{"id":"0d86e3b8-19c8-4bb5-97dc-bb2855d893c7","arxiv_id":"2508.07140","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A mask-aware transformer U-Net with mask-guided up/down-samplers and a co-feature aggregator reports state-of-the-art mural restoration on MuralDH and Dunhuang benchmarks.","lead":"This paper introduces CMAMRNet, a neural network that restores damaged murals by keeping the damage mask information active at every resolution level of the network. On two mural benchmarks it reports better restoration scores than twelve existing inpainting methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central SOTA claim is unverified because mural-specific baselines (e.g., MuralDiff [24]) are absent from Table 1; adding them could overturn the headline result.","rationale":"The reader correctly identifies that the benchmark's masks and ground truth are a key assumption, but the more directly load-bearing and testable gap is the omitted mural-specific baselines. The paper itself frames mural restoration as distinct from generic inpainting and cites methods designed for it, but the experimental table excludes them. Without those comparisons, 'outperforms state-of-the-art mural restoration' is not established even if the masks are realistic. This is not an internal inconsistency in the architecture; the ablations (Table 2) are consistent with the proposed modules providing gains. The concern is about the completeness and fairness of the empirical comparison, plus the lack of variance information for small reported margins. Since this is an addressable experimental gap rather than a demonstrated error, the existing CONDITIONAL verdict remains appropriate, so I keep the verdict unchanged.","tokens_in":8984,"tokens_out":5381,"duration_ms":54155,"concrete_test":"Re-run the comparison in Table 1 on the same MuralDH and Dunhuang train/test splits, at the same 512×512 test resolution, including MuralDiff [24] and at least one of the other cited mural-specific methods ([7], [17], or [22]) using their official implementations/checkpoints. Report PSNR, SSIM, MAE, and LPIPS under the authors' evaluation protocol. If any added method matches or exceeds CMAMRNet's numbers, the SOTA claim is unsupported; if CMAMRNet still wins, the omission was not consequential.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that CMAMRNet outperforms state-of-the-art mural restoration. Table 1, however, compares only against general inpainting methods (RN, DeepFillV2, LaMa, HINT, etc.). The paper's own introduction cites mural-specific restoration methods: line-drawing guided inpainting [7], Thanka mural inpainting [17], color-restoration [22], and MuralDiff [24]. In Section 1 the authors explicitly critique diffusion-based mural methods [24] as potentially generating historically inaccurate details, yet MuralDiff is never included in the experiments. Thus the 'previous state-of-the-art HINT' is not established as the prior best mural-restoration method. A direct comparison with MuralDiff and other mural-specific methods is required before the headline SOTA claim can be accepted. The issue is compounded by the absence of error bars or significance tests: the reported margins over HINT are only 0.77 dB PSNR on MuralDH and 0.83 dB on Dunhuang, so even a single omitted strong baseline could change the ranking.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CMAMRNet, a U-shaped Transformer network for digital mural restoration that combines two new components: MAUDS (Mask-Aware Up/Down-Sampler), which propagates mask information through resolution changes via channel selection and mask-guided fusion, and CFA (Co-Feature Aggregator), which fuses channel, spatial, and frequency features at the highest and lowest resolutions. The model is trained with an MSE+SSIM loss and evaluated on the MuralDH and Dunhuang mural datasets. The authors report consistent improvements over twelve general image-inpainting baselines on PSNR, SSIM, MAE, and LPIPS, with the largest gains over HINT, and they present ablations showing that both MAUDS and CFA contribute to the final performance. The code is publicly released.","tokens_in":9321,"tokens_out":4246,"duration_ms":42224,"significance":"If the reported results are robust, the central idea—maintaining explicit mask guidance through every sampling stage rather than letting masks attenuate through the network—is a sensible and potentially useful contribution to cultural-heritage restoration. The architecture is clearly described, the ablations indicate that both proposed modules help, and the release of code is a practical strength for reproducibility. However, the headline claim of state-of-the-art mural restoration is not yet substantiated: the comparison omits mural-specific methods cited in the paper, and no uncertainty quantification or statistical tests are provided. As it stands, the paper demonstrates a well-engineered inpainting method on two mural benchmarks, but the stronger claim of superiority over prior mural-restoration approaches requires additional evidence.","major_comments":[{"comment":"The quantitative comparison includes only general image-inpainting methods. The Introduction explicitly discusses mural-specific restoration techniques, including line-drawing guided inpainting [7], Thanka mural inpainting [17], color restoration [22], and MuralDiff [24], the last of which is critiqued at length. None of these appears in Table 1. Therefore, the statement that HINT is the 'previous state-of-the-art' and that CMAMRNet surpasses it is not established for mural restoration. The reported margins over HINT are modest (0.77 dB PSNR on MuralDH, 0.83 dB on Dunhuang), so adding MuralDiff or other mural-specific baselines could change the ranking. A direct comparison is required before the central SOTA claim can be accepted.","section":"§3.3, Table 1"},{"comment":"All quantitative results appear to come from a single training run, with no error bars, multiple seeds, or significance tests. The differences over HINT are only about 0.8 dB PSNR, which may be within run-to-run variability for deep inpainting models. Similarly, the ablation differences in Table 2 (e.g., row (1) vs. row (3)) need confidence intervals. Please report mean±std over at least three seeds for the main comparisons and ablations, and preferably a paired significance test, so readers can judge whether the reported gains are reliable.","section":"§3.2, §3.3, Tables 1–2"},{"comment":"The paper does not specify how the deterioration masks in MuralDH and Dunhuang are generated or annotated, nor whether they represent real mural damage patterns. If the masks are synthetic or based on simplified stroke simulation, the improvements may reflect performance on artificial degradation rather than authentic mural restoration. Please clarify the mask provenance in the datasets, and ideally include an evaluation on real damage masks or discuss the extent to which the benchmark masks are representative. This is important for the external validity of the central claim that CMAMRNet improves mural restoration specifically.","section":"§3.1"}],"minor_comments":[{"comment":"The caption of Figure 2 labels (b) as 'Mask-Aware Down-Sampler (MADS)' and (c) as 'Mask-Aware Up-Sampler (MAUS)', but the text states that the MAUS upsampling is 'As illustrated in Fig. 2 (b)' and the MADS downsampling is 'As depicted in Fig. 2 (c)'. These references are reversed; please correct the figure or the text.","section":"§2.2, Figure 2"},{"comment":"The interleaving notation in Eq. (7) is difficult to follow: the use of 'Finter,2i−1' and 'Finter,2i' to denote individual channels, with the mask index 'j = i mod 4', is unconventional. Please define the channel indexing more explicitly, e.g., with explicit loops or a clearer channel-assembly description.","section":"§2.2, Eq. (7)"},{"comment":"Minor typographical and style issues: 'Corresponging author' in the footnote, 'GeLu' in Figure 2 should be 'GELU', 'Misf' in Table 1 should be 'MISF', and the figure label 'EdgeConn' should be expanded consistently with the main text.","section":"General"},{"comment":"The use of 'mix-up' in data augmentation is mentioned without detail. It is unclear how mix-up is applied when a corresponding mask is required. Please specify whether the mask is also mixed, or clarify the augmentation protocol.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically coherent and the proposed modules are plausible, but the main empirical claim is currently over-stated relative to the evidence. The missing mural-specific baselines and lack of statistical validation are fixable within the scope of a revision, so I do not see this as a rejection. However, if the authors cannot include MuralDiff and other mural-specific methods, or if their results change substantially under multiple seeds, the central SOTA claim should be substantially softened. The dataset-representativeness question about mask provenance should also be addressed head-on, since it bears on whether the benchmarks actually measure mural restoration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a competent engineering paper, and the core architectural idea is legitimate. MAUDS is a clean way to keep mask information alive through resolution changes—channel selection plus pixel shuffle at up/down transitions is simple and plausible—and CFA at the two extreme resolutions is a reasonable complement. The ablation on MuralDH shows both pieces matter: removing MAUDS costs about 0.81 dB PSNR, removing CFA costs about 0.32 dB, and the full model wins on all four metrics. That is real evidence the design does something.\n\nThe paper's main claim, however, is not yet supported. Table 1 compares against general inpainting methods only. The intro cites mural-specific methods—line-drawing guided inpainting, Thanka inpainting, color restoration, and MuralDiff—and explicitly critiques MuralDiff, but none of them appear in the benchmark. Since margins over HINT are only 0.77 dB and 0.83 dB PSNR, a single omitted strong baseline could change the ranking. That is not a minor omission; it is the difference between \"SOTA for mural restoration\" and \"SOTA among the non-diffusion methods we picked.\" The authors should either include those baselines or narrow the claim.\n\nThe other weaknesses are proportional. No error bars or multiple seeds, so we cannot tell whether the gaps are stable. The code link has no commit hash or config files, which makes reproducibility partial. There are small notation issues in Eq. (7)—the interleaving indices don't quite parse—but they don't affect the observed results. The loss weight lambda is tuned on the benchmark, which is standard training, not a circularity problem. And I wouldn't lean on the \"masks are synthetic\" worry: the datasets are described as having real pixel-level deterioration masks, and testing is on held-out images.\n\nOverall: a solid architecture contribution for cultural heritage inpainting, with an unverified SOTA claim. It deserves peer review, not desk rejection. For the referee, I'd ask for mural-specific comparisons and some estimate of variance.\n\nRecommendation: send it out.","headline":"Solid mask-aware inpainting architecture with a real benchmark gap: the SOTA claim doesn't land until mural-specific baselines and variance are reported.","tokens_in":9738,"tokens_out":2231,"would_cite":true,"duration_ms":22893,"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":"The paper claims CMAMRNet, a mask-aware U-shaped transformer with Mask-Aware Up/Down-Samplers and Co-Feature Aggregators, outperforms twelve baselines on Dunhuang mural restoration benchmarks.","keywords":["mural restoration","image inpainting","mask-aware network","transformer","MAUDS","Co-Feature Aggregator","Dunhuang murals","cultural heritage"],"falsifier":"Run CMAMRNet on murals whose damage masks come from a different annotator or an automated crack detector while keeping training data fixed; if PSNR and SSIM drop sharply relative to the dataset's own masks, then the gains depend on the oracle mask distribution rather than on general mask guidance. A stronger test is expert evaluation of restorations against a known historical intact state, where synthetic-target metrics provide no ground truth.","tokens_in":8952,"feed_emoji":"🖼️","tokens_out":6930,"duration_ms":66315,"temperature":0.7,"pith_summary":"The paper argues that mural restoration improves when the deterioration mask is kept explicit at every resolution transition rather than being fed once at the input. It proposes CMAMRNet, a U-shaped transformer where Mask-Aware Up/Down-Samplers carry the mask through pixel-shuffle and channel-interleaving paths at every scale change, and Co-Feature Aggregators combine channel- and spatial-attention features at the finest and coarsest scales. On the MuralDH and Dunhuang benchmarks, the full model reports the best PSNR, SSIM, MAE, and LPIPS among twelve baselines, surpassing the previous best method. If these benchmark results hold, the practical payoff is a restoration pipeline that keeps attention pinned to genuinely damaged regions while preserving fine artistic structure, which matters for digital cultural-heritage conservation.","feed_headline":"Mask-aware network beats 12 baselines in mural restoration","feed_subtitle":"CMAMRNet keeps the damage map alive at every scale, improving PSNR, SSIM, MAE, and LPIPS on Dunhuang murals","key_machinery":"The load-bearing machinery is the pair of modules the paper names Mask-Aware Up/Down-Sampler (MAUDS) and Co-Feature Aggregator (CFA). MAUDS sits at every resolution transition in the U-shaped transformer: the Down-Sampler pixel-unshuffles the binary mask into four channels and interleaves those mask channels between feature channels before a depthwise convolution, while the Up-Sampler pixel-shuffles the mask and adds it to the odd channels of the upsampled features. CFA runs in parallel at the highest and lowest resolutions, passing image-plus-mask through two MaxViT blocks and then through Channel and Spatial Feature Focusing Blocks that use FFT to capture frequency patterns; the mask branc","core_discovery":"CMAMRNet's central claim is that persistent mask guidance, not just mask-conditioned input or loss weighting, is what makes mural inpainting work. The network maintains a binary deterioration mask alongside image features through every down- and up-sampling step, so the mask cannot be diluted by convolutions. The Mask-Aware Down-Sampler interleaves mask channels with selected feature channels after pixel unshuffle, while the Mask-Aware Up-Sampler adds a pixel-shuffled mask to odd-channel-selected features before depthwise refinement. At the coarsest and finest scales, the Co-Feature Aggregator uses parallel channel- and spatial-focusing blocks with FFT to modulate features by a mask-derived","pith_inferences":["An implication the authors leave implicit is that MAUDS propagates whatever mask it is given, so on real murals the quality of the damage annotation may matter as much as network capacity; a natural test is feeding masks from an automated crack detector and measuring the drop.","The channel-interleaving mechanism in the Mask-Aware Down-Sampler is generic: it could transfer to other image-to-image tasks with a binary prior, such as shadow removal, reflection suppression, or document cleanup.","The paper's quantitative metrics cannot fully capture historical authenticity, so a testable extension is expert art-historian evaluation of restored mural details against real conservation standards."],"forward_implications":["On MuralDH, the complete model reaches PSNR 36.2565 and SSIM 0.9683, beating the previous best method by about 0.77 dB in PSNR while also lowering MAE and LPIPS.","On Dunhuang, it reaches PSNR 33.0738 and SSIM 0.8944, a larger relative gain over the prior best than on MuralDH, suggesting mask guidance is especially valuable on the harder benchmark.","Ablations show that removing MAUDS drops PSNR by about 0.81 dB and raises MAE and LPIPS, while removing CFA also degrades perceptual quality, indicating both components contribute and work complementarily.","Because the loss is only MSE plus SSIM and training uses 256x256 crops, the reported gains come from architecture rather than heavy supervision, implying the module design may transfer to other image-restoration settings."],"supporting_citations":[{"why":"Supplies the MuralDH dataset with training/test images, pixel-level damage annotations, and deterioration masks used for the main benchmark.","marker":"[23]"},{"why":"Supplies the Dunhuang Grottoes dataset used as the second benchmark with deterioration masks.","marker":"[28]"},{"why":"Provides the Restormer transformer blocks used as the U-shaped encoder-decoder backbone for multi-scale feature representation.","marker":"[29]"},{"why":"Supplies the MaxViT blocks used inside the Co-Feature Aggregator to process concatenated image and mask features.","marker":"[16]"},{"why":"Is the prior state-of-the-art method HINT that the paper compares against and supersedes on both datasets.","marker":"[4]"},{"why":"Defines the LPIPS perceptual metric used to measure perceptual quality in the evaluation.","marker":"[30]"}],"fun_headline_variants":["CMAMRNet outperforms 12 baselines via comprehensive mask guidance","Mask-aware network restores murals by keeping damage map alive","Persistent mask guidance at every scale: CMAMRNet restores murals","Mural restoration without mask dilution: CMAMRNet","Mask-aware multi-scale sampling keeps murals intact: CMAMRNet"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the deterioration masks and intact ground-truth images in the MuralDH and Dunhuang datasets faithfully represent real mural damage and authentic restoration targets; if the masks are synthetic or unrepresentative, the reported benchmark gains measure inpainting on artificial masks rather than restoration of real murals.","fun_headline_variants_meta":{"raw":{"variants":["CMAMRNet outperforms 12 baselines via comprehensive mask guidance","Mask-aware network restores murals by keeping damage map alive","Persistent mask guidance at every scale: CMAMRNet restores murals","Mural restoration without mask dilution: CMAMRNet","Mask-aware multi-scale sampling keeps murals intact: CMAMRNet"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001561,"raw_usage":{"total_tokens":6089,"prompt_tokens":777,"completion_tokens":5312,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":5219}},"tokens_in":521,"tokens_out":5312,"duration_ms":36890,"temperature":1.0,"reasoning_tokens":5219,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:17:46.292326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CMAMRNet on murals whose damage masks come from a different annotator or an automated crack detector while keeping training data fixed; if PSNR and SSIM drop sharply relative to the dataset's own masks, then the gains depend on the oracle mask distribution rather than on general mask guidance. A stronger test is expert evaluation of restorations against a known historical intact state, where synthetic-target metrics provide no ground truth.","supporting_citations":[{"cited_title":"A comprehensive dataset for digital restoration of dunhuang murals","cited_arxiv_id":null,"evidence_quote":"Supplies the MuralDH dataset with training/test images, pixel-level damage annotations, and deterioration masks used for the main benchmark."},{"cited_title":"Dunhuang Grottoes Painting Dataset and Benchmark","cited_arxiv_id":"1907.04589","evidence_quote":"Supplies the Dunhuang Grottoes dataset used as the second benchmark with deterioration masks."},{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Provides the Restormer transformer blocks used as the U-shaped encoder-decoder backbone for multi-scale feature representation."},{"cited_title":"Maxvit: Multi-axis vision transformer","cited_arxiv_id":null,"evidence_quote":"Supplies the MaxViT blocks used inside the Co-Feature Aggregator to process concatenated image and mask features."},{"cited_title":"Hint: High-quality inpainting transformer with mask-aware encoding and enhanced attention.IEEE Trans- actions on Multimedia, 2024","cited_arxiv_id":null,"evidence_quote":"Is the prior state-of-the-art method HINT that the paper compares against and supersedes on both datasets."},{"cited_title":"The unreasonable effectiveness of deep features as a perceptual metric","cited_arxiv_id":null,"evidence_quote":"Defines the LPIPS perceptual metric used to measure perceptual quality in the evaluation."}],"review_version":1}