk that cannot be resolved by more seeds or code re-runs; it requires either a correction of the table value or a qualification of the claim. If the table value is a typo, the paper can be accepted after correction; if not, the central claim is overstated. This is therefore more load-bearing than the reproducibility concern, though the call for code verification remains appropriate.","tokens_in":21083,"tokens_out":7434,"duration_ms":66278,"concrete_test":"Cross-check the MambaIR ×4 Manga109 PSNR=33.32 dB against the original MambaIR paper (Guo et al., ECCV 2024) and its official results. If the correct value is ~32.3 dB, the table contains a typo and should be corrected; if 33.32 dB is confirmed, revise the 'superior performance across all five benchmark datasets and scale factors' claim to acknowledge MambaIR's higher PSNR on this entry, e.g., by restricting the claim to parameter-matched comparison.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that MAT achieves state-of-the-art performance in classical SR is directly contradicted by the paper's own Table VII. In the ×4 Manga109 row, MambaIR (20.57M params) scores 33.32 dB PSNR while MAT+ (self-ensemble, 9.74M params) scores 32.49 dB, a 0.83 dB deficit. Section D explicitly states 'MAT achieves superior performance across all five benchmark datasets and scale factors while requiring minimal parameters'. This is false on this entry if PSNR is the primary metric, as is standard in SR. One can argue the comparison is not parameter-matched, but such a caveat is absent. The discrepancy is large (0.83 dB) and not attributable to noise. This matters because the paper's contribution is empirical superiority; an error of this magnitude in the main comparison table undermines confidence in the reliability of the reported gains elsewhere, including the lightweight results that the reader proposed to verify.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAT, a transformer architecture for single-image super-resolution that replaces fixed-window self-attention with multi-range attention (MA) and sparse multi-range attention (SMA), and introduces the MSConvStar module for multi-range representation learning. The authors evaluate lightweight and classical variants of MAT on five benchmarks, reporting state-of-the-art PSNR/SSIM with lower parameter counts and Multi-Adds, and support the design with component ablations, LAM visualizations, ERF measurements, and runtime comparisons.","tokens_in":20,"tokens_out":6908,"duration_ms":124658,"significance":"If the empirical results are reproducible, MAT is a genuinely useful efficient SR architecture: the sparse dilated attention in SMA provides a larger effective receptive field at the same computational cost, and the reported 714K-parameter lightweight model reaches 26.83 dB on Urban100 x4, outperforming SRFormer-light with lower Multi-Adds. The paper's strengths are its component-by-component ablations (Tables I-V), the LAM/ERF analyses, the public code release, and the breadth of benchmark comparisons. However, the classical-SR state-of-the-art claim is not fully established because one baseline entry in Table VII is better in PSNR, and the reported gains are single-run values without variance; the precise magnitude of the claimed advantages is therefore uncertain.","major_comments":[{"comment":"The statement in §IV-D that \"MAT achieves superior performance across all five benchmark datasets and scale factors\" is contradicted by Table VII: on the x4 Manga109 row, MambaIR has PSNR 33.32 dB while MAT+ has 32.49 dB, a 0.83 dB deficit for the proposed method on the primary metric. Because PSNR is the standard comparison metric in SR, this sentence is false as written. The claim should be corrected or qualified (e.g., per-dataset exceptions or parameter-efficiency framing), and the MambaIR value should be verified against the original MambaIR publication because it is considerably higher than neighboring entries such as HAT-S (32.35 dB) and SwinIR (32.03 dB).","section":"§IV-D, Table VII"},{"comment":"The benchmark tables report single-run PSNR/SSIM without variance. Several of the headline gains are in the 0.08-0.16 dB range, e.g., Urban100 x4: MAT-light 26.83 vs. SRFormer-light 26.67, and Manga109 x4: MAT-light 31.38 vs. SRFormer-light 31.17; in this regime training noise of 0.1-0.2 dB is common. Please provide mean plus/minus standard deviation over at least three training runs for the main lightweight comparisons, or otherwise report seed variability, so that the claimed improvements can be distinguished from training noise.","section":"§IV-A, Tables VI-VII"},{"comment":"The abstract's \"3.3x faster\" claim is taken from the largest resolution shown in Fig. 16 (1536x1536), but at the 1280x720 setting used in Table VIII the speedup is only 162.9/71.1 about 2.29x, and the Multi-Adds ratio is 62.8/48.5 about 1.29x. The paper should state the speedup at the standard benchmark resolution and clearly qualify that the factor depends on image resolution; otherwise the headline efficiency claim is misleading.","section":"§IV-G, Fig. 16"},{"comment":"The dilation schedule delta = {9,7,5} is defined as the floor division of the 64x64 training patch size by the range sizes, and Eq. (9) samples keys at positions offset by multiples of delta. The paper does not specify how SMA is applied at test-time resolutions that are not 64x64, nor how boundary positions are handled (padding versus masking). Because the model is evaluated on arbitrary image sizes, an explicit boundary description and an experiment at a different resolution or patch size are needed to support the generalization claim for the multi-range design.","section":"Eq. (9), §IV-A"}],"minor_comments":[{"comment":"There are several typographical errors: \"MSCvonStar\" in Fig. 3 should be \"MSConvStar\", \"MSC ONV STAR\" in the Table IV heading should be \"MSConvStar\", and \"efective\" in the Fig. 15 caption should be \"effective\".","section":"Figures 3, 5, 15 and Table IV"},{"comment":"The text says \"the the number of RMAG, MAB and channel increase\" with a duplicated article; also \"Flicker2K\" should read \"Flickr2K\".","section":"§IV-A"},{"comment":"The caption says Multi-Adds are calculated for \"upscaling one image to 2562 resolution\", which appears to be a typo; the intended resolution should be stated consistently with Table VIII, where 1280x720 is used.","section":"Table VII caption"},{"comment":"Some method entries in the tables have compound venue labels (e.g., CRAFT as \"ICCV23&PAMI25\" and SRFormer cited with both ICCV and journal years); the reference list should identify the primary venue unambiguously.","section":"Tables VI-VII, References"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the Table VII discrepancy with MambaIR on Manga109 x4. If the 33.32 dB value is a typographical or copying error, a careful correction plus the requested reproducibility details could make the paper acceptable; as it stands, the empirical state-of-the-art claim is not reliable as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the useful part: MAT is a legitimately new SR architecture. The combination of multi-range regional attention (MA) and sparse multi-range global attention (SMA) with dilation in the attention region is not present in the prior work they cite, and the ablations in Tables I-V are thorough. The efficiency result is credible: 3.3x speedup over SRFormer-light is measured on an RTX 3090, and the LAM/ERF analysis shows the dilated attention actually spreads the network's attention over a wider area. The code is available, which is a plus.\n\nThe soft spot that matters: Section D states MAT achieves superior performance across all five benchmark datasets and scale factors. Table VII tells a different story. On Manga109 x4, MambaIR reports 33.32 dB while MAT+ gives 32.49 dB - a 0.83 dB deficit on one of the five datasets. The abstract and conclusion repeat the SOTA claim. This is not a rounding error. The claim needs to be qualified (e.g., 'generally superior with about half the parameters'), or the MambaIR number needs an explanation. As written, it reads as an overclaim, and it makes me slightly less confident about the other reported numbers.\n\nTwo more moderate concerns. First, all PSNR comparisons come from single runs, with no error bars or seed variance; the 0.1-0.3 dB gains over SRFormer-light on Urban100 could fall within training noise. The code helps, but multiple runs would be better. Second, the sparse sampling schedule in Eq. (9) is hand-chosen and only tested at 64x64 patch size. The assumption that this dilation schedule transfers to other resolutions is plausible but untested.\n\nOverall, this is a solid engineering paper for the SR efficiency crowd. The architecture is real, the ablations are structurally honest, and the efficiency gains look genuine. The main flaw is the overclaim in Section D, which a careful referee would have caught. I would send this to peer review, and I would ask the authors to fix the claim and add variance information before acceptance. If they do, this is a useful contribution.","headline":"New architecture, strong ablations, but the SOTA claim is contradicted by their own Table VII on Manga109 x4.","tokens_in":21818,"tokens_out":3997,"would_cite":true,"duration_ms":33216,"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":"The paper claims that a Transformer for image super-resolution can get a larger effective receptive field and richer features by mixing regional attention over several neighborhood sizes with sparse, dilated global attention—reaching…","keywords":["image super-resolution","multi-range attention","sparse attention","dilated attention","lightweight super-resolution","transformer","hierarchical feature representation"],"falsifier":"Train MAT-light with the same parameter and compute budget but with dense global attention in place of SMA, or evaluate the model on inputs whose patch size is $128\\times128$ while recomputing the dilation schedule; if the dense version matches or beats MAT, or if performance drops sharply when the schedule is re-derived for a different patch size, the multi-range sparse design is not doing the claimed work. A complementary test is to construct images whose repeating texture has a period not commensurate with the dilation step and check whether SMA misses those matches.","tokens_in":20872,"feed_emoji":"⚡","tokens_out":8184,"duration_ms":70209,"temperature":0.7,"pith_summary":"This paper tries to establish that the fixed-size attention window used by most Transformer super-resolution models is the wrong unit of efficiency. Its claim is that attending over several neighborhood sizes at once—regional attention in $7\\times7$, $9\\times9$, and $11\\times11$ blocks—plus a sparsely sampled, dilated global attention, captures hierarchical and self-similar image structure better than any single window, and at lower cost. The lightweight model MAT-light reaches $26.83$ dB on Urban100 at $\\times4$ with 714K parameters and reports the best PSNR/SSIM among lightweight methods trained on DIV2K across five benchmarks; the classical-scale model matches larger state-of-the-art transformers with about 9.6M parameters. If the claim is right, the efficiency bottleneck in SR transformers is not attention itself but how window size is chosen.","feed_headline":"Super-resolution model hits 26.83 dB with just 714K parameters","feed_subtitle":"Multi-range sparse attention reaches top quality at 3.3x the speed of fixed-window transformers.","key_machinery":"The load-bearing object is sparse multi-range attention (SMA), defined in Eqs. (8)–(10): for range size $k$ and dilation rate $\\delta$, the key-value set for pixel $(i,j)$ contains only positions $(i+x\\delta,\\ j+y\\delta)$ with $-k/2 \\le x,y \\le k/2$, so the number of attended keys stays roughly $k^2$ while the covered area grows by about $\\delta$. Multi-range attention (MA) is the non-sparse companion: several heads compute regional attention over different $k$ values and their outputs are concatenated and fused. A Local Aggregation Block uses depth-wise convolutions and channel attention for local features, and the MSConvStar module replaces the standard feed-forward network with parallel multi-scale depth-wise convolutions plus a star operation. The dilation schedule in the paper sets $\\delta$ as the floor division of the $64\\times64$ patch size by the range size, i.e., $\\{9,7,5\\}$ for ranges $\\{7,9,11\\}$, which lets SMA reach across the whole training patch at roughly constant cost.","core_discovery":"On the paper's own terms, the central discovery is that dilation can be transplanted from convolutions into self-attention to make a Transformer's receptive field flexible without raising its compute. Multi-range attention applies regional attention at several range sizes across different heads, and sparse multi-range attention samples keys on a grid stepped by a dilation rate, so a $k\\times k$ set of keys spreads across a far larger spatial footprint. Combined with a local aggregation block and the MSConvStar feed-forward module, this yields a model whose effective receptive field, measured by LAM diffusion index, covers nearly the whole input image. The paper reports that this design outperforms fixed-window transformers such as SwinIR-light and SRFormer-light on all five benchmarks at $\\times2$, $\\times3$, and $\\times4$, with the largest gains on Urban100 and Manga109, while using fewer parameters, Multi-Adds, inference time, and GPU memory.","pith_inferences":["Because the gain is tied to a dilation schedule fixed by the training patch size, a natural testable extension is to make dilation rates learned or resolution-adaptive; the paper does not report such a variant.","The same sparse-sampling logic should transfer to other dense prediction tasks such as denoising, deblurring, or segmentation, since the underlying redundancy argument concerns natural image statistics rather than super-resolution specifically.","If the benefit comes from covering a larger footprint with the same number of attended keys, then very high resolutions will thin the sampled grid and may require more range levels; the paper's experiments stop at $1536^2$ output, so this scaling behavior remains unshown.","The paper's own 'not all global dependencies are equally valuable' framing suggests that a learned gating mechanism over ranges could outperform the fixed two-heads-per-range allocation, a variant the authors leave untested."],"forward_implications":["Lightweight super-resolution can match or beat fixed-window Transformer accuracy with much less compute: MAT-light reports $26.83$ dB on Urban100 at $\\times4$ with 714K parameters and 48.5G Multi-Adds.","Scaling the same multi-range recipe works: the $9.6$M-parameter classical MAT outperforms or ties SwinIR, HAT-S, ART-S, DAT-S, SRFormer and MambaIR on most benchmark and scale combinations.","The receptive-field gain is attributable to sparse global attention: ablations show that removing SMA lowers the LAM diffusion index and PSNR more than removing MA or LAB.","The MSConvStar feed-forward module improves feature diversity while reducing parameters relative to a plain MLP or ConvFFN baseline.","Because the design avoids large padded windows, it is faster in practice: MAT-light runs about $3.3\\times$ faster than SRFormer-light and uses about 24% less GPU memory at $1536^2$ output."],"supporting_citations":[{"why":"Supplies the SwinIR baseline architecture and the fixed-window attention paradigm that MAT modifies and compares against.","marker":"[21]"},{"why":"The SRFormer baseline whose enlarged-window design is the main efficiency competitor, including the 3.3x speed comparison.","marker":"[23]"},{"why":"Provides local attribution maps and the diffusion index used to measure MAT's effective receptive field.","marker":"[20]"},{"why":"Gives the efficient neighborhood attention implementation that regional and sparse multi-range attention build on.","marker":"[52]"},{"why":"Introduces the star operation that MSConvStar uses to increase nonlinear feature expressiveness.","marker":"[28]"},{"why":"Supplies the channel attention mechanism used inside the Local Aggregation Block.","marker":"[16]"}],"fun_headline_variants":["Efficient SR with multi-range attention: 3.3x faster than SRFormer-light","Dilated attention boosts SR model quality and speed in one package","Multi-range sparse attention fixes the fixed-window bottleneck for SR","MAT: flexible attention ranges cut compute while lifting SR accuracy","Sparse multi-range attention: broad receptive fields without added cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument hinges on the idea that a sparse, evenly spaced sample of distant pixels preserves the useful non-local information that dense attention would find, at the specific dilation ratios chosen for $64\\times64$ training patches; if informative matches fall between the sampled positions, the efficiency claim loses its performance basis.","fun_headline_variants_meta":{"raw":{"variants":["Efficient SR with multi-range attention: 3.3x faster than SRFormer-light","Dilated attention boosts SR model quality and speed in one package","Multi-range sparse attention fixes the fixed-window bottleneck for SR","MAT: flexible attention ranges cut compute while lifting SR accuracy","Sparse multi-range attention: broad receptive fields without added cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000355,"raw_usage":{"total_tokens":1926,"prompt_tokens":941,"completion_tokens":985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":894}},"tokens_in":557,"tokens_out":985,"duration_ms":8680,"temperature":1.0,"reasoning_tokens":894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:21:49.030537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MAT-light with the same parameter and compute budget but with dense global attention in place of SMA, or evaluate the model on inputs whose patch size is $128\\times128$ while recomputing the dilation schedule; if the dense version matches or beats MAT, or if performance drops sharply when the schedule is re-derived for a different patch size, the multi-range sparse design is not doing the claimed work. A complementary test is to construct images whose repeating texture has a period not commensurate with the dilation step and check whether SMA misses those matches.","supporting_citations":[{"cited_title":"Swinir: Image restoration using swin transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the SwinIR baseline architecture and the fixed-window attention paradigm that MAT modifies and compares against."},{"cited_title":"Srformer: Permuted self-attention for single image super-resolution,","cited_arxiv_id":null,"evidence_quote":"The SRFormer baseline whose enlarged-window design is the main efficiency competitor, including the 3.3x speed comparison."},{"cited_title":"Interpreting super-resolution networks with local attribution maps,","cited_arxiv_id":null,"evidence_quote":"Provides local attribution maps and the diffusion index used to measure MAT's effective receptive field."},{"cited_title":"Neighborhood attention transformer,","cited_arxiv_id":null,"evidence_quote":"Gives the efficient neighborhood attention implementation that regional and sparse multi-range attention build on."},{"cited_title":"Rewrite the stars,","cited_arxiv_id":null,"evidence_quote":"Introduces the star operation that MSConvStar uses to increase nonlinear feature expressiveness."},{"cited_title":"Image super- resolution using very deep residual channel attention networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the channel attention mechanism used inside the Local Aggregation Block."}],"review_version":1}