{"id":"09a40a12-e513-4117-894e-fd3c09994496","arxiv_id":"2509.00108","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SGLC is a two-stage, global-then-local patch-based framework that improves high-resolution image dehazing and, on the authors' 4-image test set, beats downsampled Uformer by over 10 dB PSNR.","lead":"To dehaze very large photos, the authors split the work into two stages: first reconstruct coarse global structure from grid patches, then sharpen local details from overlapping window patches. They report large PSNR gains over downsampling-based baselines on high-resolution images and a top-5 finish in the NTIRE 2023 non-homogeneous dehazing challenge.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical support for the headline PSNR gain is not yet verifiable: the two NTIRE leaderboard tables contradict each other, and the local test uses 4 self-selected images.","rationale":"The reader's verdict is CONDITIONAL, and my read supports that. The strongest claim depends on the numbers in Table 1.1, so the first question is whether those numbers are externally corroborated. The paper cites the NTIRE 2023 challenge for that purpose, but the two tables reporting the challenge results disagree in both the SGLC PSNR value (22.27 vs 22.49) and its PSNR ranks (3rd vs 6th overall; 1st vs 2nd among no-extra-data entries). These are not cosmetic differences: Table 1.3 even shows a no-extra-data baseline ([Mask], 22.90) above SGLC, which directly contradicts the 'best no-extra-data PSNR' claim in Table 1.2. Until this is reconciled, the challenge placement cannot be used as independent support. The local 4-image evaluation is too fragile to carry the claim alone: with four images and no error bars, a favorable split can easily move PSNR by several dB, and the reported margin is about 10 dB over one baseline. I do not allege that the result is fabricated; the challenge entry is real evidence, and the Algorithm 1 indexing issue in §1.3.1.1 looks like a correctable typo rather than the central weakness. The central weakness is that the published quantitative record is internally inconsistent, so an independent reader cannot verify the headline claim from the paper. The reader's CONDITIONAL verdict already reflects the need for correction; I would keep it, with the leaderboard reconciliation as the first required check.","tokens_in":13069,"tokens_out":8012,"duration_ms":74411,"concrete_test":"Obtain the official NTIRE 2023 Non-Homogeneous Dehazing final leaderboard and the submitted SGLC checkpoint or scores from the organizers, then recompute SGLC's overall PSNR rank and its no-extra-data PSNR rank. If the official PSNR is 22.49 dB, Table 1.2's 'tied best no-extra-data PSNR' and 3rd-overall PSNR claims are false; if it is 22.27 dB, Table 1.3's row and sub-ranks are wrong. If the challenge server is still accessible, additionally run the same checkpoint on the official validation set and compare the returned PSNR/SSIM to both tables.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a quantitative one: SGLC markedly improves high-resolution dehazing PSNR/SSIM (§1.4.2). The two evidentiary pillars are the local comparison and the NTIRE leaderboard, and the leaderboard pillar is internally contradictory. Table 1.2 reports SGLC at 22.27 dB, tied for best (1/12) PSNR among no-extra-data solutions and 3/17 overall in PSNR; Table 1.3, presented as the same final leaderboard, lists SGLC at 22.49 dB with a 6th PSNR sub-rank and 5th overall, while [Mask] (no extra data) is shown at 22.90 dB. Both values and ranks cannot be correct, so the paper does not currently establish SGLC's actual challenge standing. The local table does not fill the gap: it is computed on a self-selected 4-image split from the 40 training pairs (§1.4.1.1), with no per-image scores, confidence intervals, or seeds. Under these conditions the reported 10 dB advantage over resized Uformer cannot be distinguished from split-selection effects. The claim may be true, but the quantitative basis as published is unverifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SGLC, a two-stage framework for high-resolution image dehazing. The first stage (GFG) divides the padded input into grid patches, dehazes each patch with a base model (Uformer), and reconstructs a global output; the second stage (LFE) refines this output on window patches and uses a blending step (MOPS) to remove seams. The authors report large PSNR/SSIM gains over downsampled Uformer, DW-GAN, and 4K-Dehazer on a 4-image local subset of HD-NH-HAZE, and claim top no-extra-data ranks in the NTIRE 2023 Non-Homogeneous Dehazing Challenge.","tokens_in":13290,"tokens_out":3572,"duration_ms":31941,"significance":"If the results hold, SGLC offers a simple, model-agnostic way to adapt existing mid-resolution dehazing networks to high-resolution inputs, with an informative ablation showing the benefit of the global-then-local ordering and the computational cost of MOPS. The paper is honest about the large inference-time overhead and does not rely on external training data. However, the central empirical claim is currently not independently verifiable: the grid-patching algorithm is described incorrectly, the two NTIRE leaderboard tables contradict each other, and the local evaluation uses a self-selected 4-image split with no variance or confidence information.","major_comments":[{"comment":"Algorithm 1's patch indexing, P_k[i,j] = I'[i+n_w, j+n_h], does not depend on the patch index k, so every generated patch is identical to the same shifted crop. As written, the grid-patching procedure cannot sample the full image and the central GFG mechanism of global coverage is not reproducible. The intended spatial offsets (e.g., based on k mod n_w and k div n_w) must be specified, or the algorithm description is incorrect.","section":"1.3.1.1, Algorithm 1"},{"comment":"The two leaderboard tables reported as the same final NTIRE 2023 challenge results are mutually inconsistent. Table 1.2 gives SGLC a PSNR of 22.27 dB with rank 1/12 among no-extra-data solutions and 3/17 overall in PSNR, while Table 1.3 lists SGLC at 22.49 dB with overall rank 5 and rank 2 among no-extra-data methods, with [Mask] at 22.90 dB. Both the PSNR values and the ranks cannot be correct simultaneously, so the paper does not currently establish SGLC's actual challenge standing.","section":"1.4.2, Tables 1.2 and 1.3"},{"comment":"The local evaluation that supports the headline gain is computed on a self-selected split of 4 images from the 40 training pairs, with no per-image scores, no standard deviation or confidence intervals, and no disclosure of how the 4 test images were chosen. Under these conditions, the reported 10 dB advantage over resized Uformer cannot be distinguished from split-selection effects. The authors should report full per-image metrics, justify the split selection, and provide seeds or repeated runs for reproducibility.","section":"1.4.1.1 and 1.4.2"},{"comment":"The text states that 'SGLC placed 5th among 13 submitted solutions' while Tables 1.2 and 1.3 refer to 17 solutions and rank 3/17 and 5/17 respectively; the counts and ranks should be reconciled with the official final leaderboard or the source should be cited consistently.","section":"1.4.3"}],"minor_comments":[{"comment":"References [10] and [11] both point to the same NTIRE 2023 challenge report with different formatting; this should be unified.","section":"References"},{"comment":"The implementation details report only ranges for batch size and learning rate (e.g., batch size 2-8, learning rate 1e-4 to 2e-4); exact hyperparameters used for the final SGLC model should be stated for reproducibility.","section":"1.4.1.2"},{"comment":"The initialization line 'zeros(G·n_w, G·n_h, 3)' and the surrounding text use width and height in an ambiguous order; the axis conventions should be clarified to avoid confusion about which dimension is horizontal and which is vertical.","section":"1.3.1.5, Algorithm 2"},{"comment":"The sentence 'The visual comparison in Figure 1.6.' is a fragment and should be completed; additionally, Figure 1.6 is referenced without detailed discussion of what it shows beyond qualitative comparison.","section":"1.4.2"},{"comment":"The method is called 'Streamlined Global and Local Features Combinator' in the abstract and 'Dual-Stage Global and Local Feature Framework' in the title; consistent naming should be used throughout.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a book chapter rather than a self-contained journal paper, and several elements (author biographies, chapter numbering) suggest it is intended for a volume. The editor should verify that the empirical claims are strengthened before considering publication, as the current evidence is internally inconsistent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. This is a plausible engineering idea stuck inside a manuscript with serious writing problems. The proposed two-stage pipeline—global grid patching followed by local window patching with overlap blending—is a reasonable way to extend mid-resolution dehazing models to 4K inputs. The order ablation (SGLC vs Inv-SGLC) is a nice touch and gives some support for the sequential design.\n\nWhat's actually new: I don't know of prior work doing exactly this sequential global-then-local patch composition for high-res dehazing. The method is modular, uses a standard backbone, and the authors limited themselves to official challenge data. That's honest practice.\n\nThe soft spots are not minor. Algorithm 1, the core of the GFG, samples every patch from the same indices—the index k never appears, so as written all patches would be identical. That is a load-bearing inconsistency. The local evaluation uses four self-selected images from the training set, with no per-image scores, no error bars, and no disclosure of how the split was chosen. And the two NTIRE leaderboard tables, presented as the same final ranking, report SGLC at 22.27 dB and 22.49 dB with different overall ranks. Both cannot be right. The external validation, which is the strongest evidence for the method, is internally contradictory.\n\nIf these were minor typos, I'd say fix and move on. But they strike at the verifiability of the main claims. A careful reviewer could still sort it out if the authors supply corrected code, test-split details, and reconciled leaderboard numbers. Given the challenge entry exists, I suspect the method actually works.\n\nWho is this for? Practitioners who want a modular way to scale a transformer dehazer to high resolution, in offline settings. Not a theory paper, and the 553s/image runtime with MOPS limits real-time use. My recommendation: send it to reviewers—the flaws are fixable—but don't cite it until the leaderboard contradiction is explained.","headline":"Plausible two-stage dehazing pipeline, but Algorithm 1 is mis-specified and the leaderboard tables contradict each other, so the claimed gains are not yet verifiable.","tokens_in":13857,"tokens_out":5423,"would_cite":false,"duration_ms":48873,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage global-then-local patch wrapper lifts high-resolution image dehazing from 14.9 dB to 25.4 dB PSNR in the paper's experiments.","keywords":["image dehazing","high-resolution imagery","global context","local detail enhancement","grid patching","window patching","transformer backbone","PSNR and SSIM evaluation"],"falsifier":"Execute Algorithm 1 as printed: because the lookup $P_k[i,j] = I'[i+n_w, j+n_h]$ does not involve the patch index $k$, every generated patch is identical, so the GFG stage cannot literally implement the claimed global coverage; a corrected sampling rule must be supplied and tested before the reported 25.43 dB can be attributed to the described mechanism.","tokens_in":12832,"feed_emoji":"🌫️","tokens_out":11748,"duration_ms":94763,"temperature":0.7,"pith_summary":"This paper argues that high-resolution dehazing fails mainly because standard practice downsizes large images or cuts them into arbitrary patches, losing either global context or local detail. To fix this, it proposes the Streamlined Global and Local Features Combinator (SGLC), a two-stage, model-agnostic wrapper: first a Global Features Generator processes minimally overlapping grid patches to produce a scene-level dehazed image, then a Local Features Enhancer refines that output with window patches. On a local test set of four 4000x6000 images from HD-NH-HAZE, SGLC reaches 25.43 dB PSNR and 0.8524 SSIM, compared with 14.90 dB / 0.6403 for Uformer on resized input and 17.46-17.48 dB for two specialized high-resolution dehazers. The paper also reports that the same framework placed among the top entries in the 2023 non-homogeneous dehazing challenge when no external training data were used.","feed_headline":"Sequential global-then-local patching lifts high-res dehazing by 10 dB","feed_subtitle":"A model-agnostic wrapper lets existing dehazing networks work on 4000x6000 images without downsampling.","key_machinery":"The load-bearing mechanism is the two-stage patch pipeline. The GFG pads the input to make height and width divisible by a patch size $G \\times G$, divides the padded image into $n_h \\times n_w$ grid patches intended to span the whole scene with minimal overlap, runs a dehazing model on each patch, and reconstructs a full-size output $\\hat{I}_F$ by reversing the grid. The LFE then tiles that output into dense window patches, runs a second Uformer-based enhancer, and assembles the result with MOPS, which averages overlapping predictions under a second-order spline window to hide seams. Both models are trained with a customized loss that combines pixel-level fidelity with Laplacian-pyramid high-frequency alignment. The sequential ordering of these two blocks, global before local, is the design choice the ablations single out as decisive.","core_discovery":"The central claim is that the global/local tension in high-resolution dehazing is best resolved sequentially rather than in parallel or by resizing. The GFG stage alone gives 24.49 dB PSNR on the local test set, already far above the 14.90 dB of Uformer applied to a resized image; adding the LFE raises this by about 0.89 dB to 25.43 dB, and the MOPS smoother adds only about 0.05 dB more while removing seams. Reversing the order (LFE before GFG) consistently underperforms, which the authors take as evidence that coarse holistic restoration should precede local refinement. The 2023 non-homogeneous dehazing challenge results are used to show the gain is not an artifact of the four-image local test: among no-extra-data solutions, SGLC tied for the best PSNR and SSIM.","pith_inferences":["If the grid-patching mechanism is corrected to sample distinct patch locations, the same sequential global-then-local wrapper could plausibly transfer to other high-resolution restoration tasks, such as deraining, deblurring, or super-resolution, that face the same global/local trade-off.","The reported 25.43 dB comes from four self-selected training-set images; until the official withheld test split is used, the true margin over the 17.5 dB baselines remains an open question.","A corrected implementation or code release would allow a direct check of whether the GFG stage is best understood as a global-context mechanism or simply as several independent dehazing passes whose outputs are stitched together."],"forward_implications":["Any existing dehazing network can be wrapped in SGLC to gain high-resolution capability, because the framework only adds patch generation, reconstruction, and blending around the model.","The global-first, local-second order is essential: every reversed configuration tested underperforms the standard order.","Dropping MOPS reduces per-image inference time from about 553 seconds to about 86 seconds while sacrificing only about 0.05 dB PSNR, giving an explicit accuracy-versus-speed trade-off for offline and near-real-time use.","Because the top no-extra-data challenge results were achieved without external training data, the structure of SGLC itself, rather than data scale, appears to drive the improvement."],"supporting_citations":[{"why":"The Uformer architecture is the backbone used for both the GFG and LFE stages, and the resized-Uformer numbers are the primary baseline the paper improves on.","marker":"(32)"},{"why":"Supplies the HD-NH-HAZE dataset and the official challenge leaderboard used to corroborate the local results.","marker":"(11)"},{"why":"DW-GAN is one of the high-resolution dehazing methods compared against in the local experiments.","marker":"(24)"},{"why":"4K-Dehazer, the specialized 4K dehazing baseline, is the other main comparison method.","marker":"(42)"},{"why":"The overlapping-patch blending algorithm adapted here becomes MOPS, the smoother that suppresses vignetting in the LFE output.","marker":"(33)"},{"why":"Provides the challenge context and evaluation protocol for one of the baseline comparisons.","marker":"(9)"},{"why":"Supplies the spatial window functions used in the MOPS blending.","marker":"(5)"}],"fun_headline_variants":["Global-first fusion sharpens high-res dehazing without downsampling","Model-agnostic two-stage dehazing gives 10 dB high-res gain","High-res dehazing: global before local beats parallel fusion","Dual-stage global-local framework targets high-res haze removal","No downsampling needed: sequential global-local dehazing wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's gains rest on the assumption that a coarse grid of minimally overlapping patches still supplies the dehazing model with enough global scene context, and that the printed patch-sampling rule actually covers the whole image rather than repeating one location.","fun_headline_variants_meta":{"raw":{"variants":["Global-first fusion sharpens high-res dehazing without downsampling","Model-agnostic two-stage dehazing gives 10 dB high-res gain","High-res dehazing: global before local beats parallel fusion","Dual-stage global-local framework targets high-res haze removal","No downsampling needed: sequential global-local dehazing wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3273,"prompt_tokens":1039,"completion_tokens":2234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":2146}},"tokens_in":655,"tokens_out":2234,"duration_ms":14814,"temperature":1.0,"reasoning_tokens":2146,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:42:40.224516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Execute Algorithm 1 as printed: because the lookup $P_k[i,j] = I'[i+n_w, j+n_h]$ does not involve the patch index $k$, every generated patch is identical, so the GFG stage cannot literally implement the claimed global coverage; a corrected sampling rule must be supplied and tested before the reported 25.43 dB can be attributed to the described mechanism.","supporting_citations":[],"review_version":2}