{"id":"aed681d6-a5d6-4484-b5ff-93ff0b960d79","arxiv_id":"2506.11768","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A state-space model for video super-resolution gains long-range spatiotemporal modeling by ordering tokens according to a content-aware, spectral-clustering-based scan.","lead":"MambaVSR is a new video super-resolution network that uses a state-space model with a content-aware scanning order, learned per video via spectral clustering, to capture long-range dependencies across frames cheaply. It reports higher PSNR than transformer-based VRT with 55% fewer parameters on REDS, but releases no code and describes the core modules without equations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 0.58 dB gain over VRT rests on Table 1 numbers that are internally inconsistent with the paper's own ablation setting, so the central comparison is not yet reproducible and the mechanism is not established as the cause.","rationale":"The reader identified the Fiedler-vector scan-order assumption as the weakest assumption, and I agree that its semantic validity and cross-frame stability are unanalyzed. However, my stress-test pass found a more load-bearing and more basic issue: the paper's own reported numbers are not self-consistent across Table 1 and the ablation tables, and the method section is too underspecified to reimplement or even to map configurations to results. The headline advantage (32.75 vs 32.17 dB) is a single unverified run with no code. The strongest claim therefore rests on reproducibility of that number, and the contribution's causal mechanism (content-aware scanning as the cause of the advantage) is supported only by a 0.09 dB ablation on REDS4 and a 0.24 dB ablation on Vid4, both at a different iteration count than the headline experiments. Because such gaps are addressable by the authors (code, exact hyperparameters, multi-seed runs, an order-stability analysis), the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT. I do not see internal logical contradiction in the architecture description; the problem is under-specification plus numerical inconsistency, not fraud or circular reasoning.","tokens_in":11528,"tokens_out":1924,"duration_ms":15326,"concrete_test":"Run the 6-frame REDS training configuration exactly as described in §4.1 (300k iterations, batch size 8, 64×64 patches, Charbonnier loss, lr 2e-4) and evaluate on REDS4; if the PSNR does not reach roughly 31.9–32.0 dB as in Table 1 and does not exceed the raster-scanning baseline (Table 2, 31.68 dB at 200k) when evaluated at the same iteration count, the claimed mechanism and headline advantage over VRT are not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central empirical claim is the 32.75 dB REDS4 result under 16-frame training with 14.1M parameters, beating VRT (32.17 dB, 30.7M). This result is a single reported run with no code release, and the method description in §3.2–3.3 is too underspecified to reimplement: spectral clustering, Fiedler-vector sorting, PatchAlignment, interleaving, and the graph construction are each described in only one or two sentences. Internal cross-checks raise further doubt. The Table 1 ablation-style comparison presented as 'MambaVSR (Ours)' under 6-frame training is 31.95 dB, but the Table 2 ablation baseline 'Raster-based scanning' (13.91M params, 200k iterations) is 31.68 dB and the full content-aware model in Table 2 is 31.82 dB (14.13M params); no reported configuration corresponds exactly to Table 1's 6-frame 'MambaVSR' (14.1M, 300k iterations). If MambaVSR's 16-frame 32.75 dB value cannot be reproduced from the supplied configuration, the headline claim of 0.58 dB over VRT has no independent support, and the attribution of that gain to content-aware scanning (CAS) is not supported: the CAS ablation in Table 2 measures a 0.09 dB REDS4 gain over raster scanning at 200k iterations, whereas the 0.86–0.58 dB gains over EDVR/VRT are claimed at the 300k/600k settings. Nothing in the text shows that the Fiedler-vector scan order is semantically stable across frames, and no plot of the learned order against motion is provided. In addition, the paper claims 'rigorous theoretical analysis' in the contributions but Section 3 contains no theorem, proof, or complexity analysis, so the linear-complexity claim is asserted, not demonstrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MambaVSR, a video super-resolution (VSR) network built on selective state-space models (SSMs). It introduces Shared Compass Construction (SCC), which builds intra-frame similarity graphs via efficient sparse attention and derives a spatial scan order by spectral clustering (Fiedler vector); Content-Aware Sequentialization (CAS), which reuses the current frame's scan order after coarse PatchAlignment and interleaves aligned temporal features; and the Global-Local State Space Block (GLSSB), which combines window self-attention with SSM-based feature propagation. Experiments on REDS4, Vimeo-90K-T, and Vid4 report 32.75 dB PSNR on REDS4 under 16-frame training with 14.1M parameters, outperforming VRT by 0.58 dB with 55% fewer parameters, along with ablations of the scanning mechanism and GLSSB components and additional LAM/DI and large-motion analyses.","tokens_in":11838,"tokens_out":4709,"duration_ms":42387,"significance":"If the reported numbers and the mechanism are correct, this is a potentially notable contribution: it would be the first state-space-model framework for VSR, and content-aware scanning would provide a way to aggregate non-local similar content across misaligned frames in linear time, a meaningful advance over optical-flow-based and quadratic-attention methods. The paper's strengths include broad benchmark coverage, comparisons against strong CNN and transformer baselines, ablations of the scanning and GLSSB modules, and the supplementary LAM/DI and large-motion experiments. However, the central mechanism is currently underspecified, the advertised 'rigorous theoretical analysis' is not delivered, and the headline Table 1 results are not reconciled with the ablation settings. These issues are load-bearing for the paper's central claims and must be resolved before the results can be accepted.","major_comments":[{"comment":"SCC and CAS are described only in prose: no equation or algorithm block defines the sparse-attention similarity matrix, the graph Laplacian, the Fiedler-vector computation, the spectral-clustering step, the PatchAlignment module, or the interleaving rule that produces the final scan sequence. Consequently, the claimed linear complexity cannot be checked, and the 'rigorous theoretical analysis' advertised in the third contribution bullet is not present anywhere in the paper. Please provide formal definitions and algorithm pseudocode, and either add the theoretical analysis or remove the claim.","section":"Sections 3.2-3.3 and Contribution bullet 3"},{"comment":"The 6-frame MambaVSR row in Table 1 reports 14.1M parameters, 300k iterations, and 31.95 dB on REDS4, while the corresponding full content-aware model in Table 2 (14.13M parameters, 200k iterations) reports 31.82 dB, with the raster baseline at 31.68 dB. These numbers are not reconciled without an explicit statement of which configuration changes from the ablation to the main table, and no Table 2 configuration exactly matches Table 1's 6-frame entry. As a result, the headline 16-frame 32.75 dB value is not tied to a reproducible configuration, and the claimed 0.58 dB advantage over VRT lacks independent support. Please report the exact configuration of each Table 1 row and align the ablation protocol with the main training schedule.","section":"Table 1 versus Table 2 (Sections 4.1 and 4.3)"},{"comment":"The causal attribution of the performance advantage to content-aware scanning is not supported by the ablation evidence. Table 2 shows only a 0.09 dB REDS4 gain (31.68 to 31.82) for content-aware over raster scanning at 200k iterations, whereas the large gains over EDVR and VRT in Table 1 are measured at 300k/600k iterations. Nothing in the text demonstrates that these larger gains are caused by CAS rather than by the longer training schedule or other components. Please provide ablations at the same training schedule as the main comparison, or explicitly quantify the effect of the schedule.","section":"Section 4.3, Table 2"},{"comment":"The reused-order assumption underlying CAS is unexamined. The authors state that PatchAlignment enables reuse of the current frame's Fiedler-vector order across neighboring frames, but no analysis or experiment tests whether this order remains semantically stable under the motions present in REDS and Vid4. If the scan order does not keep similar patches adjacent after alignment, the proposed mechanism would not produce the observed gains. Please add a quantitative stability analysis (e.g., order-overlap or adjacency-consistency metrics across frames), an ablation that recomputes the order per frame, and ideally visualizations of the learned orders against motion.","section":"Sections 3.3 and 4.3"}],"minor_comments":[{"comment":"The name 'Fielder' should be spelled 'Fiedler' throughout the paper.","section":"Throughout (e.g., Table 2, Section 4.3)"},{"comment":"The word 'Cilp' should be 'Clip' in the figure captions.","section":"Figure 2 captions"},{"comment":"The contribution list promises 'rigorous theoretical analysis', but no theoretical analysis appears in Sections 3 or 4; either provide it or reword the contribution.","section":"Contribution bullet 3"},{"comment":"The 'Frames REDS/Vimeo' column is ambiguous for rows with a single number; please clarify which dataset each frame count refers to, especially for rows like '16/14'.","section":"Table 1 and Section 4.1"},{"comment":"Axis labels and units are missing in the large-motion evaluation plot, and it is unclear whether the reported PSNR improvements are averaged over all REDS4 clips.","section":"Figure 4"},{"comment":"The conclusion contains minor grammatical issues (e.g., 'In this paper, We present' and the missing final period after 'datasets').","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The central discrepancy between Table 1 and Table 2 is my main technical concern and, if fixable through clarified configurations and added ablations, the contribution could be viable. I also encourage a careful literature check on the 'first SSM framework for VSR' claim, since several Mamba-based video restoration works are cited (e.g., VFIMamba, RainMamba) and the distinction from them is not crisply established. Given the current level of method under-specification and the absence of code, the reproducibility of the headline result is not yet established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: the content-aware scanning idea is worth paying attention to, but the paper as written cannot be used to build on. The core mechanism—computing a per-video scan order via spectral clustering on a sparse attention graph, then interleaving temporally aligned features along that order—is a genuinely fresh way to give an SSM a global receptive field that respects content rather than raster order. As far as the cited literature goes, it does appear to be the first selective-SSM VSR framework, and the reported numbers, if real, would be a solid subfield advance: 0.58 dB over VRT with roughly half the parameters on REDS4.\n\nWhat the paper does well: the idea is clearly motivated, the architecture is plausible, and the experiments include the expected benchmarks plus a LAM/DI analysis that at least attempts to show the scanning order does what it claims. The ablation structure, comparing raster, Fiedler, and full content-aware scanning, gives some evidence that the proposed mechanism helps.\n\nThe soft spots are real and several. The method sections for SCC, CAS, and GLSSB are sketches—no equations, no algorithm blocks, no tensor shapes. The contribution bullet promises 'rigorous theoretical analysis,' but Section 3 contains no theorem, no proof, and no complexity derivation; the linear-complexity claim is asserted. There is no code, and all headline numbers come from a single run with no variance or multi-seed reporting. That combination makes the paper non-reproducible as is.\n\nOne specific point: the stress-test note's claim of internal inconsistency between Table 1 and Table 2 doesn't hold up. Table 1 uses 300k iterations for the 6-frame model; Table 2 uses 200k. So the 31.95 vs 31.82 gap is expected from training length, not a contradiction. The real weakness is that no ablation at the same iteration count links the 31.95 number to the CAS mechanism, so the attribution of the headline gain to content-aware scanning is weaker than the text implies.\n\nBottom line: this is a promising idea, but the paper is a sketch. It deserves a serious referee—the concept is novel enough and the results are good enough to warrant the time—but the revision bar should be high: full mathematical specification, complexity analysis, code release, and multi-seed evaluation. I'd bring it to reading group to discuss the mechanism, but I wouldn't cite it until it's reproducible.\n\nRecommendation: send to peer review, yes, but expect major revision.","headline":"The content-aware scanning mechanism is genuinely novel, but the paper is too under-specified to reproduce and the theoretical analysis promised in the contributions is not delivered.","tokens_in":12427,"tokens_out":3517,"would_cite":false,"duration_ms":29984,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MambaVSR, a state-space video super-resolution model using content-aware scanning, beats transformer-based VRT by 0.58 dB PSNR on REDS4 with 55% fewer parameters.","keywords":["video super-resolution","state space model","content-aware scanning","spectral clustering","Fiedler vector","non-local dependencies","linear complexity","temporal alignment"],"falsifier":"Train MambaVSR on REDS with the SCC-generated scan order replaced by a random but fixed per-frame permutation, keeping everything else identical; if PSNR then matches the raster baseline instead of the reported 0.14 dB (REDS4) and 0.24 dB (Vid4) content-aware gains, the Fiedler ordering itself, not the interleaving or extra parameters, is the cause. As a second check, compute the Spearman rank correlation between Fiedler vectors of adjacent frames in a high-motion clip: a low correlation would show the shared compass is not stable enough to align content across frames.","tokens_in":11245,"feed_emoji":"🎬","tokens_out":6061,"duration_ms":51106,"temperature":0.7,"pith_summary":"The paper claims that video super-resolution can be done better and cheaper by replacing the rigid 1D scanning of state-space models with a content-aware scan order that adapts to each video's structure. It introduces MambaVSR, the first state-space model framework for VSR, which builds a per-frame semantic affinity graph, derives a scan order from it via spectral clustering, and interleaves features from coarsely aligned frames along that order so that similar content from different frames becomes adjacent before the selective SSM processes them. The paper reports that this design outperforms the transformer-based VRT by 0.58 dB PSNR on REDS4 while using 55% fewer parameters, and beats the Swin-transformer PSRT-recurrent by 0.22 dB on Vid4. A sympathetic reader should care because it suggests that the bottleneck in state-space vision models is not the SSM itself but the sequentialization order, and that a content-derived order can give linear-complexity models transformer-level non-local aggregation.","feed_headline":"MambaVSR beats transformer VSR by 0.58 dB with 55% fewer parameters","feed_subtitle":"Content-aware scanning lets a state-space model match transformer-level video quality at less than half the parameters.","key_machinery":"The load-bearing object is the content-aware scanner built from three cooperating modules. Shared Compass Construction (SCC) downsamples features, computes a sparse dual-branch attention similarity matrix, forms a Laplacian, and extracts its Fiedler vector — the eigenvector of the graph Laplacian associated with the second-smallest eigenvalue, whose sorted components give a 1D ordering that approximately groups similar nodes — to produce a shared spatial scan order used across refinement stages. Content-Aware Sequentialization (CAS) then uses patch alignment to reuse this order across frames and interleaves spatial features along it, so temporally matched similar content becomes adjacent in the SSM input sequence. The Global-Local State Space Block (GLSSB) wraps the SSM propagation inside fixed windows and fuses it with window self-attention through a learnable residual parameter, so global dependency guidance and local high-frequency recovery happen in the same block. The Fiedler-vector ordering is the mechanism that converts a spatial similarity graph into a 1D scan the selective SSM can exploit with linear complexity.","core_discovery":"On its own terms, the paper establishes that a selective state-space model can outperform both CNN- and transformer-based video super-resolution methods when the scanning order is made content-aware rather than fixed raster or multi-directional scanning. The central mechanism is that spatially similar patches, identified through sparse attention and spectral clustering, are arranged into a sequence such that the SSM's cumulative state sees globally similar content within a frame, and after coarse patch alignment, the same order interleaves temporally corresponding features across frames. This yields non-local spatiotemporal aggregation with linear complexity, which the paper argues is what enables the reported gains: 32.75 dB PSNR on REDS4 versus 32.17 dB for VRT with 55% fewer parameters, and a 0.24 dB improvement over raster scanning on Vid4 from the content-aware sequentialization alone. The paper also presents LAM-based diffusion index comparisons and temporal-downsampling tests as evidence that the model uses a larger effective receptive field than linear-complexity CNN and Swin-transformer baselines.","pith_inferences":["An untested corollary: the same Fiedler-vector sequentialization could be dropped into other state-space restoration tasks (deblurring, deraining, interpolation) where non-local self-similarity matters, as long as the similarity graph is computed on downsampled features.","One risk the paper leaves open is that the shared compass computed once per stage may drift on clips with rapid scene changes; if the Fiedler order becomes unstable across frames, CAS interleaving could mix unrelated patches. A cheap stability check would be to correlate the scan orders of adjacent frames on high-motion clips.","The ablation numbers suggest a testable prediction: a model using a random but fixed per-frame permutation should perform close to raster scanning, not to content-aware scanning, which would confirm that the ordering, not just the extra parameters, carries the benefit.","If the content-aware ordering acts mainly by placing similar textures adjacent, the method could also benefit single-image super-resolution with internal patch recurrence, where a graph over image patches already exists."],"forward_implications":["If the reported gains hold, state-space models become a viable backbone for video restoration tasks that need long-range spatiotemporal modeling, not just efficiency.","The 0.58 dB lead over VRT at 55% fewer parameters implies that content-aware sequentialization can substitute for global self-attention's quadratic complexity without sacrificing non-local aggregation.","The 0.24 dB gain of content-aware over raster scanning on Vid4 isolates the ordering as the source of improvement, not the SSM or the window attention alone.","Recomputing the scan order only once per stage and reusing it across aligned frames (the shared compass) keeps the overhead of graph construction low enough that the whole model runs at 2.46 T FLOPs per 180×320 frame.","Longer training sequences (16 frames) and larger motion produce bigger gains, suggesting the method's advantage grows with temporal extent."],"supporting_citations":[{"why":"Supplies the selective state-space model (Mamba) with linear complexity that MambaVSR adapts to video via content-aware scanning.","marker":"[7]"},{"why":"VRT is the transformer baseline the paper outperforms by 0.58 dB on REDS4 and the source of temporal mutual attention design it compares against.","marker":"[18]"},{"why":"PSRT-recurrent is the Swin-transformer baseline beaten by 0.22 dB on Vid4, and its training protocol is adopted for fair comparison.","marker":"[26]"},{"why":"BasicVSR++ provides the bidirectional propagation backbone on which MambaVSR's architecture is built.","marker":"[3]"},{"why":"Provides the activating-more-pixels analysis and diffusion index concept used to argue MambaVSR has a larger effective receptive field.","marker":"[4]"},{"why":"Supplies Local Attribution Maps (LAM) used to visualize and quantify which input pixels the model uses for reconstruction.","marker":"[10]"},{"why":"REDS is the main training and evaluation dataset; REDS4 is the benchmark split where the headline PSNR numbers are reported.","marker":"[24]"},{"why":"IART's advanced resampling module is adopted in MambaVSR+ for a fair comparison with the IART baseline.","marker":"[34]"}],"fun_headline_variants":["MambaVSR content-aware scanning beats transformer VSR","State-space VSR with content-aware scanning beats transformer","Content-aware scan order unlocks SSM for video super-resolution","Adaptive scanning state-space model outdoes transformer in VSR","Content-aware state-space VSR tops transformer with fewer params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Fiedler-vector scan order computed on downsampled features from sparse attention graphs is semantically meaningful and stable enough to reuse across frames after coarse alignment, so that similar patches become adjacent in the scan; if that fails, interleaving aligned features along the order would not aggregate similar content and the claimed 0.24 dB gain over raster scanning would vanish.","fun_headline_variants_meta":{"raw":{"variants":["MambaVSR content-aware scanning beats transformer VSR","State-space VSR with content-aware scanning beats transformer","Content-aware scan order unlocks SSM for video super-resolution","Adaptive scanning state-space model outdoes transformer in VSR","Content-aware state-space VSR tops transformer with fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001224,"raw_usage":{"total_tokens":5057,"prompt_tokens":996,"completion_tokens":4061,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":3980}},"tokens_in":612,"tokens_out":4061,"duration_ms":29023,"temperature":1.0,"reasoning_tokens":3980,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:03:17.746130+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MambaVSR on REDS with the SCC-generated scan order replaced by a random but fixed per-frame permutation, keeping everything else identical; if PSNR then matches the raster baseline instead of the reported 0.14 dB (REDS4) and 0.24 dB (Vid4) content-aware gains, the Fiedler ordering itself, not the interleaving or extra parameters, is the cause. As a second check, compute the Spearman rank correlation between Fiedler vectors of adjacent frames in a high-motion clip: a low correlation would show the shared compass is not stable enough to align content across frames.","supporting_citations":[{"cited_title":"VRT: A video restoration transformer.IEEE Transactions on Image Processing, 33:2171–2182, 2024","cited_arxiv_id":null,"evidence_quote":"VRT is the transformer baseline the paper outperforms by 0.58 dB on REDS4 and the source of temporal mutual attention design it compares against."},{"cited_title":"Rethinking alignment in video super- resolution transformers","cited_arxiv_id":null,"evidence_quote":"PSRT-recurrent is the Swin-transformer baseline beaten by 0.22 dB on Vid4, and its training protocol is adopted for fair comparison."},{"cited_title":"BasicVSR++: Improving video super- resolution with enhanced propagation and alignment","cited_arxiv_id":null,"evidence_quote":"BasicVSR++ provides the bidirectional propagation backbone on which MambaVSR's architecture is built."},{"cited_title":"Activating more pixels in image super- resolution transformer","cited_arxiv_id":null,"evidence_quote":"Provides the activating-more-pixels analysis and diffusion index concept used to argue MambaVSR has a larger effective receptive field."},{"cited_title":"Interpreting super-resolution networks with local attribution maps","cited_arxiv_id":null,"evidence_quote":"Supplies Local Attribution Maps (LAM) used to visualize and quantify which input pixels the model uses for reconstruction."},{"cited_title":"NTIRE 2019 challenge on video deblurring and super- resolution: Dataset and study","cited_arxiv_id":null,"evidence_quote":"REDS is the main training and evaluation dataset; REDS4 is the benchmark split where the headline PSNR numbers are reported."},{"cited_title":"Enhancing video super-resolution via im- plicit resampling-based alignment","cited_arxiv_id":null,"evidence_quote":"IART's advanced resampling module is adopted in MambaVSR+ for a fair comparison with the IART baseline."}],"review_version":1}