{"id":"8692e349-47f9-42ed-8673-cd0de50c148f","arxiv_id":"2501.06818","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An unpaired dehazing network combining shared Retinex priors, wavelet separable convolutions, and adaptive color correction reports modest benchmark gains, but its state-of-the-art claim is contradicted by its own tables on HSTS and I-HAZE.","lead":"UR2P-Dehaze is a network for removing haze from single images that blends Retinex-style illumination and reflectance priors, wavelet-domain convolutions, and an adaptive color corrector. It reports gains of roughly 1.4 to 1.7 dB PSNR over earlier unpaired dehazing methods, but its state-of-the-art claim is undercut by parts of its own comparison tables.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The shared-R Retinex prior in Eq. (3) is inconsistent with the paper's ASM Eq. (1); the physical motivation for the dehazing gains collapses.","rationale":"I concur with the reader's REJECT, but I would put the strongest weight on the internal inconsistency in Eqs. (1) and (3) rather than only on the overbroad SOTA wording. The abstract's unqualified 'state-of-the-art' claim is also false by the paper's own tables: UME-Net reports higher HSTS PSNR (27.35 vs 26.71), and on I-HAZE the method has the worst FID (161.615 vs 128.765 for D4+) with the arrow direction mislabeled (FID↑ should be FID↓). Nevertheless, the more load-bearing problem is that the central physical prior is not derived from the paper's own imaging model; if the shared-R assumption fails, the method's claimed contribution 'learning rich physical prior from hazy images' loses its justification, even if the numeric tables happen to be reproducible. The proposed test directly checks Eq. (3) against Eq. (1). This does not change the reader's verdict; it reinforces it with a more fundamental objection.","tokens_in":16980,"tokens_out":7761,"duration_ms":78745,"concrete_test":"Analytically check the consistency condition: substitute J = Lclear∘R into Eq. (1) and set the result equal to Lhazy∘R; the resulting identity is [Lhazy − tLclear]∘R = A(1−t). Pick any SOTS-outdoor pair with spatially varying R, compute the left and right sides numerically (R can be estimated by the authors' ΦReflect; t and A can be estimated by standard ASM solvers), and report the relative residual. If the residual is substantially above zero, Eq. (3) is refuted and the reflectance-consistency and retinex losses (Eqs. 6–7) enforce a prior that does not describe hazy image formation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The key load-bearing premise is Eq. (3), which asserts that hazy and clear images of the same scene share one reflectance R: Ihazy = Lhazy ∘ R and Iclear = Lclear ∘ R. This is internally inconsistent with the atmospheric scattering model the paper itself states in Eq. (1). If the clear image is J = Lclear ∘ R, then Eq. (1) yields Ihazy = (Lclear ∘ R)t + A(1−t). Equating this with Lhazy ∘ R requires [Lhazy − tLclear]∘R = A(1−t), where the right side is an additive airlight term independent of scene reflectance while the left side is scaled componentwise by R. For non-constant reflectance this equality cannot hold, so the shared-R prior is physically false for hazy images. The downstream priors are consequently unsupported: Lreflect (Eq. 6) compares the predicted R1 to an R2 extracted from a 'rough dehazed image' that is never defined, the 'redundant feature elimination' and 'self-monitoring mechanism' are described only in a figure caption, and Eq. (5)'s inequality is not a derivation of projection loss, because δ is undefined and the minimization is moved inside an upper bound without justification. Thus the reported SOTS gains do not demonstrate that the network learned the claimed physical prior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UR2P-Dehaze, an unpaired image dehazing method that combines a Shared Prior Estimator (SPE) based on Retinex decomposition, a Dynamic Wavelet Separable Convolution (DWSC) module, and an Adaptive Color Corrector (ACC). The authors claim state-of-the-art performance on SOTS-indoor, SOTS-outdoor, I-HAZE, and HSTS across PSNR, SSIM, LPIPS, FID, and CIEDE2000. The method is evaluated with several unpaired and paired comparison methods, and ablations are reported for each module. The central technical premise is that hazy and clear images of the same scene share a single reflectance component, enabling a reflectance-consistency loss and a color-correction loss to guide training.","tokens_in":17366,"tokens_out":4479,"duration_ms":42456,"significance":"If the claims were fully supported, the paper would offer a competitive unpaired dehazing method with explicit color and reflectance modeling, and its evaluations on four datasets with multiple metrics would be a useful empirical contribution. The paper also includes ablations for each proposed module and comparisons with many recent methods. However, the central physical prior is internally inconsistent with the atmospheric scattering model stated by the authors themselves, and the stated state-of-the-art claim is contradicted by the paper's own tables on HSTS and I-HAZE. Because these issues affect the core method and the headline result, the contribution as presented is not established.","major_comments":[{"comment":"The abstract and Section 4.2 claim that UR2P-Dehaze achieves state-of-the-art performance across PSNR, SSIM, LPIPS, FID, and CIEDE2000. This claim is contradicted by the paper's own tables. On HSTS (Table 3), UME-Net achieves a higher PSNR (27.35 vs. 26.71), and on I-HAZE (Table 4), the proposed method's FID of 161.615 is the worst in the table and its SSIM of 0.770 is below RPC-Dehaze's 0.783. In addition, Table 4 labels FID with an up arrow, which is inconsistent with the convention used in Tables 1–3 and with the fact that lower FID is better. The state-of-the-art claim must be substantially revised or removed.","section":"Abstract, Section 4.2, Tables 3 and 4"},{"comment":"The shared-reflectance Retinex prior in Eq. (3) is inconsistent with the atmospheric scattering model stated in Eq. (1). If the clear image J satisfies J = L_clear ⊙ R, then Eq. (1) gives I_hazy = (t ⊙ L_clear) ⊙ R + A ⊙ (1 − t). For this to equal L_hazy ⊙ R, one would need [L_hazy − t ⊙ L_clear] ⊙ R = A ⊙ (1 − t), a nontrivial condition that generally cannot hold for spatially varying scene reflectance and airlight. The physical motivation for the shared prior estimator is therefore not justified by the equations in the manuscript. The authors should either reconcile the two models, state the conditions under which Eq. (3) is a valid approximation, or provide empirical evidence that the approximation holds on the datasets used.","section":"Section 3.1, Eq. (1) and Eq. (3)"},{"comment":"Several key components of the SPE are not formally defined. In Eq. (6), R2 is said to be the reflectance map of the 'rough dehazed image', but this image is never defined. In Eq. (5), δ is introduced as 'the error' but its meaning, range, and estimation are not specified, and the subsequent inequality is not a derivation of the projection loss. The 'self-monitoring mechanism' and the 'redundant feature elimination module' appear only in the Figure 3 caption and are never described in the body of the paper. These undefined entities are load-bearing because the projection loss, reflectance-consistency loss, and the claimed self-monitoring behavior all depend on them.","section":"Section 3.1, Eqs. (5) and (6)"},{"comment":"The ablation study is difficult to interpret because the 'Baseline' is never defined in relation to the components. The Baseline row in Table 5 has numbers very close to those of D4 in Table 4, suggesting that the baseline may be a reproducing of D4, but this is not stated. Furthermore, the rows 'w SPE' and 'w/o SPE' represent different configurations: 'w SPE' adds SPE to the baseline, while 'w/o SPE' removes SPE from the full model. This dual use makes it unclear what each ablation actually measures. The authors should specify the baseline architecture and each ablation configuration explicitly.","section":"Section 4.3, Tables 5 and 6"}],"minor_comments":[{"comment":"The FID arrow in Table 4 is reversed; it should be a down arrow because lower FID is better.","section":"Section 4.2, Table 4"},{"comment":"The method name YOLY is written as 'LOLY' in Section 4.2 and Figure 5; please use a consistent name.","section":"Throughout"},{"comment":"There is a typo: 'The process oof performing' should be 'The process of performing'.","section":"Figure 4 caption"},{"comment":"The text says 'de-ghosting process', which appears to be a typo for 'dehazing process'.","section":"Section 4.2"},{"comment":"The paper is described as using unpaired training, but the model is trained on RESIDE (ITS and OTS), which are synthetic paired datasets. Please clarify what 'unpaired' means in this context, e.g., whether the hazy and clear images are used without explicit pair correspondence.","section":"Section 1 and Section 3"},{"comment":"The wavelet filter type and number of levels are listed as inputs in Algorithm 1 but are not specified in the implementation details; please report the exact values used.","section":"Figure 3 and Algorithm 1"},{"comment":"The color map CF is said to be predicted from the hazy feature image F, but the estimator ΦColor and the features F and F1 are not defined in the text.","section":"Section 3.3"}],"recommendation":"reject","confidential_remarks":"The manuscript is not in a publishable state. The central state-of-the-art claim is directly contradicted by Tables 3 and 4, and the physical prior in Eq. (3) is inconsistent with the paper's own atmospheric scattering model in Eq. (1). These are load-bearing issues that cannot be fixed with minor edits. If the authors substantially revise the claims, define the missing components, and provide a defensible justification for the Retinex prior, a future resubmission might be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an incremental unpaired-dehazing paper that does something real—it beats D4/D4+ on SOTS-indoor/outdoor—but the headline 'state-of-the-art' claim is too broad, and the physical prior that gives the paper its identity doesn't survive contact with the paper's own atmospheric scattering model.\n\nThe genuinely new part is the package: a shared Retinex prior estimator with an iterative reflectance-consistency loss, wavelet-domain separable convolution, and an adaptive color corrector. The ablations cover the three modules on four datasets, and the color-histogram analysis is a nice complement to PSNR/SSIM. If the numbers are reproducible, this is a legitimate data point in the unpaired block: 26.82/0.948 on SOTS-indoor and 27.53/0.967 on SOTS-outdoor, both ahead of D4+ and RPC-Dehaze.\n\nThe soft spots are serious. First, Eq. (3) says hazy and clear images share one reflectance R, with I = L ∘ R. But the paper's own Eq. (1) is the atmospheric scattering model, which has an additive airlight term A(1−t). Substituting J = L_clear ∘ R into Eq. (1) gives I = (L_clear ∘ R)t + A(1−t), which cannot equal L_hazy ∘ R for non-constant reflectance. The 'rich physical prior' is therefore not physically grounded; the loss functions may still guide the network, but the claimed mechanism does not. The projection loss (Eqs. 4–5) compounds this: I_project and δ are never defined, and the inequality in Eq. (5) doesn't justify the minimization. The 'rough dehazed image' that supplies R2 in Eq. (6) is also undefined.\n\nSecond, the abstract's unqualified SOTA claim is contradicted by the paper's own tables. On HSTS, UME-Net has higher PSNR (27.35 vs 26.71); on I-HAZE, the method's FID is the worst in the table and the arrow is mislabeled. The gains over D4+ are real on SOTS, but they are 1.0–1.7 dB, not a broad sweep.\n\nThird, the 'self-monitoring mechanism' promised in the abstract never appears in the methods or experiments. The code is not released, and the wavelet filter type/levels and SPE iteration count are under-specified. The ablation 'Baseline' is numerically identical to D4, which should have been stated.\n\nNet: the empirical contribution is modest but usable, and the modules are plausible. The framing, the physics, and the reproducibility need major work. For peer review, I would not desk-reject—the unpaired block deserves a check on these numbers—but I would send it back with a strong request to fix the physical story, define the undefined quantities, release code, and draw the claims down to what the tables actually support.","headline":"A modest empirical gain on SOTS is real, but the physical prior story doesn't survive contact with the paper's own ASM, and the SOTA claim is overbroad.","tokens_in":17876,"tokens_out":4788,"would_cite":false,"duration_ms":41682,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68U10","68T45"],"pacs":[],"model":"deepseek-v4-flash","headline":"Shared reflectance prior lifts unpaired dehazing to new SOTA scores.","keywords":["unpaired image dehazing","Retinex theory","shared prior estimator","dynamic wavelet separable convolution","adaptive color corrector","atmospheric scattering model","SOTS benchmark","color restoration"],"falsifier":"Render a fixed 3D scene under two different synthetic haze densities (varying $\\beta$ in $t(x) = e^{-\\beta d(x)}$) with known ground-truth reflectance, run the trained shared prior estimator on both hazy images, and check whether the two predicted reflectance maps coincide; if they diverge systematically with $\\beta$, the shared-$R$ assumption underlying $L_{reflect}$ is violated.","tokens_in":16733,"feed_emoji":"🌫️","tokens_out":7290,"duration_ms":59891,"temperature":0.7,"pith_summary":"This paper argues that unpaired image dehazing—where no clean reference is available during training—can be substantially improved by learning several physical priors at once instead of relying on one hand-crafted prior. It proposes UR2P-Dehaze, which uses a shared prior estimator to enforce that hazy and clear versions of the same scene share a single reflectance map under Retinex theory, while a projection step removes 'redundant features.' On the SOTS-outdoor benchmark the reported PSNR is 27.53 dB and SSIM 0.967, leading all unpaired methods and beating the previous best unpaired method by 1.68 dB PSNR and 0.036 SSIM; on SOTS-indoor the reported PSNR is 26.82 dB and SSIM 0.948. If these numbers hold, the method also improves LPIPS, FID, CIEDE2000, and color histogram alignment, which matters because unpaired dehazing is the practical, data-scarce setting.","feed_headline":"Shared reflectance prior lifts unpaired dehazing to a new high","feed_subtitle":"UR2P-Dehaze gains 1.68 dB PSNR on SOTS-outdoor and restores reference colors.","key_machinery":"The load-bearing identity is the Retinex assumption $I = L \\circ R$ (element-wise product), extended to image pairs in Eq. 3: the hazy image and the clear image share one reflectance $R$, differing only in illumination $L$. On this basis the Shared Prior Estimator (SPE) runs three estimators—$\\Phi_{Light}$, $\\Phi_{Reflect}$, $\\Phi_{Color}$—optimized by four losses: a projection loss $L_{project} = \\|I_{hazy} - I_{project}\\|_2^2$ with an asserted bound $\\delta$ in Eq. 5, a reflectance consistency loss $L_{reflect} = \\|R_1 - R_2\\|_2^2$ that ties the predicted reflectance to a 'rough dehazed image,' a Retinex loss $L_{retinex}$ enforcing $R \\circ L \\approx I_{project}$ plus an illumination smoothness term, and a color loss $L_{color} = \\|C_F - \\hat{C}_F\\|_1$. Reconstruction uses Dynamic Wavelet Separable Convolution (DWSC), which applies depthwise separable and pointwise convolutions across wavelet subbands (Eqs. 8–11), and an Adaptive Color Corrector (ACC) to align colors. The total loss is $L_{loss} = 50 L_{project} + 0.1 L_{reflect} + 0.1 L_{retinex} + 1 L_{color}$, and the SPE is iteratively trained so that the estimated priors stabilize.","core_discovery":"The core claim is that a Retinex decomposition $I_{hazy} = L_{hazy} \\circ R$ and $I_{clear} = L_{clear} \\circ R$, with a single shared reflectance $R$ for hazy and clear images of the same scene, gives an unpaired network enough structure to outperform all prior unpaired dehazing methods on SOTS-indoor and SOTS-outdoor. The paper reports that the full model reaches 26.82 dB PSNR / 0.948 SSIM indoors and 27.53 dB / 0.967 outdoors, reduces LPIPS, FID, and CIEDE2000 relative to the best unpaired baselines, and brings the RGB histograms of dehazed images measurably closer to the clear reference. It further reports that dehazing with UR2P-Dehaze increases the number of correctly detected objects in hazy images, showing a benefit for downstream vision.","pith_inferences":["The paper leaves the 'rough dehazed image' that supplies $R_2$ undefined; a concrete next step is to pin down that source, since different choices change the reflectance-consistency loss and could alter the reported gains.","The projection inequality in Eq. 5 hints at a certification path: if the error $\\delta$ can be measured or bounded on a validation set, the method could be extended to output a confidence or failure warning per image.","The wavelet-domain separable convolution is a generic building block; it could be dropped into other restoration tasks (denoising, super-resolution) where a large receptive field is wanted at low parameter cost, though the paper does not test this.","The ACC's close histogram match to the reference suggests the module could be used as an automatic color-grading or white-balancing component even when full dehazing is not desired."],"forward_implications":["UR2P-Dehaze becomes the new leading unpaired dehazing method on SOTS-indoor and SOTS-outdoor if the reported metrics reproduce, with the largest gains on outdoor scenes.","The improved CIEDE2000 and histogram alignment indicate that unpaired dehazing can restore color rather than merely increase contrast, making it more suitable for photography and color-sensitive applications.","Because the method is trained on unpaired RESIDE splits and tested on real I-HAZE and HSTS images, the approach is a practical option when paired clean/hazy data are unavailable.","The reported downstream detection improvements suggest that dehazing with UR2P-Dehaze can serve as a preprocessing step for object detection and recognition in foggy conditions."],"supporting_citations":[{"why":"Supplies the Retinex decomposition $I = L \\circ R$ that the shared prior estimator is built on.","marker":"[35]"},{"why":"Provides the RESIDE training/test datasets (ITS, OTS, SOTS) used for all reported comparisons.","marker":"[37]"},{"why":"The strongest unpaired baseline (D4) that UR2P-Dehaze claims to surpass by 1.4–1.68 dB PSNR.","marker":"[11]"},{"why":"The D4+ follow-up whose published LPIPS, FID, and CIEDE2000 numbers are used as recent unpaired baselines.","marker":"[43]"},{"why":"A strong supervised method (FFANet) that frames the gap between paired and unpaired dehazing performance.","marker":"[47]"}],"fun_headline_variants":["Shared reflectance prior raises unpaired dehazing bar","Unpaired dehazing gains from rich physical prior","Physical prior enhances unpaired dehazing fidelity","Reflectance prior sharpens unpaired dehaze details"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole training procedure assumes that a hazy image and its clear counterpart of the same scene differ only in illumination and therefore share one true reflectance map $R$; if haze changes apparent reflectance beyond a pure lighting scaling, the shared-$R$ prior is wrong and the reported gains would not transfer to new scenes.","fun_headline_variants_meta":{"raw":{"variants":["Shared reflectance prior raises unpaired dehazing bar","Unpaired dehazing gains from rich physical prior","Physical prior enhances unpaired dehazing fidelity","Reflectance prior sharpens unpaired dehaze details"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000302,"raw_usage":{"total_tokens":1772,"prompt_tokens":1012,"completion_tokens":760,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":700}},"tokens_in":628,"tokens_out":760,"duration_ms":7110,"temperature":1.0,"reasoning_tokens":700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:51:48.992755+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a fixed 3D scene under two different synthetic haze densities (varying $\\beta$ in $t(x) = e^{-\\beta d(x)}$) with known ground-truth reflectance, run the trained shared prior estimator on both hazy images, and check whether the two predicted reflectance maps coincide; if they diverge systematically with $\\beta$, the shared-$R$ assumption underlying $L_{reflect}$ is violated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Retinex decomposition $I = L \\circ R$ that the shared prior estimator is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the RESIDE training/test datasets (ITS, OTS, SOTS) used for all reported comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The strongest unpaired baseline (D4) that UR2P-Dehaze claims to surpass by 1.4–1.68 dB PSNR."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The D4+ follow-up whose published LPIPS, FID, and CIEDE2000 numbers are used as recent unpaired baselines."}],"review_version":1}