{"id":"dedd62f1-5aed-47ec-ba37-77da3dc8063c","arxiv_id":"2506.03740","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"SAAT, an alternating channel-spatial-window attention Transformer, reports slight PSNR/SSIM gains over SwinIR and HAT, but the evidence is weakened by inconsistent baselines and test-set tuning.","lead":"This paper introduces SAAT, a Transformer for image super-resolution that alternates channel attention, spatial attention, and shifted-window attention groups. It reports top benchmark scores, but the comparisons contain misreported baselines and the model was tuned on a test dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA-parity claim rests entirely on Table 1, and Table 1's SRFormer rows match published values at ×2 but fall 0.5–1.5 dB below them at ×3/×4 (even below the published light variant), so the comparison baselines — and the claimed gains — are not trustworthy.","rationale":"The reader's weakest assumption — that Table 1's fairness assumes a common training/evaluation protocol — points at the same defect I identify, and my pass sharpens it from 'unverified' to 'demonstrably violated.' The decisive evidence is internal to the table: ×2 SRFormer numbers match the published literature digit-for-digit, while ×3/×4 numbers do not, and at ×4 the full SRFormer model is reported below the published light variant on all five datasets. A uniform protocol problem would distort all rows; instead the distortion is selective, and it targets the baseline the paper most needs to beat. The separate text/table contradiction (§4.4.1: 'improves by 0.10 dB' vs the table's 0.67–0.69 dB over SwinIR) shows the quantitative reporting is not internally consistent, independent of any external comparison. I do not treat the absence of code as a standalone objection; that would be process criticism. It becomes a correctness risk precisely because the one checkable baseline row is already wrong, so no avenue remains for external verification. The architecture itself is described in enough detail that the modules are implementable, and the internal equations are mutually consistent; I also credit the ablation for showing the SMSAB/ECAB branches contribute something (≈1.2 dB in the ablation configuration), although the magnitude is larger than is typical for lightweight attention branches and the hyperparameters (α, β, μ) were tuned on Urban100, which later appears as a headline benchmark. These positives do not rescue the central claim, because the claim is comparative and the comparison is unreliable. The reader's REJECT verdict stands; the constructive path is a revision that releases code and checkpoints, reports parameter counts, and rebuilds Table 1 from official baseline checkpoints under one documented protocol. Hence verdict_should_be is UNCHANGED.","tokens_in":14530,"tokens_out":12521,"duration_ms":103970,"concrete_test":"Reproduce the baseline rows: download the official SRFormer, SwinIR, and HAT checkpoints from the authors' public repositories, evaluate them at ×3 and ×4 on Set5, Set14, BSD100, Urban100, and Manga109 under one fixed protocol (Y-channel PSNR/SSIM with the standard 8-pixel border crop, no self-ensemble), and diff every entry against Table 1. If the official SRFormer values exceed the paper's by more than 0.3 dB on any dataset, or if the control baselines disagree, the comparison is misreported and the SOTA-parity claim is unestablished. Separately, instantiate the described SAAT configuration (3 SWSAG + 3 CWSAG groups, 4 SWSAB + 4 CWSAB blocks per group, 180 channels, window size 16, OCAB with μ = 0.5) and count its parameters against SwinIR's 11.9M and HAT's published counts; if SAAT exceeds HAT's parameter count, the 'same quantity of parameters' premise fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SAAT's central claim, 'performance comparable to the state-of-the-art under the same quantity of parameters,' is substantiated only by Table 1, and Table 1 is demonstrably unreliable for the transformer baselines that define SOTA. The SRFormer row at ×2 (Set5 38.51, Set14 34.44, BSD100 32.57, Urban100 34.09, Manga109 40.07) matches the published SRFormer values exactly, which shows the authors had access to the published numbers; yet at ×3 and ×4 the same baseline is reported 0.5–1.5 dB below the published values (e.g., ×4 Set5 32.19 vs 33.00, Set14 28.69 vs 29.23, Urban100 26.39 vs 27.69, Manga109 30.75 vs 32.24), and is in fact below the published SRFormer-light variant on every ×4 dataset. Since the SwinIR, HAT, and EDT rows in the same table match their published numbers, the discrepancy is not a uniform protocol difference; the SRFormer baseline specifically is depressed, inflating SAAT's apparent margins. Two supporting inconsistencies compound this: §4.4.1 states SAAT improves over SwinIR by 0.10 dB on Urban100 and Manga109, while Table 1 shows 0.67–0.69 dB at ×2 and ×4 — both cannot be right; and the paper reports no parameter counts for any model, so the 'same quantity of parameters' qualifier that defines the claim's scope is unverifiable, while the described architecture (24 blocks, each containing (S)W-MSA, OCAB, an MLP, and an ECA or SMSAB branch) plausibly exceeds HAT's parameter budget. With no code or checkpoints released, Table 1 is the entire empirical basis, and at least one of its rows is wrong.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAAT, a Transformer-based single-image super-resolution model that alternates between groups combining shifted-window multi-head self-attention with efficient channel attention (CWSAG) and with a multi-scale shared spatial attention block (SWSAG), adding overlapping cross-attention blocks and a ConvFFN. The authors claim that SAAT achieves performance comparable to state-of-the-art (SOTA) under the same parameter count, and they support this with quantitative and visual comparisons on Set5, Set14, BSD100, Urban100, and Manga109 in Table 1, plus ablation studies on Urban100. The central evidence is Table 1, which reports SAAT as best on most datasets and scales.","tokens_in":14946,"tokens_out":4284,"duration_ms":40255,"significance":"If the reported results were reliable, the architecture would represent a plausible incremental combination of window attention, channel attention, and spatial attention for SR, and the claimed parity with, or improvement over, SwinIR and HAT at equal parameter budgets would be of interest to the low-level vision community. However, the empirical foundation is not trustworthy: the SRFormer baseline in Table 1 is inconsistent with published values at ×3 and ×4 while matching at ×2; the text and the table contradict each other on the size of the SwinIR gain; the ablation table does not reconcile with the main table; and hyperparameters are tuned on a test set. The paper also reports no parameter counts for any model and releases no code or checkpoints, so the central 'same quantity of parameters' claim is unverifiable. The core architectural idea is plausible but not new in isolation, and the paper does not convincingly demonstrate its contribution.","major_comments":[{"comment":"The SRFormer baseline is internally inconsistent with published results. At ×2, the reported values (e.g., Set5 38.51, Urban100 34.09, Manga109 40.07) match the published SRFormer numbers exactly, but at ×3 and ×4 they fall roughly 0.5–1.5 dB below the published values (e.g., ×4 Set5 32.19 vs 33.00, Urban100 26.39 vs 27.69, Manga109 30.75 vs 32.24), even falling below the published light variant on every ×4 dataset. Since the SwinIR, HAT, and EDT rows in the same table match their published numbers, this is not a uniform protocol difference. This specifically depresses a key transformer baseline and inflates SAAT's apparent margins, so the SOTA comparison and the abstract's central claim are not supported.","section":"Table 1, SRFormer rows"},{"comment":"The text states that SAAT improves over SwinIR by 0.10 dB on Urban100 and Manga109, but Table 1 shows differences of 0.69 dB (×2 Urban100: 34.50 vs 33.81), 0.67 dB (×4 Urban100: 28.12 vs 27.45), and 0.51–0.54 dB on Manga109 at ×2/×4. These two claims cannot both be correct, indicating an error in either the text or the table.","section":"§4.4.1 vs Table 1"},{"comment":"The ablation in Table 2 reports that the full model with both SMSAB and ECAB achieves PSNR/SSIM 28.47/0.8459 on Urban100, yet Table 1 lists SAAT at ×4 Urban100 as 28.12/0.8391. If both evaluations use the same Urban100 test set and the same protocol, the main result is lower than the ablation's 'best' configuration, which is a direct inconsistency that needs explanation.","section":"Table 2 vs Table 1"},{"comment":"The hyperparameters α, β, and μ are selected by maximizing PSNR on Urban100 (Tables 3 and 4), and Urban100 is also one of the evaluation datasets in Table 1. Part of the reported Urban100 gain is therefore a tuned result rather than a predicted performance, so the Urban100 margins in Table 1 are not an honest measure of generalization.","section":"§4.3.2–§4.3.3, Tables 3–4"},{"comment":"The central claim, 'performance comparable to the state-of-the-art under the same quantity of parameters,' is unverifiable because no parameter counts are reported for SAAT or any baseline. Given the architecture description (24 blocks, each containing (S)W-MSA, OCAB, an MLP, and an ECA or SMSAB branch), it is not obvious that the parameter count is comparable to HAT or SwinIR; the authors need to provide exact parameter counts and FLOPs to substantiate the claim.","section":"Abstract and §4.2"},{"comment":"No training or evaluation details are given for any baseline method, nor is any code or checkpoint released. The paper does not state which official implementations, patch sizes, augmentation schemes, or evaluation codes were used for EDSR, RCAN, SwinIR, SRFormer, EDT, or HAT. Given the SRFormer discrepancy, the fairness of the comparison cannot be assessed, and the reader cannot reproduce the baseline numbers.","section":"§4.2 and Table 1"}],"minor_comments":[{"comment":"The notation reuses f_LN for both the first and second LayerNorm without distinguishing them; this makes the equations ambiguous and should be fixed with subscripts such as f_LN^1 and f_LN^2.","section":"Eqs. (3)–(4)"},{"comment":"The left-hand side and the right-hand side both use the symbol X_i^H (or X_i^W) with different meanings; different symbols should be used for the input sub-feature and the convolved output.","section":"Eqs. (5)–(6)"},{"comment":"The text says 'we first average the input X across each dimension to obtain two one-dimensional sequences: X_H ∈ ℝ^{B×C×H×W} and X_W ∈ ℝ^{B×C×H×W}.' Averaging should produce lower-dimensional tensors (e.g., B×C×H and B×C×W); the stated shapes are inconsistent with the description.","section":"§3.3.1, SMSAB description"},{"comment":"Several typographical errors appear: 'Flicker2K' should be 'Flickr2K' (§4.1), 'SAA T' and 'HA T' are artifacts in Table 1, and 'LayNorm' appears in Figure 1.","section":"Throughout"},{"comment":"Reference [40] (SCSA) is cited for the synergistic use of spatial and channel attention but is not discussed in the related-work section; the relationship between SCSA and the proposed alternating SWSAG/CWSAG design should be clarified to position the novelty.","section":"Related work and §3"},{"comment":"The conclusion says SAAT is 'outperforming state-of-the-art techniques,' while the abstract claims only 'comparable to the state-of-the-art'; these claims should be reconciled.","section":"§5 Conclusion"}],"recommendation":"reject","confidential_remarks":"The pattern in Table 1—where SRFormer matches published numbers at ×2 but is 0.5–1.5 dB below published values at ×3 and ×4, while SwinIR, HAT, and EDT match their published numbers—strongly suggests the table was assembled without careful verification of the baselines. Combined with the internal contradictions between §4.4.1, Table 2, and Table 1, as well as the test-set tuning of α, β, and μ on Urban100, the reported results cannot be considered reliable. If the authors resubmit, the editor should require raw per-image logs, exact parameter counts, and a clear statement of the training/evaluation protocol for every baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: SAAT is an engineering combination of existing attention modules (ECA, shifted-window attention, a shared multi-scale depthwise-1D spatial attention, and HAT's OCAB) arranged in alternating groups. That arrangement is sensible, and the spatial-attention block is a neat lightweight idea. But the paper's central claim—SOTA parity under equal parameters—rests on Table 1, and Table 1 does not hold up.\n\nThe SRFormer rows at x3 and x4 are 0.5–1.5 dB below the published values, while the SwinIR, HAT, and EDT rows in the same table match their published numbers. That pattern is not a protocol difference; someone transcribed or generated the SRFormer numbers independently, and they are wrong. Since the paper's margins over SOTA are only a few tenths of a dB, this is load-bearing.\n\nThere are also internal contradictions. Section 4.4.1 says SAAT improves over SwinIR by 0.10 dB on Urban100 and Manga109; Table 1 shows 0.47–0.70 dB at every scale. Both cannot be true. The ablation tables use a setup that lands at ~27.8 dB PSNR on Urban100, while the full model in Table 1 reaches 28.12 dB; the paper never explains that gap, and the ablation gains of more than a dB for a single module are implausible unless the baseline was undertrained. No parameter counts are reported anywhere, so the 'same quantity of parameters' qualifier is unverifiable; with 24 blocks plus OCABs, SAAT likely exceeds HAT's budget. Finally, no code or checkpoints are provided.\n\nWhat this paper does well: it reads cleanly, the motivation is clear, and the alternate-group idea is worth exploring. If the numbers were solid, a 0.1–0.7 dB gain over SwinIR would be a useful incremental step for the SR community.\n\nBut as is, the evidence is not trustworthy. The authors need to recheck all baseline values, reconcile the text with the tables, report parameter counts, and release code so the comparisons can be reproduced. A serious referee would return this for data-integrity fixes before any scientific evaluation.\n\nRecommendation: do not send to peer review in its current form. Ask for a corrected and code-released version, then reconsider.","headline":"A reasonable architecture combination undermined by inconsistent baseline numbers and a table that contradicts its own text; not ready for publication.","tokens_in":15447,"tokens_out":5008,"would_cite":false,"duration_ms":46130,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SAAT claims a super-resolution Transformer that beats SwinIR and HAT at equal parameter count by alternating channel and spatial attention with shifted windows.","keywords":["image super-resolution","Transformer","efficient channel attention","multi-scale spatial attention","shifted window attention","overlapping cross-attention","single image super-resolution","PSNR and SSIM"],"falsifier":"Evaluate the released SwinIR, HAT, SRFormer, and EDT checkpoints with the exact evaluation code used for SAAT on Set5, Set14, BSD100, Urban100, and Manga109, and retrain SAAT in the same training loop; if SAAT's reported margin over HAT and SwinIR at equal parameters drops to within run-to-run noise, roughly 0.05 dB, the paper's central claim fails.","tokens_in":14332,"feed_emoji":"🖼️","tokens_out":10319,"duration_ms":97263,"temperature":0.7,"pith_summary":"The paper proposes SAAT, a Transformer for single-image super-resolution whose deep feature extractor alternates between two attention groups. The CWSAG group pairs shifted-window self-attention with a cheap efficient channel attention block, and the SWSAG group pairs it with a shareable multi-scale spatial attention block; an overlapping cross-attention block closes each group. Its central claim is that this alternation recovers global channel information and spatial structural detail that window-only attention discards, and that it pushes PSNR and SSIM above the compared models on nearly every scale of Set5, Set14, BSD100, Urban100, and Manga109 at the same parameter count. A sympathetic reader would care because it suggests the synergy between channel and spatial attention, rather than simply more depth or parameters, is a practical route to sharper texture and edge reconstruction.","feed_headline":"SAAT beats SwinIR and HAT at equal parameter budget","feed_subtitle":"Alternating channel and spatial attention with shifted windows lifts restored-image quality on five benchmarks.","key_machinery":"The load-bearing mechanism is the alternating pair of attention groups. CWSAG computes (S)W-MSA in parallel with ECAB, whose average-pooled features are reweighted by a fast 1D convolution with an adaptively chosen kernel size, adding global channel context without a dimension-reducing bottleneck. SWSAG computes (S)W-MSA in parallel with SMSAB, which averages features along height and width, splits them into four parts, applies shared depthwise 1D convolutions with kernels 3, 5, 7, and 9, concatenates the parts through GroupNorm, and multiplies the resulting horizontal and vertical attention masks onto the features. Each group ends with an Overlapping Cross-Attention Block so adjacent windows exchange information, and the whole stack alternates SWSAG and CWSAG three times. The scalar factors $\\alpha$ and $\\beta$, both set to 0.01 in the experiments, temper the parallel branches so they do not fight the window-attention branch during optimization.","core_discovery":"On the paper's own terms, the central discovery is that standard window self-attention wastes useful signal, and that two parallel attention branches placed in alternating groups can recover it without a parameter blow-up. In the Efficient Channel & Window Synergistic Attention Group, (S)W-MSA is computed in parallel with ECAB, whose average-pooled features pass through a fast 1D convolution with an adaptive kernel size to reweight channels. In the Spatial & Window Synergistic Attention Group, (S)W-MSA is computed in parallel with SMSAB, which averages the input along height and width, splits each into four parts, runs shared depthwise 1D convolutions with kernels 3, 5, 7, and 9, and multiplies the resulting vertical and horizontal attention maps onto the features. The paper reports the best numbers in its table on almost all scales, for example 28.12 dB on Urban100 at scale 4 versus 27.46 dB for HAT, and interprets the margin as evidence that alternating channel and spatial attention with window attention is what drives accuracy.","pith_inferences":["If the reported gains hold under a controlled re-run, the alternation implies that channel and spatial attention need not be merged inside a single block; sequencing them in separate groups may be enough, a claim the paper does not directly test.","The same alternating-group recipe is a plausible candidate for other image restoration tasks such as denoising and JPEG artifact removal, which the paper does not experiment with.","The Table 1 numbers for SRFormer at scale 3 are far below that model's published results, so re-evaluating with official checkpoints is the quickest check on whether the state-of-the-art comparison is reliable."],"forward_implications":["At the same parameter budget, SAAT reports PSNR gains over SwinIR of roughly 0.3 to 0.7 dB across the tested scales, with the largest improvements on the texture-heavy Urban100 and Manga109 sets.","Because the channel branch uses an adaptive 1D convolution rather than a channel-reduction bottleneck, the reported gains come without adding a large parameter overhead compared with standard channel attention.","The reported visual comparisons show cleaner edges and repeated textures, indicating that the multi-scale spatial branch contributes structural detail rather than only nudging a metric.","If reproduced, the alternating-group design offers a way to improve transformer-based restoration that does not depend on the same-task pretraining strategy used by HAT."],"supporting_citations":[{"why":"Supplies the SwinIR-style shifted-window attention and MLP structure that SAAT extends with parallel channel and spatial branches.","marker":"[12]"},{"why":"SRFormer is the permuted self-attention baseline whose parameter efficiency and spatial retention SAAT claims to surpass.","marker":"[13]"},{"why":"HAT provides the channel-attention and overlapping cross-attention (OCAB) components that SAAT recombines into its two attention groups.","marker":"[15]"},{"why":"ECA-Net supplies the efficient channel attention scheme, an adaptive 1D convolution without dimensionality reduction, used in ECAB.","marker":"[42]"},{"why":"DIV2K is the training dataset that, together with Flickr2K, defines the data protocol for the reported results.","marker":"[43]"},{"why":"Flickr2K is the second training dataset, and its inclusion in the same protocol is load-bearing for comparing against baselines trained on DIV2K alone.","marker":"[44]"}],"fun_headline_variants":["SAAT: Alternating attention groups sharpen images","Channel-spatial synergy beats standard window attention","SAAT: New attention mix matches SOTA at same size","Alternating channel-spatial attention lifts SR quality","SAAT: Efficient attention pairing for super-resolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that SAAT beats the compared models at equal parameters rests on the assumption that every baseline in Table 1 was trained and evaluated under the same protocol as SAAT, including the same data, patch size, augmentation, iterations, and metric code, which the paper does not demonstrate and which its SRFormer entries contradict.","fun_headline_variants_meta":{"raw":{"variants":["SAAT: Alternating attention groups sharpen images","Channel-spatial synergy beats standard window attention","SAAT: New attention mix matches SOTA at same size","Alternating channel-spatial attention lifts SR quality","SAAT: Efficient attention pairing for super-resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1429,"prompt_tokens":1052,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":304}},"tokens_in":668,"tokens_out":377,"duration_ms":4368,"temperature":1.0,"reasoning_tokens":304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:55:51.147771+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the released SwinIR, HAT, SRFormer, and EDT checkpoints with the exact evaluation code used for SAAT on Set5, Set14, BSD100, Urban100, and Manga109, and retrain SAAT in the same training loop; if SAAT's reported margin over HAT and SwinIR at equal parameters drops to within run-to-run noise, roughly 0.05 dB, the paper's central claim fails.","supporting_citations":[{"cited_title":"Swin transformer: Hierarchical vision transformer us- ing shifted windows","cited_arxiv_id":null,"evidence_quote":"Supplies the SwinIR-style shifted-window attention and MLP structure that SAAT extends with parallel channel and spatial branches."},{"cited_title":"SRFormerV2: Taking a Closer Look at Permuted Self-Attention for Image Super-Resolution","cited_arxiv_id":"2303.09735","evidence_quote":"SRFormer is the permuted self-attention baseline whose parameter efficiency and spatial retention SAAT claims to surpass."},{"cited_title":"Activating more pixels in image super resolution transformer","cited_arxiv_id":null,"evidence_quote":"HAT provides the channel-attention and overlapping cross-attention (OCAB) components that SAAT recombines into its two attention groups."},{"cited_title":"ECA-net: Ef- ficient channel attention for deep convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"ECA-Net supplies the efficient channel attention scheme, an adaptive 1D convolution without dimensionality reduction, used in ECAB."},{"cited_title":"chal- lenge on single image super-resolution: Methods and results","cited_arxiv_id":null,"evidence_quote":"DIV2K is the training dataset that, together with Flickr2K, defines the data protocol for the reported results."},{"cited_title":"Enhanced deep residual net- works for single image super-resolution","cited_arxiv_id":null,"evidence_quote":"Flickr2K is the second training dataset, and its inclusion in the same protocol is load-bearing for comparing against baselines trained on DIV2K alone."}],"review_version":1}