{"id":"f811726c-4c75-4bcf-a9ed-80c634ee5b45","arxiv_id":"2505.19668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A Transformer-Mamba hybrid with multi-cross attention and multi-scan state-space fusion reports state-of-the-art PSNR and SSIM on synthetic burst super-resolution benchmarks.","lead":"A burst-super-resolution team combined overlapping window attention, cross-frame attention, and a state-space model to turn a burst of shaky, noisy phone images into one sharper high-resolution photo. On synthetic benchmark data they report the best peak signal-to-noise ratio so far, though the real-world ranking depends on how baseline models were retrained.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-world SOTA claim is contradicted by Table 1: Ours PSNR 48.60 is below BurstM 48.65/49.12 and AFCNet 48.63, and SSIM 0.986 is below BurstM/SBFBurst 0.987, so 'highest SSIM' and 'second in PSNR' are unsupported.","rationale":"The reader's conditional verdict is well calibrated. The architecture is plausible, and the synthetic comparison is internally consistent: Ours has the best PSNR/SSIM/LPIPS in Table 1's synthetic block, and the ablations show the proposed modules improve over the Swin baseline. The load-bearing weakness is the real-world ranking claim. Section 4.3 says Ours achieves the highest SSIM and LPIPS and ranks second in PSNR, but the table shows BurstM and SBFBurst with SSIM 0.987 at patch 80, above Ours 0.986, and several methods above Ours in PSNR (BurstM 48.65 at 64, AFCNet 48.63 at 80, Burstormer 48.82 at 80, SBFBurst 48.87 at 80, BurstM 49.12 at 80). The only reading that makes the text true is to restrict the comparison to methods retrained at patch 64, but that restriction is not stated in the abstract or in the unqualified sentence in Section 4.3. Because the paper's headline claim includes real-world SOTA, this is a substantive defect. It does not destroy the synthetic contribution, but it requires correction: qualification of the real-world claim, a clear comparison protocol, and ideally a controlled retraining of all baselines at a common patch size with variance estimates. Therefore the disposition should remain conditional rather than accept or reject.","tokens_in":17840,"tokens_out":7861,"duration_ms":78907,"concrete_test":"Sort the three real-world columns of Table 1 by PSNR, SSIM, and LPIPS, including entries marked (80) and (64). If Ours is not strictly first in SSIM and not strictly second in PSNR under the full table, revise the abstract and Section 4.3 to state the restricted comparison (e.g., best among models retrained at patch size 64). Then rerun the real-world evaluation with all methods trained at patch size 80 under identical settings, reporting mean and standard deviation over three seeds; if the claimed ranks do not hold under that controlled protocol, withdraw or explicitly limit the real-world SOTA claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: synthetic and real-world SOTA. The synthetic part is internally consistent: in Table 1, Ours (43.20 PSNR, 0.972 SSIM, 0.028 LPIPS) beats all listed synthetic entries, including SeBIR (42.86/0.967/0.035). The vulnerable part is the real-world claim. Section 4.3 states that 'in the real-world dataset, it achieves the highest scores for SSIM and LPIPS, while ranking second in PSNR,' and the abstract repeats 'highest scores for SSIM and LPIPS.' Table 1 does not support this. Sorting the real-world columns shows Ours PSNR 48.60 is lower than BurstM 48.65 (same patch 64), AFCNet 48.63 (80), Burstormer 48.82 (80), SBFBurst 48.87 (80), and BurstM 49.12 (80). Ours SSIM 0.986 is lower than BurstM 0.987 (80) and SBFBurst 0.987 (80); it is only tied at the 64-patch setting. Thus, under a straightforward reading of the table, the 'highest SSIM' and 'second in PSNR' statements are false unless the comparison is restricted to 64-patch retrained models, a restriction the text does not state. Because the real-world SOTA claim is part of the paper's headline, this is a load-bearing weakness: the claimed ranking does not follow from the evidence presented. The issue is not that the architecture is invalid; the synthetic result and ablations give it credibility. The problem is that the paper overstates its real-world standing, mixing published patch-80 numbers with retrained patch-64 numbers and then asserting ranks that do not match the table.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a burst super-resolution network that combines a Multi-Cross Attention (MCA) encoder, built from overlapping cross-window attention and cross-frame attention, with a decoder based on a Multi-Scan State-Space Module (MS-SSM) and cross-frame attention, while retaining DCN-based alignment. The method is evaluated on the NTIRE2022 synthetic and real-world burst-SR benchmarks and on an ISO 12233 resolution chart. The central claims are that the method achieves state-of-the-art results on the synthetic dataset across PSNR, SSIM, and LPIPS, and that on the real-world dataset it achieves the highest SSIM and LPIPS with second-best PSNR.","tokens_in":18318,"tokens_out":8404,"duration_ms":83153,"significance":"If the reported synthetic result is reproducible, the proposed architecture is a credible integration of window attention, cross-frame attention, and state-space models for burst super-resolution, offering competitive performance at a moderate parameter count (12.60M). The synthetic comparison in Table 1 supports the claim of improved PSNR/SSIM/LPIPS over the listed baselines, including SeBIR. The real-world claims, however, are not supported by the table as presented, because the comparison mixes published patch-80 numbers with retrained patch-64 numbers and the text asserts rankings that the table does not show. The ISO 12233 experiment is a useful practical extension, but its protocol is not described in enough detail to verify. Overall, the work is of moderate significance for the burst-SR community, provided the comparison claims are corrected and the retraining protocol is documented.","major_comments":[{"comment":"The real-world ranking claims are not supported by Table 1. The text states that on the real-world dataset the method 'achieves the highest scores for SSIM and LPIPS, while ranking second in PSNR,' and the abstract repeats this. Sorting the real-world columns shows that Ours (PSNR 48.60, SSIM 0.986, LPIPS 0.021) is below BurstM at patch 64 (48.65 PSNR) and below several patch-80 entries (BurstM 49.12, SBFBurst 48.87, Burstormer 48.82, AFCNet 48.63). For SSIM, Ours 0.986 is below BurstM (0.987) and SBFBurst (0.987), and for LPIPS Ours 0.021 is tied with BSRT (0.021) and SeBIR (0.021), not strictly best. The claims are therefore only defensible if restricted to the 64-patch retrained models, but the text does not state that restriction. This is a load-bearing issue because the abstract, §4.3, and the conclusion all repeat the real-world state-of-the-art claim. The authors should either revise the claims to match the table or provide a consistent comparison protocol that justifies the stated ranking.","section":"Table 1 caption, §4.3"},{"comment":"The retraining protocol for the entries labeled '(64)' is not adequately specified. The caption says these values are obtained by retraining on the real-world dataset with the same 64-patch size as the proposed method, 'based on the open-sourced code from the paper,' but the manuscript does not report the retraining epochs, learning-rate schedule, loss function (aligned or not), data splits, or any other details needed to reproduce those numbers. Because the real-world comparison mixes published patch-80 numbers with retrained patch-64 numbers in a single table, the ranking of methods across these columns is not interpretable without knowing whether the retraining was faithful and comparable. The authors should either report full retraining details or present patch-80 and patch-64 results in separate, clearly labeled comparisons.","section":"Table 1 caption / §4.2"},{"comment":"The ablation study reports quantitative gains of 0.78 dB (MCA versus baseline) and 0.11 dB (full model versus MCA only) but does not provide a table with final PSNR, SSIM, or LPIPS values for the baseline, MCA-only, RMB-only, and full model. Only training curves are shown, and the sole numeric value, '≈43.07' for the RMB-only model, appears in the caption of Fig. 13. Without a numeric ablation table, the claimed contributions of the individual modules cannot be verified. Please add a table with final metrics and parameter counts for each ablated configuration.","section":"§4.5"},{"comment":"The ISO 12233 resolution-chart experiment does not describe a controlled comparison protocol. It states that low-resolution images were upsampled using 'several state-of-the-art BurstSR methods' but does not say whether those methods were retrained or fine-tuned on the same short-wave infrared data as the proposed method. The conclusion later asserts that 'the results, obtained using the same training data,' but this is not described in §4.4. Without a clear statement of how the competing models were adapted to the SWIR domain, the reported LP/mm differences cannot be attributed to the proposed architecture rather than to training conditions.","section":"§4.4"}],"minor_comments":[{"comment":"The manuscript header reads 'Preprint submitted to Nuclear Physics B,' which is not an appropriate journal for a computer-vision paper; this appears to be a template error and should be corrected.","section":"Header"},{"comment":"Equation numbering is inconsistent: Eq. (1) is used twice, first for Fd = E(F0) in §3 and again for the four equations inside the encoder description in §3.1. The equations in §3.1 should be renumbered sequentially.","section":"§3, §3.1"},{"comment":"The loss in Eq. (10) is written as L = ||ISR − IHR|| without specifying the norm; given the text mentions L1 loss, the equation should state the L1 norm explicitly and clarify whether the aligned L1 loss is used for real-world training.","section":"Eq. (10)"},{"comment":"The patch-size description is confusing: 'For synthetic training, we used a patch size of 384 × 384 for high-resolution images. However, due to memory constraints, the patch size for low-resolution images was reduced to 64 × 64 during training on real-world data.' Please clarify which patch size applies to synthetic versus real-world training and the relationship between the HR and LR patch sizes.","section":"§4.2"},{"comment":"The ablation text says the orange line uses 'the proposed MCA modules in the decoder,' but MCA is described as an encoder module; this is likely a typo and should be corrected. Also, the baseline 'original BSRT' should be described precisely so the reader knows which configuration is used as the blue line.","section":"§4.5"},{"comment":"The conversion factor of 9.26 used to convert LW/PH to LP/mm is stated without derivation. Please provide the formula and the sensor parameters used, so the reported resolution values can be checked.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid synthetic result and a plausible architecture, but the real-world state-of-the-art claim is not supported by the table as presented. The mixed patch-size comparison and the insufficiently documented retraining protocol are the main obstacles. The ISO 12233 section also needs a clearer protocol. These issues are fixable with a careful revision; I do not see a reason for rejection. The 'Nuclear Physics B' header suggests a submission-template error that the editor may want to flag."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the synthetic result is real and interesting, but the real-world \"SOTA\" claim is not supported by Table 1. The architecture is a sensible mashup of known pieces—overlapping cross-window attention, cross-frame channel gating, and a Mamba-style multi-scan SSM decoder—and the ablations show each piece earns its keep. The 43.20 dB PSNR on the synthetic benchmark beats the listed prior best by 0.34 dB, and the training curves (Figs. 12-13) show the full model converging better than the RMB-only variant. The ISO 12233 chart experiment is a nice, if small, addition: real captured bursts, clear resolution measurements, and the proposed method wins on all three directions. That is real work and worth acknowledging.\n\nThe soft spot is the real-world table and the language around it. The paper says Ours \"achieves the highest scores for SSIM and LPIPS, while ranking second in PSNR\" on the real-world dataset. Looking at Table 1, Ours PSNR is 48.60, which is below BurstM's 48.65 (same 64-patch retrain) and below Burstormer 48.82, SBFBurst 48.87, AFCNet 48.63, and BurstM 49.12 (all at patch 80). Ours SSIM is 0.986, tied with several methods at patch 80 and below BurstM and SBFBurst at 0.987. LPIPS is 0.021, tied with BSRT and SeBIR. So \"highest SSIM\" is false under any unrestricted reading, and \"second in PSNR\" is only true if you restrict to the 64-patch retrained models—which the text never says. The footnote explains the (64) values are retrained, but the claim itself is not scoped. That is a load-bearing overstatement because it appears in the abstract.\n\nThe deeper protocol issue is the mixing of patch sizes and published numbers in one table. Synthetic numbers are copied from papers; real-world numbers are a mix of published patch-80 and retrained patch-64 results. That makes the table hard to read and the ranking fragile. The fix is straightforward: retrain all baselines at the same patch size, report mean±std across seeds, and release code and hyperparameters. The synthetic result and ablations suggest the architecture is genuinely competitive; the paper just needs to present the real-world comparison honestly.\n\nThe citation pattern looks fine—prior work (BSRT, Burstormer, BurstM, MambaIR) is credited appropriately. No formal verification or released code, so this is an empirical paper only.\n\nWho is this for? People building burst-SR backbones. It deserves a serious referee, but with a clear brief to check the evaluation protocol. I would ask for major revision, not desk reject.","headline":"A plausible burst-SR architecture with a genuine synthetic SOTA claim, but the real-world SOTA claims in the abstract and Section 4.3 are contradicted by Table 1.","tokens_in":18819,"tokens_out":3096,"would_cite":false,"duration_ms":29198,"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":"A burst super-resolution architecture that pairs overlapping cross-window and cross-frame attention with a multi-scan state-space decoder claims state-of-the-art results on the NTIRE2022 benchmarks.","keywords":["burst super-resolution","multi-image super-resolution","multi-cross attention","cross-window attention","cross-frame attention","state-space model","Mamba","RAW burst images"],"falsifier":"Retraining every compared method under identical settings (patch size 64, same epochs, same aligned-L1 loss, same data) and recomputing the real-world PSNR column would settle the claim; as printed, Table 1 already shows BurstM at 48.65 and AFCNet at 48.63, both above the paper's 48.60, so the asserted 'second in PSNR' ranking does not follow from the table.","tokens_in":17649,"feed_emoji":"📷","tokens_out":6851,"duration_ms":43117,"temperature":0.7,"pith_summary":"Burst super-resolution aims to recover a high-resolution image from a short burst of noisy, slightly misaligned RAW frames, using the sub-pixel shifts between frames as extra information. This paper proposes an encoder-decoder network in which a Multi-Cross Attention feature extractor combines overlapping cross-window attention with cross-frame attention, and a Multi-Scan State-Space decoder fuses the aligned features while correcting residual misalignment left by deformable-convolution alignment. On the NTIRE2022 synthetic benchmark the paper reports 43.20 dB PSNR, beating the prior best SeBIR by 0.34 dB, and it reports the best SSIM and LPIPS on the real-world benchmark. The authors also evaluate on an ISO 12233 resolution chart with a SWIR camera and report higher resolvable line pairs per millimeter than Burstormer and BSRT. The significance, if the comparisons hold, is that combining windowed attention with linear-complexity state-space scanning is a practical route to extracting and fusing sub-pixel detail across frames.","feed_headline":"Burst-SR network beats prior best by 0.34 dB PSNR","feed_subtitle":"Overlapping cross-window and cross-frame attention plus multi-scan SSM fusion improve fine-detail reconstruction in burst photos.","key_machinery":"The paper's central objects are Multi-Cross Attention (MCA) and the Multi-Scan State-Space Module (MS-SSM). MCA combines overlapping Cross-Window Attention, in which the key/value windows are larger than the query window so attention reaches beyond the fixed window, with Cross-Frame Attention, a lightweight channel-compressed convolutional gating block that pools global frame information. MS-SSM is a four-direction scan of the flattened feature sequence inside a Residual Mamba Block, so long-range dependencies are modeled with linear complexity while a local convolution restores neighborhood regularity after the 1D scan. Together they carry the argument: CWA and CFA widen the receptive field during feature extraction, and MS-SSM plus CFA compensates for the limited receptive field of the DCN alignment in the fusion stage.","core_discovery":"The central claim is that the standard burst-SR pipeline of feature extraction, flow-guided deformable alignment, fusion, and reconstruction can be improved by replacing the Swin Transformer stages with two purpose-built modules. In the encoder, Multi-Cross Attention (MCA) runs overlapping Cross-Window Attention and Cross-Frame Attention in parallel: the query uses a $P\\times P$ window while keys and values use overlapping $P'\\times P'$ windows, so each query token can attend to a wider neighborhood and to information from other frames. In the decoder, a Multi-Scan State-Space Module (MS-SSM) scans the fused features in four directions inside a Residual Mamba Block, and a Cross-Frame Attention block then re-weights the frame-wise information. The paper reports that this architecture achieves 43.20 PSNR, 0.972 SSIM, and 0.028 LPIPS on the synthetic NTIRE2022 dataset, ahead of all compared methods, and 48.60 PSNR, 0.986 SSIM, 0.021 LPIPS on the real-world dataset, where it says it is first in SSIM and LPIPS and second in PSNR. The ablation attributes 0.78 dB of the synthetic gain to MCA and a further 0.11 dB to the RMB fusion module.","pith_inferences":["The attention recipe may transfer to other multi-frame low-level vision tasks: any alignment-plus-fusion pipeline that suffers from fixed-window attention could use overlapping cross-window attention and cross-frame gating to pull sub-pixel information from neighboring frames.","Since the multi-scan SSM explicitly scans the dithering range of each pixel, it could be tested as a generic post-alignment error corrector, potentially reducing dependence on accurate optical flow estimates in video restoration.","The real-world PSNR column of Table 1 lists BurstM (48.65) and AFCNet (48.63) above the paper's 48.60, so the 'second in PSNR' statement is an editorial reading of the table rather than a direct numerical consequence.","The ISO 12233 result uses a short-wave infrared camera with 14-frame bursts and controlled 2-pixel dithering; those LP/mm numbers indicate fine-detail recovery in that specific setup and should not be treated as a direct RGB burst-SR measurement."],"forward_implications":["On the synthetic NTIRE2022 burst benchmark, the method claims 43.20 dB PSNR, a 0.34 dB gain over the previous best SeBIR, together with the best SSIM and LPIPS among the compared methods.","On the real-world benchmark, it reports the best SSIM and LPIPS scores, reaching 48.60 dB PSNR while training at a smaller patch size than several baselines.","On the ISO 12233 chart experiments, it reports 66 LP/mm horizontal, 59 LP/mm vertical, and 65 LP/mm diagonal resolution, exceeding the BSRT and Burstormer reconstructions.","The ablation study reports that replacing the Swin extraction with MCA adds 0.78 dB PSNR on synthetic bursts, and adding the MS-SSM-based RMB adds a further 0.11 dB.","The design implies that retaining DCN-based alignment is not a liability when a multi-scan state-space fusion module is available to correct the residual misalignment."],"supporting_citations":[{"why":"Supplies the BSRT baseline, the NTIRE2022 evaluation protocol, and the post-processing scripts used for training.","marker":"[6]"},{"why":"Burstormer is the main CNN/Transformer comparison and provides the aligned-L1 protocol and patch-64 retraining baseline.","marker":"[2]"},{"why":"BurstM is the optical-flow/Fourier-space alternative that the decoder's MS-SSM is designed to outperform.","marker":"[15]"},{"why":"SBFBurst provides the simple base-frame-guidance residual baseline and is retrained at patch 64 for comparison.","marker":"[7]"},{"why":"SeBIR is the previous synthetic-dataset state of the art that the reported 43.20 PSNR is measured against.","marker":"[55]"},{"why":"MambaIR supplies the vision state-space module design on which the Residual Mamba Block is built.","marker":"[20]"},{"why":"Deformable convolution provides the alignment module the paper retains and whose limitations MS-SSM corrects.","marker":"[28]"},{"why":"SpyNet computes the multi-scale optical flows used for coarse alignment before DCN refinement.","marker":"[45]"},{"why":"The NTIRE2022 challenge defines the synthetic and real-world burst datasets and the 4x evaluation task.","marker":"[51]"}],"fun_headline_variants":["Burst SR gains 0.34 dB with overlapping cross-attention","Multi-scan state-space and cross-frame attention lift burst resolution","Overlapping attention windows and MS-SSM improve burst image SR","Burst SR: novel attention and state-space decoding beat prior best","0.34 dB PSNR jump in burst super-resolution from new modules"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported state-of-the-art ranking assumes that all numbers in the comparison table are directly comparable, even though the baselines were trained with different patch sizes (80 versus 64) and some were retrained while others were taken from their original papers.","fun_headline_variants_meta":{"raw":{"variants":["Burst SR gains 0.34 dB with overlapping cross-attention","Multi-scan state-space and cross-frame attention lift burst resolution","Overlapping attention windows and MS-SSM improve burst image SR","Burst SR: novel attention and state-space decoding beat prior best","0.34 dB PSNR jump in burst super-resolution from new modules"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000809,"raw_usage":{"total_tokens":3600,"prompt_tokens":1044,"completion_tokens":2556,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":2464}},"tokens_in":660,"tokens_out":2556,"duration_ms":15927,"temperature":1.0,"reasoning_tokens":2464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:08:43.907981+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retraining every compared method under identical settings (patch size 64, same epochs, same aligned-L1 loss, same data) and recomputing the real-world PSNR column would settle the claim; as printed, Table 1 already shows BurstM at 48.65 and AFCNet at 48.63, both above the paper's 48.60, so the asserted 'second in PSNR' ranking does not follow from the table.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BSRT baseline, the NTIRE2022 evaluation protocol, and the post-processing scripts used for training."},{"cited_title":"Dudhane, S","cited_arxiv_id":null,"evidence_quote":"Burstormer is the main CNN/Transformer comparison and provides the aligned-L1 protocol and patch-64 retraining baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BurstM is the optical-flow/Fourier-space alternative that the decoder's MS-SSM is designed to outperform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SBFBurst provides the simple base-frame-guidance residual baseline and is retrained at patch 64 for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SeBIR is the previous synthetic-dataset state of the art that the reported 43.20 PSNR is measured against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Deformable convolution provides the alignment module the paper retains and whose limitations MS-SSM corrects."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The NTIRE2022 challenge defines the synthetic and real-world burst datasets and the 4x evaluation task."}],"review_version":1}