{"id":"287068bb-c26b-4f0c-8ced-7eb62221578f","arxiv_id":"2411.11906","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"S3Mamba applies scale-modulated state space models to arbitrary-scale super-resolution, reporting marginal PSNR gains over prior INR-based methods.","lead":"Researchers present S3Mamba, a network that adapts Mamba-style state space models to upscale images to any size with one model. It reports slightly higher PSNR than previous arbitrary-scale models, but the paper overlooks a closely related Mamba-based method and the gains are tiny.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper cites MambaSR [66], an SSM-based ASSR method, but omits it from all comparisons; this directly undercuts both the 'first SSM-based ASSR' novelty claim and the SOTA claim.","rationale":"The reader's weakest_assumption correctly identifies the omission of MambaSR as the most load-bearing issue. The paper's own reference list contains the exact contradiction to its novelty claim, and the empirical tables do not include a comparison that could justify the SOTA claim. This is not a matter of differing community consensus; it is an internal inconsistency between the cited related work and the reported experiments. The proposed SSSM equations appear technically plausible, and the ablation studies show small but consistent gains over the MLP and vanilla-SSM choices, but those ablations do not address whether the full method beats an already-published SSM-based ASSR system. Given the paper's central claims, the missing comparison is so close to the claims that it warrants rejection in the current form. A head-to-head experiment would settle the matter, but until then the evidence does not support the stated novelty or SOTA.","tokens_in":997,"tokens_out":657,"duration_ms":64382,"concrete_test":"Reproduce both S3Mamba and MambaSR under the same training and evaluation protocol used in the paper (same backbones EDSR and RDN, same DIV2K and COZ splits, same scales ×2–×30) and report PSNR/SSIM and computational cost. Also confirm the publication date of MambaSR (Mathematics, vol. 12, issue 15, 2024) relative to the arXiv submission. If MambaSR achieves comparable or better performance, or if it predates S3Mamba, the central claims of 'first' and 'state-of-the-art' are unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's two central claims—(1) pioneering the State Space Model in arbitrary-scale super-resolution and (2) achieving state-of-the-art performance—both rest on the completeness and fairness of the comparison baseline set. Section 2.2 explicitly lists MambaSR [66] (Yan et al., Mathematics 12(15):2370, 2024) among recent SSM-based vision works, and the title itself announces 'Arbitrary-scale super-resolution integrating mamba with fast fourier convolution blocks.' This is a direct, published SSM-based ASSR method that predates the arXiv submission (Nov 2024). Yet Tables 1 and 2 omit MambaSR entirely from all quantitative comparisons. The text in Section 4.1 then claims 'we introduce the State Space Model into arbitrary-scale super-resolution for the first time.' If MambaSR is functional, this novelty claim is false and the performance lead over SSM-based competitors is unverified. The omission is not an obscure oversight; the authors cite the very paper they fail to compare against, making the empirical SOTA claim internally inconsistent with the cited literature.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes S3Mamba, an arbitrary-scale super-resolution method built on a 'Scalable State Space Model' (SSSM) that modulates the discretization parameters Δ and B with scale and continuous coordinate information (Eqs. 6–13), together with a 'scale-aware self-attention' mechanism (Eq. 16). The method is evaluated on synthetic DIV2K and real-world COZ benchmarks using EDSR and RDN backbones, reporting state-of-the-art PSNR/SSIM at multiple in-scale and out-of-scale magnifications. The paper also claims to be the first to introduce SSMs into arbitrary-scale super-resolution and to maintain linear computational complexity.","tokens_in":14924,"tokens_out":5348,"duration_ms":53827,"significance":"If the claims were established, the SSSM would be a useful architectural contribution: it adapts the selective-scan mechanism of Mamba to continuous coordinate and scale conditioning, potentially giving an efficient, context-aware decoder for arbitrary-scale SR. The ablation studies (Tables 3–4) show small but consistent gains when the SSSM, global feature extraction, and the proposed gating module are added, suggesting the core design is not empty. However, the headline contributions—novelty, state-of-the-art performance, and linear-complexity advantage—are currently under-supported by the evidence in the manuscript. The paper is empirical, and the main results are quantitative comparisons, so the missing baseline, missing error bars, and absent complexity analysis are load-bearing.","major_comments":[{"comment":"The paper cites MambaSR [66], titled 'Arbitrary-scale super-resolution integrating mamba with fast fourier convolution blocks', in Section 2.2, yet omits it from all quantitative comparisons in Tables 1 and 2. Section 4.1 then claims 'we introduce the State Space Model into arbitrary-scale super-resolution for the first time.' This is internally inconsistent: if MambaSR is a functional SSM-based ASSR method, the novelty claim is false, and the claimed performance lead over SSM-based competitors is untested. The authors should add MambaSR to the comparison (or justify its exclusion) and revise the 'first' claim accordingly.","section":"Section 2.2, Section 4.1, Tables 1-2"},{"comment":"The claimed state-of-the-art performance on DIV2K is not supported by the reported numbers. Against the strongest baseline CiaoSR, the differences are at most 0.02–0.03 dB, and the direction is not consistent: at ×3 the proposed method is worse on both backbones (31.13 vs 31.15 for EDSR; 31.40 vs 31.42 for RDN), and similar reversals appear at ×24 and ×30. No error bars, standard deviations, or statistical significance tests are reported, so these differences are within typical run-to-run noise for SR training. The statement in Section 5.2 that 'our method achieves the best performance in most scenarios' needs quantitative support, such as multiple seeds with variances or paired tests.","section":"Table 2, Section 5.2"},{"comment":"The module named 'scale-aware self-attention' is not an attention mechanism. Equation (16) computes a coordinate- and scale-dependent modulation αweight from SSSM, multiplies it elementwise with features, and then applies SSSM twice more. There are no attention scores, no queries/keys/values, no normalization over feature positions, and no learned dependence on the feature content. This is a multiplicative gating operation, and describing it as 'self-attention' misrepresents the contribution. The authors should either rename the module or provide a formal definition that justifies the 'attention' terminology.","section":"Section 4.2, Eq. (16)"},{"comment":"The paper repeatedly claims linear computational complexity and states in Section 5.2 that 'our computational complexity is only half of its, as shown in the supplementary materials,' but the supplementary material and appendix are not included in the submitted manuscript. No complexity analysis, FLOPs table, inference-time measurement, or formal scaling argument appears in the main text. Since the efficiency claim is one of the stated headline contributions, it needs to be substantiated in the manuscript itself, or explicitly deferred to a version that contains the appendix.","section":"Abstract, Section 4.1, Section 5.2"}],"minor_comments":[{"comment":"The heading 'Introdution' is a typo and should be 'Introduction'.","section":"Section 1 heading"},{"comment":"The table note 'Bold and Bold indicate the best performance and second-best performance, respectively' appears corrupted; it likely should describe bold and underline formatting.","section":"Table 2"},{"comment":"The notation for reference [7] is inconsistent: Table 1 uses 'LIT', while Table 2 uses 'CLIT'. Use one consistent abbreviation.","section":"Tables 1 and 2"},{"comment":"The caption contains 'discredited low-resolution images', which should be 'discretized low-resolution images'.","section":"Figure 1 caption"},{"comment":"The notation 'αweight = SSSM(coordHR, scale)' is unclear because SSSM was previously defined as a sequence model on features; clarify how coordinates and scale are prepared as inputs to the state space model.","section":"Eq. (16)"},{"comment":"The paper refers repeatedly to an appendix and supplementary material for additional comparisons and analyses, but this version contains no such sections; either include them or adjust the references.","section":"Section 4.2 and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The omitted MambaSR baseline is the most serious issue. It directly affects both the novelty and the state-of-the-art claims, and the authors themselves cite the paper in related work. I would ask the editor to require the authors to engage with MambaSR before considering the paper further. The 'self-attention' naming also needs to be corrected or justified; as written, it overstates the contribution. The manuscript may be viable after these revisions and a careful re-running of the comparisons with error bars."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the scale- and coordinate-conditioned discretization in Eqs. (6)-(13) is a genuine new idea, and the ablation tables suggest it does something. But the two headline claims—first SSM in arbitrary-scale SR, and state-of-the-art performance—are not supported as presented. Section 2.2 cites MambaSR [66], whose title literally says arbitrary-scale super-resolution with Mamba and FFT blocks, and then Tables 1 and 2 omit it entirely. That is a load-bearing problem, not a cosmetic one.\n\nWhat the paper does well: the module is simple, the idea of modulating Delta and B by scale and coordinate is clearly motivated, and the experiments include both synthetic and real-world benchmarks with a public code release. The ablations show consistent, if small, gains from each component, and the linear-complexity claim is plausible. The real-world COZ results, especially the SSIM gains, are the most encouraging part.\n\nThe soft spots are concentrated in the comparison and the claims. Missing MambaSR is the big one. The paper also reports no error bars, and most PSNR gains over CiaoSR are 0.01-0.06 dB, which is within run-to-run noise for SR. Calling the scale-aware module \"self-attention\" oversells it: Eq. (16) shows an SSSM-generated modulation map applied multiplicatively, with no Q/K/V. Rename it or rework it. And the repeated \"for the first time\" claims in Section 1 and 4.1 are simply false if MambaSR is a working SSM-based ASSR method, which the paper itself acknowledges by citing it.\n\nIs MambaSR a valid baseline? The title and venue are real, so the burden is on the authors to explain why it does not appear in the tables. If they added it and the gains are as marginal as those over CiaoSR, the practical contribution weakens to an incremental architecture tweak. If they added it and it does not fit for some technical reason, that reason needs to be stated in the paper.\n\nWho this is for: researchers working on efficient arbitrary-scale SR architectures. The SSSM modulation idea could be worth borrowing even if the paper's framing needs correction. The paper deserves a serious referee, but with a clear mandate: add the missing baseline, add variance estimates or at least multiple runs, soften the novelty claim, and stop calling the modulation module self-attention. I would send it to peer review with a request for major revision.\n\nRecommendation: engage with it, but demand the comparison be made complete before accepting any SOTA claim.","headline":"A credible scale-modulated SSM module is buried under a self-inflicted novelty problem: the paper claims to be first SSM-based ASSR while citing MambaSR and never comparing against it.","tokens_in":15395,"tokens_out":2147,"would_cite":false,"duration_ms":24075,"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":"S3Mamba claims to bring state space models to arbitrary-scale super-resolution, achieving state-of-the-art results on DIV2K and COZ with linear computational complexity.","keywords":["arbitrary-scale super-resolution","state space model","implicit neural representation","scale-aware attention","continuous representation","linear computational complexity","Mamba","image restoration"],"falsifier":"Run S3Mamba and MambaSR on the DIV2K validation set and the COZ test set under identical training protocols; if MambaSR attains comparable or higher PSNR/SSIM, or if its publication predates S3Mamba's submission as a valid SSM-based ASSR method, the paper's core claims of novelty and state-of-the-art performance would be falsified.","tokens_in":14531,"feed_emoji":"🖼️","tokens_out":4022,"duration_ms":43853,"temperature":0.7,"pith_summary":"The paper tries to establish that state space models (SSMs), known for linear-complexity global modeling, can be adapted to arbitrary-scale super-resolution (ASSR) by making the discretization process scale-aware. It proposes S3Mamba, whose Scalable State Space Model (SSSM) modulates the state transition matrix and sampling step size using scale and coordinate information, plus a scale-aware self-attention mechanism. If correct, this would provide an efficient alternative to Transformer-based ASSR decoders, which incur quadratic complexity, while improving reconstruction quality and generalization to unseen magnifications.","feed_headline":"Mamba model upscales images at any scale in linear time","feed_subtitle":"S3Mamba modulates state-space discretization by scale, building one continuous space for every magnification.","key_machinery":"The central object is the Scalable State Space Model (SSSM), which modifies the standard SSM discretization by computing scale- and coordinate-aware modulation factors via MLPs, forming $\\Delta'_{x_k} = \\Delta_{x_k} \\cdot \\Delta^{\\text{scale}}_{x_k}$ and $B'_{x_k} = B_{x_k} \\cdot B^{\\text{scale}}_{x_k}$ before applying zero-order hold. This carries the argument by making the recurrence adapt to the physical spacing between pixels at different scales, thereby unifying representations across magnifications; a scale-aware self-attention map, also generated by the SSSM from coordinates and scale, further refines the fused high-resolution features.","core_discovery":"On its own terms, the paper's central discovery is that injecting scale and continuous coordinate information into the SSM discretization process—specifically modulating the sampling interval $\\Delta_{x_k}$ and the input matrix $B_{x_k}$ by multiplying them with scale- and coordinate-dependent factors (Eqs.\\ 6–13)—creates a scale-adaptive continuous representation. This allows a single SSM-based decoder to handle arbitrary magnification factors while retaining the linear computational complexity of SSMs and outperforming existing INR-based methods on both synthetic and real-world benchmarks.","pith_inferences":["The paper cites MambaSR [66], an earlier SSM-based arbitrary-scale super-resolution method, but excludes it from all comparisons; if MambaSR is a valid baseline, the claim of being 'first' to introduce SSM to ASSR is questionable and the performance lead over SSM-based prior work is unverified.","The scale-conditional discretization idea could transfer to other continuous signal reconstruction tasks, such as arbitrary-scale inpainting or video frame interpolation, where coordinate- and scale-conditioned recurrence may help maintain temporal or spatial consistency.","The scale-aware self-attention map is generated from coordinates and scale alone, without image content; testing whether this content-agnostic conditioning generalizes across very different scene types would clarify its role versus learned content-based attention."],"forward_implications":["If SSSM works as claimed, SSM-based decoders can replace MLP and Transformer decoders in ASSR, providing linear complexity that scales to very large output resolutions.","The scale-modulated discretization could generalize beyond trained magnification factors, as the out-of-scale results on DIV2K suggest continuous representation across unseen scales.","The method is designed to be a plug-in module for existing SR backbones such as EDSR and RDN, so it could improve arbitrary-scale capability of many fixed-scale SR models.","The scale-aware self-attention mechanism may reduce artifacts in real-world degraded images, as indicated by the COZ experiments where SSIM improvements are noticeable."],"supporting_citations":[{"why":"LIIF introduced INR into arbitrary-scale super-resolution, establishing the continuous representation paradigm this paper extends.","marker":"[10]"},{"why":"CiaoSR is the main SOTA attention-based INR baseline that S3Mamba claims to outperform on DIV2K and COZ.","marker":"[2]"},{"why":"MambaIR pioneered the use of SSMs in image restoration, providing the basis for adopting SSMs in ASSR.","marker":"[21]"},{"why":"Vision Mamba's VSS block structure is used to construct the SSSM block, following its design for bidirectional state space modeling.","marker":"[76]"},{"why":"COZ provides the real-world dataset and benchmark settings used to evaluate generalization to realistic degradation.","marker":"[18]"},{"why":"MambaSR is an SSM-based ASSR method cited in the related work but omitted from comparisons, making it central to assessing the novelty and SOTA claims.","marker":"[66]"}],"fun_headline_variants":["S3Mamba: Arbitrary-scale super-resolution in linear time","Scale-aware SSM upscales images to any magnification","Mamba state space model achieves any-scale super-resolution","One model, every scale: S3Mamba's continuous representation","S3Mamba: Linear-time upscaling at arbitrary scales"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim of being the first to introduce SSM into ASSR and of achieving state-of-the-art performance rests on the assumption that all relevant baselines, including MambaSR which the paper itself cites, are included in the comparisons.","fun_headline_variants_meta":{"raw":{"variants":["S3Mamba: Arbitrary-scale super-resolution in linear time","Scale-aware SSM upscales images to any magnification","Mamba state space model achieves any-scale super-resolution","One model, every scale: S3Mamba's continuous representation","S3Mamba: Linear-time upscaling at arbitrary scales"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001111,"raw_usage":{"total_tokens":4632,"prompt_tokens":953,"completion_tokens":3679,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":3593}},"tokens_in":569,"tokens_out":3679,"duration_ms":27975,"temperature":1.0,"reasoning_tokens":3593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:18:42.911365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run S3Mamba and MambaSR on the DIV2K validation set and the COZ test set under identical training protocols; if MambaSR attains comparable or higher PSNR/SSIM, or if its publication predates S3Mamba's submission as a valid SSM-based ASSR method, the paper's core claims of novelty and state-of-the-art performance would be falsified.","supporting_citations":[{"cited_title":"Learning contin- uous image representation with local implicit image function","cited_arxiv_id":null,"evidence_quote":"LIIF introduced INR into arbitrary-scale super-resolution, establishing the continuous representation paradigm this paper extends."},{"cited_title":"Ciaosr: Continuous implicit attention-in- attention network for arbitrary-scale image super-resolution","cited_arxiv_id":null,"evidence_quote":"CiaoSR is the main SOTA attention-based INR baseline that S3Mamba claims to outperform on DIV2K and COZ."},{"cited_title":"Continuous optical zooming: A benchmark for arbitrary-scale image super-resolution in real world","cited_arxiv_id":null,"evidence_quote":"COZ provides the real-world dataset and benchmark settings used to evaluate generalization to realistic degradation."},{"cited_title":"Mambasr: Arbitrary-scale super-resolution integrating mamba with fast fourier convolution blocks","cited_arxiv_id":null,"evidence_quote":"MambaSR is an SSM-based ASSR method cited in the related work but omitted from comparisons, making it central to assessing the novelty and SOTA claims."}],"review_version":1}