{"id":"332ee2f3-eae2-40d1-9c8c-ab540c54a5d5","arxiv_id":"2412.07222","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A Mamba-plus-Transformer network with a bidirectional Channel-Mamba Block and a recursive channel-weight module reports slightly higher PSNR and SSIM than several lightweight super-resolution baselines.","lead":"This paper proposes a new super-resolution network, MPSI, that mixes Transformer and Mamba state-space blocks to model pixel interactions across long image sequences. It reports small PSNR and SSIM gains over several lightweight super-resolution baselines on five standard benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The state-of-the-art claim is not secure: baselines are taken from published lightweight results without retraining, no error bars or parameter counts are reported, and the margins over DAT are small enough to be seed noise; the x2 Manga109 result even trails DAT.","rationale":"The reader's weakest assumption is correct and load-bearing. The paper's central claim of SOTA is an empirical claim supported only by a single table with no uncertainty quantification and with baselines taken from prior publications. In image super-resolution, published baseline numbers depend on training details, and the paper itself contains an internal inconsistency in the loss description (MSE + perceptual vs L1). The reported gains over the strongest lightweight baseline, DAT, are around 0.07 dB, which is within common seed-to-seed variation for these models. The x2 results are not even uniformly better, showing the effect is not robust across scales. A controlled retraining comparison with multiple seeds and error bars would settle the issue; until then, the SOTA claim should be treated as conditional rather than taken at face value. I agree with the reader's CONDITIONAL verdict, and my stress-test does not uncover a reason to change it. The internal inconsistency in the loss function should nevertheless be corrected, and parameter/FLOPs reporting would help future verification.","tokens_in":13541,"tokens_out":6677,"duration_ms":64990,"concrete_test":"Retrain the published lightweight DAT and SwinIR models using MPSI's exact training recipe (500K iterations, Adam beta1=0.9, beta2=0.99, initial LR 2e-4 halved at 250K/400K/450K/475K, patch size 64, batch size 8, random 90/180/270 rotations and horizontal flips, L1 loss) and evaluate on the five test sets; run each with 3 seeds and report mean +/- std, and also report parameter counts for MPSI and all baselines. If MPSI's mean x4 Urban100 PSNR is within one standard deviation of DAT's retrained mean, or if MPSI has substantially more parameters than the lightweight baselines, the SOTA claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Section 1, Contribution 3; Section 4.2) that MPSI attains state-of-the-art performance depends on the comparability of Table 1 numbers. The table note says SwinIR, ELAN, MambaIR, and DAT are lightweight versions with results taken from their papers; none are retrained under MPSI's protocol, so differences in loss, iteration count, LR schedule, patch size, and augmentation are uncontrolled. MPSI's gains are tiny (e.g., x4 Urban100 PSNR 32.58 vs DAT 32.51, +0.07 dB; SSIM +0.0024), and no multiple-seed statistics are reported, so these could be within run-to-run variance. The x2 results already contradict the abstract's unqualified SOTA claim: on Manga109, MPSI (39.47) trails DAT (39.49). Additionally, the loss description is internally inconsistent: the abstract and Section 3.1 say MSE + perceptual loss, while Section 4.1 says L1 loss is minimized. This makes reproduction ambiguous and further weakens the basis for comparison. No parameter counts or FLOPs are reported, so it is unclear whether the comparison is parameter-fair.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MPSI, a single-image super-resolution network built on the Mamba state-space model. The architecture combines Spatial Transformer Blocks and Channel-Mamba Blocks into Spatial Attention Mamba Groups, plus a Mamba-based channel recursion module (MCRM) intended to propagate early-layer features. The authors report experiments on Set5, Set14, B100, Urban100, and Manga109 for ×2, ×3, and ×4 upscaling, comparing against published results of several SR models, and claim state-of-the-art performance. Ablation studies on Urban100 and Manga109 assess the contributions of CMB, MCRM, DDBM, and the Mamba recursive process.","tokens_in":13908,"tokens_out":2428,"duration_ms":23742,"significance":"If the central claim were fully supported, MPSI would be a meaningful contribution to lightweight SR, combining Mamba's long-sequence modeling with a novel feature-recursion mechanism. The paper's strengths include a clearly described architecture, a reasonable set of ablations, and an explicit comparison with several recent Transformer-based SR models. However, the state-of-the-art assertion is not yet secure: the comparison protocol restricts baselines to lightweight versions without retraining them, no error bars or parameter counts are provided, and the reported gains over DAT are small enough to be within run-to-run variation. The inconsistency between the loss stated in Section 3.1 and the loss used in Section 4.1 further undermines reproducibility. The core ideas have merit, but the evidence as presented does not justify the unqualified SOTA claim.","major_comments":[{"comment":"The claim that MPSI 'notably surpasses the existing methods, attaining state-of-the-art performance' (Section 1, Contribution 3) is not supported by the comparison as conducted. The table caption states that SwinIR, ELAN, MambaIR, and DAT results are from lightweight versions and are taken from their papers; none of these baselines is retrained under MPSI's training protocol (batch size 8, patch 64×64, 500K iterations, L1 loss, specific LR schedule). Differences in loss, iteration count, patch size, augmentation, and schedule are uncontrolled confounds. To substantiate a SOTA claim, the authors should either retrain the baselines under the same protocol or report parameter counts, FLOPs, and multiple-seed statistics with significance testing, and restrict the SOTA claim to the controlled comparison.","section":"Section 4.2 / Table 1"},{"comment":"The loss function is described inconsistently. Section 3.1 states 'Our model is optimized by MSE loss and perceptual loss,' while Section 4.1 states that the Adam optimizer is used 'to minimize the L1 loss function.' These are different objectives, and the discrepancy makes reproduction ambiguous and weakens the comparability with baselines trained with their own published losses. The authors must state exactly which loss (or combination) is used and correct the inconsistent passage.","section":"Section 3.1 vs. Section 4.1"},{"comment":"The empirical support for the SOTA claim is internally inconsistent. At ×2, MPSI trails DAT on Manga109 in both PSNR (39.47 vs. 39.49) and SSIM (0.9787 vs. 0.9788), contradicting the abstract's unqualified statement of state-of-the-art performance. In addition, the reported margins over the second-best method are very small (e.g., ×4 Urban100 PSNR +0.07 dB, SSIM +0.0024), and no error bars or multiple-seed results are reported, so these differences may not be significant. The claims in the abstract and conclusion should be qualified to reflect the actual pattern of results.","section":"Section 4.2 / Table 1"},{"comment":"No parameter counts or computational cost are reported for MPSI or the comparison models. Since the paper emphasizes designing a lightweight model and compares only with lightweight versions of baselines, the absence of parameter/FLOP tables makes it impossible to judge whether MPSI's gains are obtained within a comparable parameter budget. The definition of 'lightweight' in Section 2.2 is therefore unsubstantiated, and the comparison may not be parameter-fair.","section":"Section 4.1 / Table 1"}],"minor_comments":[{"comment":"There is a typo in Table 3: 'SIMM' should be 'SSIM'. This appears in the table body and in the surrounding text ('both PSNR and SIMM dropped').","section":"Table 3"},{"comment":"The text says 'The way SW-SA handles features in SAT is very similar to DAT [6].' The acronym 'SAT' is not defined; this should likely be 'STB' (Spatial Transformer Block).","section":"Section 3.2"},{"comment":"The baseline 'MambaIR' is included in Table 1 without a citation in the reference list. A citation to the MambaIR paper should be added.","section":"Table 1 / References"},{"comment":"The list of compared models in the text mentions CARN, EDSR, IMDN, LAPAR-A, RDN, ESRT, SwinIR, ELAN, and DAT, but Table 1 also includes MambaIR, and CARN does not appear in Table 1. The model list should be reconciled with the table.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible architectural contribution and the ablations are a positive feature. However, the central SOTA claim rests on a comparison protocol that is not rigorous enough for the claim as written. The authors should be asked to retrain baselines under a common protocol or substantially temper the SOTA language, and to resolve the loss inconsistency. If these issues are addressed, the paper could become acceptable; in its current form it is not ready for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nMPSI is a Mamba-based SR network that combines a spatial transformer block with a bidirectional channel Mamba block, then adds a channel-recursion module that reweights features using information from all earlier layers. That specific combination is not in the literature, and the paper does something many Mamba papers skip: it actually ablates the introduced modules. The ablation tables show consistent gains on Urban100 and Manga109, and replacing DDBM with channel attention hurts, which supports the design choices. At x3 and x4 the reported numbers beat the listed baselines on every dataset, and the visual comparisons look plausible.\n\nThat is the good part. The problem is the \"state-of-the-art\" claim, which does not hold up. The baseline numbers come from published lightweight versions of SwinIR, ELAN, MambaIR, and DAT; none are retrained under MPSI's training protocol. The margins are tiny — +0.07 dB at x4 Urban100, +0.0024 SSIM — and no error bars, seeds, parameters, or FLOPs are reported, so we cannot tell whether the comparison is even parameter-fair. More damning, at x2 on Manga109 MPSI trails DAT, which directly contradicts the abstract's unqualified SOTA statement. There is also an internal inconsistency: the abstract and Section 3.1 say the model is trained with MSE plus perceptual loss, but Section 4.1 says L1. That is exactly the kind of detail a reader needs to reproduce the numbers, and it is ambiguous. Finally, MambaIR appears in Table 1 but is never cited or discussed in the text — a notable gap given that it is the closest existing Mamba-based SR work.\n\nNone of this kills the architecture. With tempered claims, matched-protocol baselines, and proper uncertainty reporting, this could be a modest but honest contribution to lightweight SR. As it stands, it reads like a solid starting point in need of revision, not a finished SOTA claim.\n\nWho is it for? People working on Mamba-based low-level vision, and referees who want to verify whether the bidirectional Mamba trick genuinely helps under controlled comparison. It deserves a serious referee, but I would want the authors to fix the loss description and either retrain baselines or soften the claim before accepting.\n\nMy recommendation: send it to peer review, with a request for major revision focused on comparison methodology and reporting.","headline":"The Mamba blocks are plausibly designed and the ablations are decent, but the state-of-the-art claim is unsupported by the lightweight non-retrained baselines and is directly contradicted by the x2 Manga109 result.","tokens_in":14375,"tokens_out":2105,"would_cite":false,"duration_ms":23006,"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":"MPSI uses bidirectional Mamba state-space modeling to beat prior lightweight super-resolution models at 3x and 4x upscaling.","keywords":["single image super-resolution","Mamba","state space model","pixel-wise sequential interaction","lightweight super-resolution","channel attention","image restoration"],"falsifier":"Retrain lightweight SwinIR, ELAN, DAT, and MambaIR under exactly MPSI's training protocol—500K iterations, L1 loss, Adam with $\\beta_1=0.9$, $\\beta_2=0.99$, learning rate $2\\times10^{-4}$ with the stated halving schedule, and the same random rotation/flip augmentation—and compare on Urban100 at $\\times 4$. If any of them reaches or exceeds MPSI's reported 32.58 dB PSNR, the paper's state-of-the-art claim would fail; a second check would replace DDBM's reverse Mamba path with an additional forward pass to see whether bidirectionality itself contributes.","tokens_in":13342,"feed_emoji":"🖼️","tokens_out":7767,"duration_ms":67348,"temperature":0.7,"pith_summary":"Single-image super-resolution models built on convolutional or windowed-Transformer backbones struggle to connect distant pixels, so the paper introduces MPSI, a network that treats image features as sequences and models their long-range interactions with Mamba-style state-space blocks. MPSI combines spatial-window self-attention with a Channel-Mamba Block that runs features forward and backward through two Mamba layers, plus a Mamba channel recursion module that pools features from every block and reweights the group output by channel. In experiments with bicubic degradation, MPSI reports the best PSNR and SSIM among the compared lightweight methods on all five benchmarks at $\\times 3$ and $\\times 4$ upscaling, and strong but not uniformly best results at $\\times 2$. If those numbers hold under matched training conditions, the paper would show that selective state-space sequence models are a practical backbone for lightweight super-resolution.","feed_headline":"Beats prior lightweight SR models at 3x and 4x upscaling","feed_subtitle":"Bidirectional state-space modeling plus recursive channel weighting improves reconstruction on Urban100 and Manga109.","key_machinery":"The Channel-Mamba Block (CMB) is the central new object: it uses the Dual Direction Bi-Mamba module (DDBM) to run the feature sequence in both directions through two Mamba state-space layers, governed by $h_t = A h_{t-1} + B x_t$ and $y_t = C h_t$, then sums the forward and backward outputs through a linear projection. Mamba, a selective state-space model, gives linear-time dependency across arbitrarily long feature sequences without position embeddings, which is why the paper expects it to fix the long-range modeling gap of windowed self-attention. The second mechanism, the Mamba channel recursion module (MCRM), applies adaptive average pooling to the input and each block's output in a group, runs those pooled vectors through Mamba, and uses a sigmoid-gated multilayer perceptron to produce a channel-weight vector that multiplies the group's final feature map. MCRM is what lets early-layer features influence the final result without dense skip connections across all layers.","core_discovery":"The paper's central claim is that MPSI notably surpasses existing methods and attains state-of-the-art performance, with the reported evidence concentrated in the $\\times 3$ and $\\times 4$ upscaling tasks, where it achieves the highest PSNR and SSIM on Set5, Set14, B100, Urban100, and Manga109. Against EDSR, IMDN, LAPAR-A, RDN, ESRT, SwinIR, MambaIR, ELAN, and DAT, the largest recorded margin is on Urban100 at $\\times 4$, where MPSI beats the second-best result by 0.07 dB in PSNR and 0.0024 in SSIM. The paper attributes the gain to two mechanisms: bidirectional Mamba sequence modeling in the Channel-Mamba Block, which captures global pixel interactions that windowed attention misses, and the Mamba channel recursion module, which preserves and reweights information from earlier layers. It also notes that the SwinIR, ELAN, DAT, and MambaIR comparison numbers come from lightweight versions of those models.","pith_inferences":["One consequence the paper leaves implicit is that the state-of-the-art claim is tied to lightweight baselines whose published numbers were not reproduced under MPSI's training protocol; retraining full-size baselines under identical conditions could narrow or reverse the reported margins.","A testable extension would replace DDBM's reverse Mamba pass with an additional forward pass of the same cost; the paper's ablation replaces DDBM with channel attention, not with a unidirectional Mamba, so the value of bidirectionality itself is not isolated.","MCRM's recursive gating could be grafted onto other super-resolution backbones, such as SwinIR or DAT, as a plug-in module; the paper only tests it inside MPSI, so its transferability is unknown."],"forward_implications":["MPSI's reported results imply that selective state-space models can replace global self-attention as the long-range modeling component in a lightweight super-resolution network.","The pairing of windowed spatial self-attention (STB) with channel-wise bidirectional Mamba (CMB) indicates that spatial and channel interactions can be aggregated without full-image attention, which is cheaper at high resolution.","The MCRM design implies that recursive channel weighting can recover some of the benefit of dense feature reuse while keeping the network graph simpler.","The Urban100 margin at $\\times 4$ suggests MPSI is especially effective on images with repeated structured texture, where long-range pixel dependencies carry real information.","The $\\times 2$ results, where MPSI is not uniformly best, imply the advantage of the Mamba-based design is more pronounced at larger upscaling factors."],"supporting_citations":[{"why":"Supplies the Mamba state-space model that both CMB and MCRM are built on.","marker":"[12]"},{"why":"Provides the STB/SGFN design and the L1 training recipe, and serves as a lightweight comparison baseline.","marker":"[6]"},{"why":"Defines the SwinIR baseline whose windowed self-attention MPSI is designed to surpass.","marker":"[27]"},{"why":"Defines the ELAN baseline; its window-confined long-range attention motivates the Mamba choice.","marker":"[44]"},{"why":"Supplies the squeeze-and-excitation channel-weighting idea that MCRM adapts.","marker":"[17]"},{"why":"Provides the DIV2K training set used for the main experiments and ablations.","marker":"[1]"},{"why":"Provides the Flickr2K training set and the EDSR baseline.","marker":"[28]"}],"fun_headline_variants":["Mamba-based SR surpasses existing models on 3x and 4x upscaling","Pixel-wise sequential modeling with Mamba achieves top SR results","Mamba recursion keeps early features, boosting SR results","Mamba SR tops Urban100 at 4x, beating runner-up by 0.07 dB"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim rests on the comparability assumption that published PSNR/SSIM numbers for the lightweight baselines can be compared directly with MPSI's single training run, even though no baseline was retrained under MPSI's optimizer, loss, schedule, and augmentation.","fun_headline_variants_meta":{"raw":{"variants":["Mamba-based SR surpasses existing models on 3x and 4x upscaling","Pixel-wise sequential modeling with Mamba achieves top SR results","Mamba recursion keeps early features, boosting SR results","Mamba SR tops Urban100 at 4x, beating runner-up by 0.07 dB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001355,"raw_usage":{"total_tokens":5531,"prompt_tokens":1006,"completion_tokens":4525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":622,"completion_tokens_details":{"reasoning_tokens":4442}},"tokens_in":622,"tokens_out":4525,"duration_ms":32837,"temperature":1.0,"reasoning_tokens":4442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:56:16.759477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain lightweight SwinIR, ELAN, DAT, and MambaIR under exactly MPSI's training protocol—500K iterations, L1 loss, Adam with $\\beta_1=0.9$, $\\beta_2=0.99$, learning rate $2\\times10^{-4}$ with the stated halving schedule, and the same random rotation/flip augmentation—and compare on Urban100 at $\\times 4$. If any of them reaches or exceeds MPSI's reported 32.58 dB PSNR, the paper's state-of-the-art claim would fail; a second check would replace DDBM's reverse Mamba path with an additional forward pass to see whether bidirectionality itself contributes.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Provides the STB/SGFN design and the L1 training recipe, and serves as a lightweight comparison baseline."},{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Defines the SwinIR baseline whose windowed self-attention MPSI is designed to surpass."},{"cited_title":"In: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Defines the ELAN baseline; its window-confined long-range attention motivates the Mamba choice."},{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the squeeze-and-excitation channel-weighting idea that MCRM adapts."},{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops","cited_arxiv_id":null,"evidence_quote":"Provides the DIV2K training set used for the main experiments and ablations."},{"cited_title":"In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops","cited_arxiv_id":null,"evidence_quote":"Provides the Flickr2K training set and the EDSR baseline."}],"review_version":1}