{"id":"a7d903dc-c643-4d85-a357-f7812f73a5f9","arxiv_id":"2411.18003","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HAAT merges SDRCB from DRCT with HGAB from HMANet, reporting marginal PSNR gains on Set5/Set14 but omitting harder benchmarks, ablations, and code.","lead":"HAAT is a transformer model for image super-resolution that combines two existing attention blocks from DRCT and HMANet. Reported on two small benchmarks, its improvements over the closest baseline are about 0.02 dB, and no code or ablations are provided.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"HAAT's SOTA claim is invalid as presented: Table 1 compares HAAT's RGB-channel PSNR against published Y-channel baseline numbers, so the 0.01–0.03 dB margins over DRCT are not interpretable.","rationale":"The paper's only quantitative evidence for its headline claim is Table 1. The methodological statement in Section 3 reveals that HAAT is evaluated on RGB channels, whereas every baseline in the table was published with Y-channel scores. Since Y-channel PSNR is the field's standard and generally differs from RGB PSNR by a nontrivial, image-dependent amount, the comparison in Table 1 is invalid on its face. The training patch size and dataset differences are additional confounders, but the metric mismatch alone is sufficient to invalidate the cross-paper comparison. A simple re-evaluation under the standard Y-channel protocol would settle whether HAAT genuinely outperforms DRCT. Until that is done, the claim is unverdictable, and the reader's REJECT is appropriate. We agree partially with the reader's weakest assumption: the root issue is comparability, but the specific RGB-vs-Y inconsistency is more concrete and self-inflicted than patch size.","tokens_in":5659,"tokens_out":7750,"duration_ms":63372,"concrete_test":"Recompute HAAT's PSNR and SSIM on Set5 and Set14 using the standard Y-channel protocol (convert to YCbCr, use luminance only, remove the same border, e.g., 2×scale) as used by DRCT, and compare against DRCT's published Y-channel numbers (or re-run a public DRCT checkpoint with the same evaluation code). If HAAT's Y-channel PSNR does not exceed DRCT's 38.72/34.96/33.09 dB on Set5 ×2/×3/×4 (and corresponding Set14 values), the reported margins are an artifact of the mismatched metric and the SOTA claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 states 'For evaluation, we use all RGB channels and exclude the outermost (2 × scale) border pixels.' The baseline entries in Table 1 (EDSR, RCAN, SwinIR, HAT, DRCT, etc.) are quoted from their original papers, which evaluate PSNR and SSIM on the luminance (Y) channel of YCbCr, the standard protocol in SISR. RGB-channel and Y-channel PSNR are different metrics; RGB PSNR is generally lower than Y PSNR for the same reconstruction, and the gap is image-dependent. Therefore, HAAT's 38.74 dB on Set5 ×2 is not numerically comparable to DRCT's 38.72 dB. The table also mixes training data (DIV2K-only for older methods vs. DF2K for newer ones) and training patch size (256×256 for HAAT vs. 192×192 commonly used by DRCT and HAT). Any of these differences alone can explain the 0.01–0.03 dB margins. The paper provides no re-evaluation of baselines under its own protocol, no Y-channel numbers for HAAT, and no code; hence the central claim 'HAAT surpasses state-of-the-art' is unsupported by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HAAT, a hybrid attention transformer for single-image super-resolution that combines Swin-Dense-Residual-Connected Blocks (SDRCB) with Hybrid Grid Attention Blocks (HGAB). SDRCB integrates Swin Transformer layers with dense-residual connections, while HGAB fuses channel attention, sparse (grid) attention, and window attention. The model is trained on DF2K and evaluated on Set5 and Set14, reporting PSNR/SSIM margins of 0.01–0.03 dB over DRCT and claiming state-of-the-art performance. The paper contains no code, no ablations, no error bars, and evaluates only two small benchmarks.","tokens_in":5983,"tokens_out":5489,"duration_ms":44893,"significance":"If validated, HAAT would be an incremental improvement over existing Transformer-based SR models such as HAT and DRCT. The architecture is a plausible combination of known components, and the dense-residual design is a reasonable direction. However, the only empirical support is a cross-paper comparison on two datasets under mismatched evaluation protocols, so the central significance claim is not currently established. The paper should be credited for clearly describing the architecture and the training/evaluation settings, but it does not provide reproducibility artifacts such as code or detailed per-component analysis.","major_comments":[{"comment":"The comparison in Table 1 is not a like-for-like comparison. The text in Section 3 states that HAAT is evaluated on all RGB channels with the outermost (2 × scale) border pixels excluded, while the baseline entries (EDSR, RCAN, SwinIR, HAT, DRCT, etc.) are taken from their original papers, which report PSNR/SSIM on the Y channel of YCbCr. RGB-channel and Y-channel PSNR are different metrics with image-dependent gaps, so HAAT's 38.74 dB on Set5 ×2 is not numerically comparable to DRCT's 38.72 dB. Please re-evaluate all baselines under HAAT's protocol or report HAAT under the standard Y-channel protocol, and then re-examine the 0.01–0.03 dB margins.","section":"3, Table 1"},{"comment":"Baseline training settings are not controlled. The table mixes models trained on DIV2K only (EDSR, RCAN, SAN, IGNN, HAN, NLSN) with models trained on DF2K (SwinIR, CAT-A, HAT, DAT, DRCT, HAAT), and HAAT uses 256×256 HR patches while HAT and DRCT typically use 192×192. Any of these differences can plausibly explain the reported 0.01–0.03 dB gains. Please provide a controlled comparison, for example by re-running baselines under identical training data, patch size, and evaluation protocol, or by using a shared public implementation with the same settings.","section":"3, Table 1"},{"comment":"The claim that HAAT surpasses state-of-the-art methods is supported only on Set5 and Set14, two small benchmarks. The standard SISR evaluation also includes B100, Urban100, and Manga109, where the relative ordering of HAT/DRCT-type models can differ. Without results on these datasets, the abstract's 'benchmark datasets' claim is an overgeneralization. Please add these datasets or qualify the claim.","section":"3, Table 1"},{"comment":"The definition of HGAB appears to contain a technical error: Equation (8) computes FM = LN(MAL(Fin)) + Fin, and Equation (9) then computes FM = LN(MAL(FM)) + FM. Since the text says HGAB consists of an MAL followed by an MLP layer, the second equation should use an MLP rather than another MAL. As written, the MLP does not appear in the block, making the architecture description inconsistent and the results hard to reproduce.","section":"2.2, Eqs. (8)–(9)"},{"comment":"No ablation study isolates the contributions of SDRCB, HGAB, channel attention, sparse attention, or window attention. Because these are existing components combined in a new arrangement, the claim that the combination yields complementary gains is not evidenced. Please include ablations over the main components and attention branches.","section":"3, Experiments"}],"minor_comments":[{"comment":"The training dataset column in Table 1 spells 'DIFK' in several rows; this should be 'DF2K'.","section":"3, Table 1"},{"comment":"The list of compared methods contains 'CDRT' and 'IGN'; these should be 'DRCT' and 'IGNN', respectively.","section":"3, Table 1"},{"comment":"The notation 'XM AL' in Equation (7) appears to be a formatting error; it should be 'X_MAL' or a similar subscript.","section":"2.2, Eq. (7)"},{"comment":"The block is called 'Hybrid Grid Attention Block' in the abstract and Section 2.2 title, but 'Hybrid Global Attention Block' in the body text; please unify the terminology.","section":"2.2"},{"comment":"Reference [8] for DRCT is cited as an arXiv preprint; if a peer-reviewed version exists, it should be cited instead.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a very short submission with a large central claim based on a single table. The protocol mismatch between RGB and Y-channel evaluation, the uncontrolled training settings, and the lack of ablations mean the current evidence does not support the abstract's claim. These issues are fixable in principle, but would require a substantial experimental revision including re-evaluation of baselines, additional benchmarks, and ablations. The editor may wish to consider whether the journal's standards for reproducibility and validation can be met in revision, especially given that no code is provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xiang, quick take on HAAT: the SOTA claim is not interpretable as presented. Table 1 mixes HAAT's RGB-channel PSNR (all channels, border of 2×scale pixels excluded) with published Y-channel numbers from prior papers, which is the standard in SISR. RGB and Y PSNR are different metrics; the 0.01–0.03 dB margins over DRCT could easily be a protocol artifact. That alone sinks the headline claim.\n\nWhat is genuinely new: nothing beyond a direct combination. The paper takes SDRCB from DRCT and HGAB from HMANet, bolts them together, and reports numbers. To its credit, it is explicit about the provenance—the introduction says it \"enhances the DRCT architecture,\" and the equations match the cited blocks. No new mechanism, no analysis, no ablation.\n\nThe other soft spots are proportionate. Only Set5 and Set14 are reported, the two easiest benchmarks; B100, Urban100, and Manga109 are absent, where the margin might reverse. Training patch size is 256×256 versus 192×192 for most baselines; training data mixes DIV2K-only and DF2K. No code, no error bars, no controlled re-evaluation of baselines under HAAT's protocol. The phrase \"significantly better\" is an overstatement for 0.02 dB.\n\nThe math and equations are consistent; there is no internal contradiction. The citation pattern is honest. But the empirical support is not.\n\nBottom line: desk reject. The central claim is unsupported by the evidence in the table, and the novelty is too low to warrant referee time. I would not cite it, and I would not bring it to a journal club except as an example of protocol mismatch. If the authors re-evaluate baselines under identical conditions and add the harder benchmarks, the story might be recoverable, but as it stands the paper does not substantiate its title claim.","headline":"HAAT's headline SOTA claim is unsupported because Table 1 mixes RGB-channel PSNR for HAAT against published Y-channel baseline numbers, and the reported margins are tiny.","tokens_in":6496,"tokens_out":2753,"would_cite":false,"duration_ms":24719,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"HAAT combines channel, sparse, and window attention to push image super-resolution accuracy ahead of prior state-of-the-art models on standard benchmarks.","keywords":["image super-resolution","transformer","attention mechanism","channel attention","sparse attention","window attention","Swin transformer","dense residual connection"],"falsifier":"Train DRCT and HAAT from scratch under the same protocol (patch size 256×256, same DF2K training data, same border handling) and compare PSNR on Set5 and Set14; if HAAT no longer leads at all three scales, the claimed advantage is an artifact of training setup.","tokens_in":5485,"feed_emoji":"🖼️","tokens_out":4439,"duration_ms":37579,"temperature":0.7,"pith_summary":"The paper introduces HAAT, a transformer for single-image super-resolution that combines Swin-Dense-Residual-Connected Blocks (SDRCB) with Hybrid Grid Attention Blocks (HGAB). The authors argue that existing window-based transformer models ignore information across channels, and that a hybrid attention that adds channel attention and sparse (grid) attention to shifted-window attention can better fuse nonlocal features. On the Set5 and Set14 benchmarks, HAAT reports the highest PSNR among compared methods at scales x2, x3, and x4, surpassing DRCT by 0.02–0.03 dB. The significance is that it shows a modest but consistent accuracy gain from attending to channel and grid structure rather than only local windows.","feed_headline":"Hybrid attention transformer beats DRCT on SR benchmarks","feed_subtitle":"HAAT mixes channel, sparse, and window attention to edge ahead on Set5 and Set14 scores.","key_machinery":"The load-bearing mechanism is the attention split inside HGAB. Input features are copied into a channel-attention branch and split into two halves: one half goes to window-based multi-head self-attention (W-MSA), the other to shifted-window attention (SW-MSA), while a grid multi-head self-attention (Grid-MSA) branch processes a divided feature map with sparse attention. The outputs are concatenated and added to the channel-attention output, then layer-normalized and residually added to the input. SDRCB wraps Swin Transformer Layers in dense residual connections, concatenating features from multiple transition layers before a 1×1 convolution fuses them, with a residual scale of 0.2.","core_discovery":"The central discovery claimed is that a hybrid attention aggregation design improves super-resolution above current transformer baselines. HAAT builds on DRCT's deep residual groups but replaces the standard attention block with HGAB, which splits features into parallel branches: channel attention over the whole input, window and shifted-window attention on one half, and grid (sparse) attention on the other half, then merges them with a residual connection. With SDRCB's dense-residual connections expanding the receptive field, the network reaches 38.74 dB PSNR on Set5 at x2, 35.17 dB on Set5 at x3, and 33.12 dB on Set5 at x4, compared with DRCT's 38.72, 35.15, and 33.09 dB, respectively.","pith_inferences":["The reported margins over DRCT are small enough that a fair re-test with matched training data, patch size, and evaluation border handling could shift which model leads.","Applying the same hybrid attention split to larger windows or to multi-scale grid attention could further improve performance beyond what the paper reports.","Because the evaluation covers only Set5 and Set14, whether the gains generalize to larger or more diverse benchmarks such as Urban100 or Manga109 remains untested."],"forward_implications":["The consistent 0.02–0.03 dB advantage over DRCT suggests the hybrid attention branches capture useful information the dense-residual base misses.","Because SDRCB and HGAB are modular, they can be dropped into other transformer-based super-resolution or restoration models and would be expected to transfer their gains.","The reported results establish that channel and grid attention can be combined with shifted windows at a training patch size of 256×256 without instability."],"supporting_citations":[{"why":"DRCT is the direct architecture HAAT extends, supplying the residual dense group structure and the baseline numbers that HAAT claims to surpass.","marker":"[8]"},{"why":"HAT is the hybrid attention transformer whose window-plus-channel design HAAT modifies, and its reported scores serve as a key comparison point.","marker":"[7]"},{"why":"SwinIR is the window-attention super-resolution baseline that HAAT improves upon, providing the Swin Transformer Layer framework.","marker":"[1]"},{"why":"Swin Transformer introduces the shifted-window self-attention mechanism that SDRCB and the W-MSA/SW-MSA branches rely on.","marker":"[2]"},{"why":"HMANet supplies the grid multi-head self-attention (Grid-MSA) design that HGAB adapts for sparse global attention.","marker":"[9]"}],"fun_headline_variants":["HAAT: Hybrid attention aggregation transformer tops DRCT on SR","Merging channel, sparse, and window attention edges out DRCT","SDRCB and HGAB help transformer super-resolve past DRCT","HAAT leverages cross-channel attention to beat DRCT on SR","Hybrid attention with grid branch improves SR beyond DRCT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed state-of-the-art depends on comparing numbers measured by the authors with numbers from other papers, which used different training patches and possibly different evaluation details; under identical conditions the small margins may not hold.","fun_headline_variants_meta":{"raw":{"variants":["HAAT: Hybrid attention aggregation transformer tops DRCT on SR","Merging channel, sparse, and window attention edges out DRCT","SDRCB and HGAB help transformer super-resolve past DRCT","HAAT leverages cross-channel attention to beat DRCT on SR","Hybrid attention with grid branch improves SR beyond DRCT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000853,"raw_usage":{"total_tokens":3665,"prompt_tokens":860,"completion_tokens":2805,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":2716}},"tokens_in":476,"tokens_out":2805,"duration_ms":20097,"temperature":1.0,"reasoning_tokens":2716,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:36:18.635739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DRCT and HAAT from scratch under the same protocol (patch size 256×256, same DF2K training data, same border handling) and compare PSNR on Set5 and Set14; if HAAT no longer leads at all three scales, the claimed advantage is an artifact of training setup.","supporting_citations":[{"cited_title":"Activating more pixels in image super-resolution transformer,","cited_arxiv_id":null,"evidence_quote":"HAT is the hybrid attention transformer whose window-plus-channel design HAAT modifies, and its reported scores serve as a key comparison point."},{"cited_title":"Swinir: Image restoration using swin transformer,","cited_arxiv_id":null,"evidence_quote":"SwinIR is the window-attention super-resolution baseline that HAAT improves upon, providing the Swin Transformer Layer framework."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows,","cited_arxiv_id":null,"evidence_quote":"Swin Transformer introduces the shifted-window self-attention mechanism that SDRCB and the W-MSA/SW-MSA branches rely on."},{"cited_title":"HMANet: Hybrid Multi-Axis Aggregation Network for Image Super-Resolution","cited_arxiv_id":"2405.05001","evidence_quote":"HMANet supplies the grid multi-head self-attention (Grid-MSA) design that HGAB adapts for sparse global attention."}],"review_version":1}