{"id":"0c0722bf-5f69-4b55-bac6-b88016a51901","arxiv_id":"2507.17892","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DiNAT-IR, a transformer that alternates local and dilated neighborhood attention and adds a channel-aware module, matches or exceeds state-of-the-art image restoration on several benchmarks, particularly motion deblurring.","lead":"This paper presents DiNAT-IR, an image restoration network that pairs dilated neighborhood attention with a lightweight channel-aware module. It reports top or near-top results on deblurring, defocus deblurring, deraining, and denoising benchmarks, including the best GoPro motion deblurring PSNR listed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Checkpoint selection for DPDD, deraining, and denoising is not validated against a held-out split; with margins as small as 0.01 dB on DPDD, the multi-benchmark claim may rest on test-set peeking.","rationale":"I read the paper as an empirical architecture study: Restormer's U-Net backbone with alternating neighborhood and dilated neighborhood attention plus a channel-wise gating module, evaluated on deblurring, defocus deblurring, deraining, and denoising. The strongest evidence is Table 1, where DiNAT-IR reports 33.80 dB on GoPro and 31.57 dB on HIDE. The reader's weakest assumption identifies the missing validation protocol for checkpoint selection; I agree that this is the most concrete threat, but I regard it as only partially load-bearing because the GoPro model is the last iteration and HIDE is not tuned, so the headline motion-deblurring claim is not what the concern attacks. I would not move the verdict to reject: no internal inconsistency in the reported tables was found, Table 5 is an internally consistent small-scale ablation, and the GoPro/HIDE margins over Restormer are large enough that the qualitative direction (local plus dilated neighborhood attention helps deblurring) is plausible. Non-central issues that should be corrected but do not change the verdict: the DiNAT citation points to Hua et al. 2019 (a depth-estimation CNN) instead of Hassani and Shi 2022; Table 1 FLOPs are not 'comparable' to Restormer/NAFNet (45.62 G vs. 35.31 G and 15.85 G), so the conclusion's 'similar or reduced computational costs' is overstated; and the Limitation section admits ablations are GoPro-only, which limits the causal story across tasks. These support the reader's conditional verdict rather than overturning it.","tokens_in":13416,"tokens_out":12039,"duration_ms":117903,"concrete_test":"Retrain each affected task with a fixed validation split held out from the training set (for example, 10% of DPDD training scenes, a standard Rain13K subset, and 10% of SIDD training images), keep the published schedule, and select the final checkpoint by best validation PSNR instead of the reported fixed iterations. Then report test metrics at that checkpoint. If DPDD combined PSNR drops below GRL-BD's 27.04, or Rain13K/SIDD gaps to Restormer widen beyond the text's claims, the multi-benchmark claim is not supported; if the published numbers are reproduced, the checkpoint-selection concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The broad claim that DiNAT-IR is competitive across multiple benchmarks depends on the honesty and protocol of per-task checkpoint selection. Section 4 lists exact chosen iterations for DPDD-D (290K), DPDD-S (140K), Rain13K (40K of fine-tuning), and SIDD (220K) but never mentions a held-out validation split. On DPDD, the headline advance is a 0.01 dB combined-PSNR margin over GRL-BD (27.05 vs. 27.04 in Table 2); if those checkpoints were selected by inspecting test-set curves, that margin could be an artifact of selection rather than architecture. The deraining and denoising results are similarly close to Restormer (e.g., 38.93 vs. 38.99 dB on Rain100L, 39.89 vs. 40.02 dB on SIDD), so late-checkpoint selection could change ordinal claims. The GoPro/HIDE headline is less exposed: the GoPro model is explicitly the last training iteration and HIDE is an untuned generalization evaluation, so the reader's core motion-deblurring claim survives this concern. The vulnerability is specific to the abstract's multi-benchmark claim and to Section 4.2's 'state-of-the-art' wording for DPDD, not to Table 1 alone.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"DiNAT-IR is presented as a Transformer architecture for image restoration. It builds on Restormer's U-Net and replaces the self-attention blocks by alternating Neighborhood Attention (NA, dilation δ=1) and Dilated Neighborhood Attention (DiNA) with stage-dependent dilation pairs, plus a lightweight channel-aware module (CAM) that multiplies the attention output by channel attention computed from global average pooling and a 1D convolution. The paper evaluates the method on motion deblurring (GoPro/HIDE), defocus deblurring (DPDD single and dual pixel), deraining (five benchmarks), and denoising (SIDD). It reports PSNR 33.80/31.57 dB on GoPro/HIDE, 27.05 dB combined PSNR for dual-pixel DPDD, close-to-Restormer deraining results, and 39.89 dB on SIDD. An ablation on a 16-channel GoPro setup shows that CAM improves NA, DiNA, and NA-DiNA variants and that the hybrid NA-DiNA with CAM has the best FID. The paper states that code will be released but does not provide it; the authors also note that the GoPro model may not have converged and that the ablation is limited to GoPro.","tokens_in":13669,"tokens_out":8724,"duration_ms":84085,"significance":"If the reported results are reliable, the contribution is meaningful: a local+global attention mechanism with a cheap channel-aware module reaches state-of-the-art or near-state-of-the-art results on deblurring at comparable parameters and FLOPs. The GoPro/HIDE evaluation is relatively robust because HIDE is an unseen benchmark and the GoPro checkpoint is the last iteration, not a test-set-selected checkpoint. The ablation in Table 5 is internally consistent and supports the design choices, and the candid statements about deraining PSNR and possible GoPro under-convergence are to the authors' credit. The evidence is entirely empirical; no theoretical analysis or code is supplied. The main significance is a new architecture with competitive results, but the multi-benchmark claim needs a stronger evaluation protocol to be fully convincing.","major_comments":[{"comment":"The final checkpoints for DPDD-D (290K iterations), DPDD-S (140K), Rain13K (40K of fine-tuning), and SIDD (220K) are selected by training iteration, but the paper does not state that these choices were made on a held-out validation split. Because the reported margins over the next-best methods are as small as 0.01 dB on DPDD combined PSNR (27.05 vs 27.04 for DiNAT-IRD vs GRL-BD in Table 2), 0.06 dB on Rain100L (38.93 vs 38.99 for Restormer in Table 3), and 0.13 dB on SIDD (39.89 vs 40.02 for Restormer in Table 4), test-set-informed checkpoint selection could change the ordinal claims made in Sections 4.2-4.4 and in the abstract. Please specify the validation protocol, or if none was used, adopt a pre-specified selection rule and report the sensitivity of the conclusions to the choice of checkpoint.","section":"Section 4, Implementation Details; Tables 2-4"},{"comment":"All reported benchmark numbers appear to come from a single training run, with no error bars, confidence intervals, or seed-level statistics. In comparisons where the differences are 0.01-0.13 dB, it is not possible to determine whether the headline improvements (e.g., 33.80 vs 33.69 dB on GoPro in Table 1) are within run-to-run variation. Please report the mean and standard deviation over at least three independent training runs for the proposed method and, if available, for the closest baselines, or otherwise justify why single runs are sufficient for the ordinal claims.","section":"Tables 1-4"},{"comment":"The architecture description does not fully specify the proposed model: the number of Transformer blocks and channel widths per stage, the mapping of the dilation pairs δ∈{1,36},{1,18},{1,9},{1,4} to encoder and decoder levels, the location of the channel-aware module within the block, and the input fusion for the dual-pixel variant DiNAT-IRD are not stated. Since the code is only announced for future release ('Our codes will be released soon'), these omissions make the main architectural claim difficult to verify or reproduce as written.","section":"Section 3, Figure 2"}],"minor_comments":[{"comment":"The introduction attributes 'the original DiNAT' to Hua et al. (2019), but that reference describes a dilated fully convolutional network for depth estimation; the relevant DiNAT work is Hassani & Shi (2022), which is cited in Section 2. Please correct the citation and the corresponding sentence.","section":"Section 1 and Section 3.1"},{"comment":"There are several typographical errors: 'DiNA T-IR' in Table 3 has an extra space, 'receptive filed' in Section 3.2 should be 'receptive field', and the Figure 2 caption 'sigmoid summation' should be 'sigmoid'.","section":"Table 3, Section 3.2, Figure 2"},{"comment":"Section 5 states that NIQE is used as a perception metric, but Table 5 reports FID and LPIPS only; please add NIQE to the table or delete the mention.","section":"Section 5"},{"comment":"The GoPro paragraph states that 'DiNAT-IR may not fully converge to an optimal solution'; please state explicitly how the final 600K+200K training schedule was chosen and whether continuing training changed the reported 33.80 dB result, or position the reported number as a lower bound.","section":"Section 4, GoPro implementation details"},{"comment":"On HIDE, DiNAT-IR ties MaIR's PSNR but has slightly lower SSIM (0.945 vs 0.946), and on GoPro MaIR has higher SSIM (0.969 vs 0.967); the wording 'matching or surpassing all compared methods' should be made explicitly about PSNR to avoid overstating the result on both metrics.","section":"Abstract and Section 4.1"},{"comment":"The limitation section acknowledges that the ablation was performed only on GoPro; the final version should discuss whether the 1.74 dB improvement over the 16-channel Restormer baseline in Table 5 transfers to the full-resolution models and to deraining and denoising tasks.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the evaluation protocol, not the architectural idea. If the authors can provide a validation-based selection rule and error bars, the paper is likely acceptable. The 'soon' code release should be replaced by an actual link or a reproducibility appendix. The citation error in the introduction should be corrected; I do not see evidence of deliberate misattribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the core result on GoPro motion deblurring (33.80 dB, matching or slightly beating MaIR and NAFNet) is credible, and the ablation story is internally consistent. But the abstract says \"competitive across multiple benchmarks,\" which overreaches: on deraining and denoising DiNAT-IR is consistently below Restormer, sometimes by close to a decibel.\n\nWhat's actually new: applying DiNA to low-level restoration is not novel by itself, but the specific empirical finding is. The paper shows that the hybrid NA-DiNA design from high-level vision hurts deblurring compared to DiNA-only, and that a cheap ECA-style channel-aware module closes the gap. Table 5 supports this narrative cleanly. Credit also for honesty: they state the GoPro model may not have converged, and the deraining table caption openly says \"noticeably lower PSNR.\"\n\nSoft spots:\n- Evaluation protocol. They pick a specific checkpoint per task (290K DPDD-D, 140K DPDD-S, 40K fine-tune for Rain13K, 220K SIDD) without mentioning a held-out validation split. On DPDD the combined PSNR edge over GRL-BD is 0.01 dB. If those checkpoints were chosen by watching test curves, that margin is meaningless. The GoPro/HIDE numbers are safe because GoPro is the last iteration and HIDE is untuned, but the multi-benchmark claim rests on shaky ground. This is the paper's main weakness.\n- Text vs. table mismatch on deraining. Section 4.3 calls the gaps \"negligible,\" but on Test1200 DiNAT-IR gets 32.31 dB vs. Restormer's 33.19 dB, a 0.88 dB drop. That is not negligible, and the framing needs correcting.\n- Citation error: \"The original DiNAT (Hua et al., 2019)\" cites the wrong paper; DiNAT is Hassani & Shi 2022.\n- No error bars, no code yet, and the best-configuration choice in the ablation leans on visual quality claims that are hard to verify.\n\nThe stress-test note is on target: checkpoint selection is the load-bearing concern, but it does not break the core deblurring claim. I would send this to peer review. A good referee should ask for a validation-based checkpoint protocol, error bars or repeated runs, a corrected deraining discussion, and a toned-down abstract. The paper deserves a hearing.","headline":"Solid architecture paper with a credible GoPro deblurring result, but the multi-benchmark claim is softer than the abstract says.","tokens_in":14230,"tokens_out":2878,"would_cite":false,"duration_ms":28472,"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 DiNAT-IR, which alternates local and dilated neighborhood attention with a channel-aware module, reaches 33.80 dB PSNR on GoPro and 31.57 dB on HIDE, matching or surpassing all compared methods.","keywords":["image restoration","motion deblurring","dilated neighborhood attention","channel-aware module","transformer","defocus deblurring","image deraining","image denoising"],"falsifier":"Retrain DiNAT-IR on GoPro with a held-out validation split, choose the checkpoint by validation PSNR, and then compare test PSNR with NAFNet and MaIR under identical training budgets; if the 0.11 dB margin over NAFNet shrinks or reverses, the paper's central improvement claim fails. The same validation-based selection should be applied to the DPDD and SIDD checkpoints, where the stated iterations (290K, 140K, 40K fine-tune, 220K) were selected without a reported validation split.","tokens_in":13202,"feed_emoji":"📷","tokens_out":8686,"duration_ms":80466,"temperature":0.7,"pith_summary":"DiNAT-IR is an attempt to transplant Dilated Neighborhood Attention, a sparse sliding-window attention from high-level vision, into low-level image restoration, and the paper's central claim is that it works if the local-attention branch is paired with a channel-aware module that supplies the global context it lacks. On the GoPro motion-deblurring benchmark the architecture reaches 33.80 dB PSNR, ahead of Restormer, NAFNet, and MaIR; on HIDE it ties MaIR at 31.57 dB, with 25.90 M parameters. A sympathetic reader would care because the result offers a concrete alternative to channel-wise self-attention for high-resolution restoration, and because the paper documents a failure mode of the original NA-DiNA hybrid that would otherwise be invisible. The authors are candid that deraining PSNR falls slightly below Restormer, so the claimed advantage is strongest where global context and local detail both matter.","feed_headline":"Deblurring with dilated attention: 33.80 dB on GoPro","feed_subtitle":"Mixing local and dilated neighborhood attention with a channel-aware module matches or beats larger restoration transformers.","key_machinery":"The central object is the Dual Transformer Block with channel-aware self-attention. Each block combines a spatial-attention branch that alternates between local Neighborhood Attention ($\\delta=1$) and Dilated Neighborhood Attention with stage-specific larger dilations ($\\delta\\in\\{36,18,9,4\\}$ paired with 1), and a channel-aware branch that takes the layer-normalized input, applies global average pooling, a 1-D convolution of kernel size 3 along the channel dimension, and a sigmoid; the two branches are merged by element-wise multiplication. This mechanism carries the paper's argument because it isolates the failure mode of the vanilla NA-DiNA hybrid, limited global context from the local NA branch, and provides the compensating global signal.","core_discovery":"The paper's central claim is that a U-Net restoration model can beat channel-wise self-attention baselines by alternating Neighborhood Attention ($\\delta=1$) with Dilated Neighborhood Attention (stage dilation pairs $\\{1,36\\},\\{1,18\\},\\{1,9\\},\\{1,4\\}$), provided each block is gated by a lightweight channel-aware module that recovers global context. On motion deblurring this architecture reports 33.80 dB PSNR on GoPro and 31.57 dB on HIDE, matching or surpassing Restormer, NAFNet, MaIR, and the other compared methods, and it reports the highest combined PSNR (27.05 dB) for dual-pixel defocus deblurring on DPDD. The authors also state a negative result that motivates the design: the vanilla NA-DiNA hybrid, as used for high-level vision, degrades deblurring performance compared with a DiNA-only design, which they attribute to the local NA branch restricting global context; the channel-aware module is introduced to fix exactly that. The paper is explicit that the advantage is not universal: on deraining benchmarks DiNAT-IR trails Restormer on PSNR while nearly matching SSIM.","pith_inferences":["Editorial inference: the reported margin over NAFNet on GoPro is 0.11 dB and over Restormer is 0.88 dB, but the latter comparison uses very different training schedules and patch sizes; a reader should treat cross-model gaps as protocol-dependent until a common training recipe is used.","Editorial inference: the channel-aware module is a near-free global context signal, so the same GAP-Conv1d-sigmoid gate could be tested as a drop-in addition to other windowed or neighborhood attention models, including super-resolution and video restoration.","Editorial inference: because the paper selects checkpoints at fixed iterations without a validation split, the published numbers should be read as upper-bound estimates of what the architecture can do; independent code release would allow re-evaluation with validation-based early stopping.","Editorial inference: the qualitative argument that NA improves text sharpness despite near-identical PSNR and LPIPS suggests perception metrics may not capture the visual claim; a user study on restored text could test whether the hybrid branch is worth its cost."],"forward_implications":["If the reported GoPro and HIDE numbers hold, Dilated Neighborhood Attention becomes a viable spatial-attention backbone for restoration, not just for detection and segmentation.","The ablation result, with the channel-aware module adding 0.19 dB to the full NA-DiNA configuration on a 16-channel GoPro baseline, implies that other local-attention transformers for low-level vision could gain from the same cheap global-context fix.","The dual-pixel DPDD result (27.05 dB combined PSNR, best among the compared methods) suggests the alternation of local and dilated attention is especially beneficial for defocus patterns, which have structured spatial extent.","Because the model was trained only on GoPro and then evaluated on HIDE, the reported HIDE performance supports the claim that the architecture generalizes to new blur distributions without task-specific retraining.","At 25.90 M parameters, performance comparable to MaIR is reached with fewer parameters, so the design is a parameter-efficient alternative for deployment."],"supporting_citations":[{"why":"Supplies the Restormer U-Net backbone that DiNAT-IR builds on and the channel-attention baseline it is compared against.","marker":"Zamir et al., 2022"},{"why":"Introduces Dilated Neighborhood Attention and the NA-DiNA hybrid that the paper adapts and then critiques for limited global context.","marker":"Hassani & Shi, 2022"},{"why":"Defines Neighborhood Attention, the dilation-value-1 local branch used in the alternating scheme.","marker":"Hassani et al., 2023"},{"why":"Provides the GoPro motion-deblurring dataset used for training, main evaluation, and ablations.","marker":"Nah et al., 2017"},{"why":"Provides the HIDE dataset used to test cross-dataset generalization of the GoPro-trained model.","marker":"Shen et al., 2019"},{"why":"NAFNet, a strong convolutional baseline whose 33.69 dB GoPro PSNR the paper reports surpassing by 0.11 dB.","marker":"Chen et al., 2022"},{"why":"MaIR, the recent restoration model whose GoPro and HIDE PSNR DiNAT-IR claims to match or surpass.","marker":"Li et al., 2025"},{"why":"ECA-Net, the source of the channel-attention design used in the channel-aware module.","marker":"Wang et al., 2020"}],"fun_headline_variants":["Dilated neighborhood attention boosts deblurring to 33.80 dB","Channel-aware dilated attention matches Restormer at 33.80 dB","DiNAT-IR: mixing dilations for sharper deblurring","Neighborhood attention with channel-aware global context hits 33.80 dB","Outperforms channel-only attention via dilated neighbors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported test numbers come from an honest evaluation protocol, meaning the model checkpoints were picked at the stated training iterations without looking at test-set metrics; if that premise is false, the small reported margins, like 0.01 dB on the combined DPDD PSNR, could disappear under a correct protocol.","fun_headline_variants_meta":{"raw":{"variants":["Dilated neighborhood attention boosts deblurring to 33.80 dB","Channel-aware dilated attention matches Restormer at 33.80 dB","DiNAT-IR: mixing dilations for sharper deblurring","Neighborhood attention with channel-aware global context hits 33.80 dB","Outperforms channel-only attention via dilated neighbors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000644,"raw_usage":{"total_tokens":3004,"prompt_tokens":1031,"completion_tokens":1973,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":1895}},"tokens_in":647,"tokens_out":1973,"duration_ms":14177,"temperature":1.0,"reasoning_tokens":1895,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:17:09.101670+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain DiNAT-IR on GoPro with a held-out validation split, choose the checkpoint by validation PSNR, and then compare test PSNR with NAFNet and MaIR under identical training budgets; if the 0.11 dB margin over NAFNet shrinks or reverses, the paper's central improvement claim fails. The same validation-based selection should be applied to the DPDD and SIDD checkpoints, where the stated iterations (290K, 140K, 40K fine-tune, 220K) were selected without a reported validation split.","supporting_citations":[{"cited_title":"Restormer: Efficient transformer for high-resolution image restoration","cited_arxiv_id":null,"evidence_quote":"Supplies the Restormer U-Net backbone that DiNAT-IR builds on and the channel-attention baseline it is compared against."},{"cited_title":"Neighborhood attention transformer","cited_arxiv_id":null,"evidence_quote":"Defines Neighborhood Attention, the dilation-value-1 local branch used in the alternating scheme."}],"review_version":1}