{"id":"8b42fde5-6973-4df9-8ca8-395a49816e57","arxiv_id":"2509.05374","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper derives a composite atmospheric scattering model for non-ideal clean data and uses a four-loss committee to train a dehazing network, claiming state-of-the-art results on real-world haze benchmarks.","lead":"An image dehazing method that treats the 'clean' photos used in training as still containing some haze, and uses a set of losses to unify synthetic and real domains. It reports modest but consistent quality improvements on real-world haze benchmarks, but with caveats in comparisons and tuning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unverified premise: RESIDE 'clean' images are assumed to obey Eq. (2), but no measurement supports this; if false, L_cc enforces a mis-specified relation and the reported gains are not evidence for the proposed mechanism.","rationale":"The paper's central claim is explanatory: the synthetic-to-real gap stems from non-ideal clean data, and the proposed loss committee unifies domains based on Eq. (4). The load-bearing assumption is the validity of Eq. (2) for the training set. The authors provide no quantitative evidence for it. Since L_cc alone does not help and the other losses are not specific to the ASM, the current experiments cannot distinguish the physical-mechanism hypothesis from generic regularization. This concern is more central than the comparison-protocol issues or the Eq. (7) typo, because it attacks the theoretical foundation rather than the empirical protocol. The proposed test isolates the contribution of the non-ideal terms and would settle the matter.","tokens_in":9157,"tokens_out":5482,"duration_ms":54137,"concrete_test":"Retrain the full model (m4) on ITS+OTS but replace L_cc (Eq. 8) with the standard ASM form that assumes Ic clean: L_cc' = L1(e^{-βh z} J + (1 - e^{-βh z}) A_h, I_h), keeping all other losses, weights, and the training schedule identical. Evaluate on I-Haze and O-Haze. If PSNR remains within ~0.1 dB of 17.971/18.237, the extra terms involving Ac and βc in Eq. (4) are not responsible for the gains, undercutting the core mechanism. If PSNR drops substantially, the non-ideal modeling is empirically relevant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The derivation of Eq. (4) rests entirely on the assumption that every training 'clean' image Ic is a hazy observation of an ideal image J following Eq. (2) with estimable Ac, βc (Section III-A). This is never tested. The only support is Figure 2, which shows some depth-dependent haziness in selected images; it does not establish that the ASM holds globally per image. If Ic contains sensor noise, color shifts, or scattering not representable by a single Ac and βc, then Eq. (4) is not the true relation between Ih and J, and L_cc (Eq. 8) encourages physically wrong reconstructions. The ablation (Table III) compounds this: adding L_cc alone worsens PSNR (m2: 15.736 vs m1: 16.293), and the gains only appear after adding L_dc and L_is, which are generic depth and feature-distribution regularizers. Thus the observed improvement may stem from these regularizers rather than from the domain-unification model. The claim that non-ideal clean data cause the synthetic-to-real gap and that the loss committee addresses it is therefore not established by the experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims that the synthetic-to-real domain gap in image dehazing is caused by non-ideal collection of 'clean' training images, which themselves retain residual haze. It derives a composed atmospheric scattering model (Eq. 4) and proposes a loss committee comprising synthesis-domain consistency (Lsc), cross-domain consistency (Lcc), depth consistency (Ldc), and implicit supervision (Lis, a KL divergence between the estimated ideal image J and the observed clean image Ic). The model is trained on RESIDE (ITS/OTS) and evaluated on real-world I-Haze, O-Haze, and RTTS. The reported PSNR/SSIM values on I-Haze and O-Haze are the best among the compared methods (e.g., PSNR 17.971 on I-Haze vs. 17.346 for FCDM), but NIQE is consistently worse than FCDM on all three datasets.","tokens_in":9531,"tokens_out":6860,"duration_ms":64067,"significance":"If validated, the paper offers a useful and principled perspective: that apparent 'clean' training images are themselves hazy observations of an ideal image, and that dehazing should account for this composition. Eq. (4) is a correct algebraic consequence of the assumed model, and the idea of treating the real clean image as an implicit variable is interesting. The authors also evaluate generalization across several backbones and include downstream segmentation/detection examples. However, the current evidence does not support the causal claim: the key premise is never measured, the ablation attributes most of the performance gain to a generic depth-distillation loss, and the comparison protocol has potential confounds. With additional validation and controlled experiments, the contribution could be solid.","major_comments":[{"comment":"The entire method rests on the assumption that every collected clean image Ic is itself an ASM observation of an ideal image J: Ic = e^{-βc z}J + (1 - e^{-βc z})Ac. This premise is asserted but never tested on the RESIDE training data. Figure 2 shows a few qualitative examples with depth-dependent haziness; it does not establish that a single (Ac, βc) per image fits the residuals, nor that those residuals follow the scattering model as opposed to sensor noise, color shifts, or other degradation. If the premise fails, Eq. (4) and Lcc (Eq. 8) are mis-specified, and the reported gains cannot be attributed to the proposed mechanism. Please provide a quantitative validation, e.g., estimate Ac and βc per RESIDE clean image and measure the residual against Eq. (2), or compare training with an alternative 'clean' prior.","section":"Section III-A, Eq. (2) and Eq. (4)"},{"comment":"The ablation does not support the claim that the cross-domain loss is the source of improvement. Adding Lcc alone (m2) decreases PSNR from 16.293 to 15.736 and SSIM from 0.828 to 0.817, and the result is marked 'unstable'. Adding Lis (m3) raises PSNR to 16.814, but the large jump to 17.971 occurs only after Ldc is added in m4. Ldc is a generic depth-consistency regularizer using DepthAnything and is not specific to domain unification. Thus the experiments do not isolate the effect of Lcc. Please report multiple-seed results with variance, and provide a controlled ablation, e.g., m1+Ldc+Lis versus m1+Lcc+Ldc+Lis, to show the marginal contribution of Lcc.","section":"Section IV-C, Table III"},{"comment":"The loss weights αsc, αcc, αdc, and αis are obtained via AutoML (Bayesian optimization), but no validation set is described. If the AutoML procedure used I-Haze or O-Haze test metrics, the comparison with fixed-weight baselines is not fair, and the reported gains of about 0.6–0.8 dB over FCDM could be partly due to selection bias. Please specify the validation split used for weight tuning and state whether the same tuning procedure was applied to the m1 baseline and to the retrained competitors. If the weights were tuned on the test datasets, the headline results are not blind.","section":"Section IV-A, Implementation details"},{"comment":"The comparison protocol is inconsistent across methods. The paper retrains FFANet, GCANet, DEA-Net, and PSD variants on the same training dataset, but uses provided pre-trained models for MSBDN, DehazeFormer, and FCDM. If those pre-trained models were trained on different data or at different resolutions, the comparisons are confounded. This matters especially because FCDM outperforms the proposed method on NIQE on all three datasets. Please retrain all competitors on the same ITS/OTS split (or document the exact pre-training data and resolution), and report statistical significance or error bars for the PSNR/SSIM differences.","section":"Section IV-A, Table I and Table II"},{"comment":"The definition of z in Lcc is ambiguous. The text says 'we treat both zc and zh as z', but Eq. (7) uses zc and Eq. (9) uses zh and zc separately. It is not clear whether z in Eq. (8) is zh, zc, or some combination, and whether gradients flow through both Dehaze calls. Since the two-stage architecture (Eqs. (5)–(6)) feeds the output Ic of the first Dehaze call into a second Dehaze call, the use of stop-gradient is an important implementation detail. Please clarify with exact equations or pseudocode.","section":"Section III-B, Eq. (8)"}],"minor_comments":[{"comment":"The SSIM values (17.261, 17.294, 16.806) are impossible because SSIM is bounded above by 1 and they are identical to the PSNR values in the row above. This appears to be a copy-paste error; please correct the table.","section":"Table IV"},{"comment":"The implicit supervision loss is defined as Ldiv in Eq. (10) but appears as Lis in Eq. (11). Please align the notation.","section":"Section III-B, Eqs. (10)–(11)"},{"comment":"The claim of 'significantly outperforming' is too strong given that NIQE is consistently worse than FCDM (I-Haze: 3.857 vs. 3.817; O-Haze: 2.473 vs. 2.268; RTTS: 3.117 vs. 2.968). Please qualify the claim and discuss this trade-off explicitly.","section":"Abstract and Section IV-B"},{"comment":"Reference [19] (Tang et al., 'Investigating the optimal size of anticancer nanomedicine') appears to be a citation error; it is not related to dehazing. Please verify the citation.","section":"References"},{"comment":"The downstream-task evaluation is only qualitative. Adding quantitative segmentation/detection metrics (e.g., mIoU or mAP) would strengthen the claim.","section":"Section IV-B, Figures 4–5"}],"recommendation":"major_revision","confidential_remarks":"The paper has a reasonable core idea but the evidence is currently insufficient to establish the proposed mechanism. The main concerns are (1) the untested premise of Eq. (2), (2) the ablation attributing most gains to Ldc rather than Lcc, and (3) the unspecified AutoML validation set. These are addressable with additional experiments and clarifications, so I recommend major revision rather than rejection. Please monitor that the authors provide a controlled ablation and clarify the comparison protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: there is a real idea here — the composite ASM relation in Eq. (4) is new as far as I can tell, and the training recipe gives small but consistent PSNR/SSIM gains on I-Haze and O-Haze. But the paper’s central explanation for those gains is not backed by the experiments. The proposed cross-domain loss alone makes things worse (m2: 15.7 vs m1: 16.3 PSNR), and the full improvement comes only after adding depth (DepthAnything) and KL-based implicit supervision, which are generic regularizers. No measurement shows that the RESIDE 'clean' images actually obey Eq. (2) with a single A_c and β_c; the depth-dependent haze in Figure 2 is not enough. If the premise is wrong, Eq. (4) and L_cc are mis-specified, and the gains do not validate the stated mechanism.\n\nWhat is genuinely useful: the paper formalizes the intuition that synthetic clean images are not ideal observations, proposes a two-stage inference where the network also dehazes the clean image, and evaluates on external real-world benchmarks with several competing methods. The ablation is at least reported, including the negative result of m2. I also credit the authors for noting in Section V that training is prone to collapse.\n\nSoft spots beyond the premise: the loss weights are tuned with AutoML but no validation split is described; Table IV lists SSIM values like 17.26, which is impossible and suggests a copy-paste error; Eq. (7) appears to use the wrong subscripts (β_c and A_c where β_h/A_h seem intended); and comparison protocols are not uniform (pre-trained vs retrained baselines). None of these are fatal by themselves, but together with the ablation they mean the 'domain unification' mechanism is currently a hypothesis, not an established explanation.\n\nWho is this for? Someone working on synthetic-to-real dehazing or using dehazing as a downstream pre-processing step. The exact method may not be the lasting part; the observation and the training recipe might be.\n\nMy verdict: send to peer review. The novel formalization and the honest ablations are worth refereeing, but the paper should be pushed to either test the residual-haze assumption directly (e.g., fit ASM to RESIDE clean images and report residuals) or reframe the contribution as an empirical loss recipe without the causal story. As it stands, I would not cite the causal claim, but I might cite the method with caveats.","headline":"A plausible but unproven causal story bolted onto a modestly effective training recipe; worth a serious referee, but the mechanism needs testing.","tokens_in":9948,"tokens_out":4012,"would_cite":false,"duration_ms":39931,"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 that residual haze in supposedly clean training images breaks the standard atmospheric scattering model, and that a loss committee built on the derived composite model closes the synthetic-to-real gap in single-image dehazi","keywords":["image dehazing","synthetic-to-real domain gap","atmospheric scattering model","non-ideal clean data","domain unification","loss committee","depth consistency","real-world dehazing"],"falsifier":"Estimate residual haze parameters (atmospheric light and transmission) directly on RESIDE clean images with a standard dehazing prior, then re-haze them under Equation (4) and measure reconstruction error against the actual synthetic hazy images; if the error is large, the two-layer model is mis-specified. A cleaner controlled test: retrain on clean images that are genuinely haze-free (digitally cleared or rendered with no atmosphere); under the paper's explanation, the real-world gain over plain training should substantially disappear.","tokens_in":9098,"feed_emoji":"🌫️","tokens_out":6170,"duration_ms":58351,"temperature":0.7,"pith_summary":"The paper tries to establish that the synthetic-to-real performance drop in dehazing is caused mainly by non-ideal clean data: the 'clean' images used to synthesize training pairs still contain depth-dependent residual haze. Because of that, the standard atmospheric scattering model does not connect synthetic hazy images to the ideal haze-free image, so the authors derive a two-layer composite model and train a network with a four-part loss committee to implicitly recover the ideal image. The method reports the best PSNR and SSIM among compared methods on I-Haze and O-Haze, and better MUSIQ scores on RTTS, using only synthetic RESIDE training data. If correct, this means real-world dehazing can be improved by fixing the clean side of synthetic training data rather than by collecting real paired hazy/haze-free images. The authors note in Section V that reconstructing the ideal clean image is non-trivial and the model is prone to training collapse without extra feature- or pixel-level constraints.","feed_headline":"Residual haze in 'clean' images explains dehazing domain gap","feed_subtitle":"A loss committee on the composite scattering model lifts real-world PSNR to 17.97 on I-Haze and 18.24 on O-Haze.","key_machinery":"The key object is the composite atmospheric scattering relation of Equation (4): Ih = e^{-(β_h+β_c)z}J + e^{-β_h z}(1-e^{-β_c z})Ac + (1-e^{-β_h z})Ah, obtained by treating the clean image as a hazy observation of an ideal image. This relation reframes synthetic-to-real dehazing as a two-layer scattering problem and is enforced through a loss committee: Lsc constrains the synthetic pair via the standard model on Ic, Lcc constrains Ih through the composite model, Ldc distills depth from a pretrained estimator to tie zc and zh together, and Lis aligns the implicit J with the visible clean-image distribution by KL divergence.","core_discovery":"The central claim is that every collected 'clean' image Ic is itself a hazy observation of an ideal image J, so the training synthesis Ih = e^{-β_h z}Ic + (1-e^{-β_h z})Ah should be composed with Ic = e^{-β_c z}J + (1-e^{-β_c z})Ac. Substituting gives Equation (4), in which Ih and J do not satisfy the simple atmospheric scattering model. The paper builds a dehazing method on this composite relation: it runs the same dehazing network twice, once on Ih and once on Ic, and applies a loss committee—synthesis-domain consistency Lsc, cross-domain consistency Lcc, depth consistency Ldc using a pretrained depth estimator, and implicit KL supervision Lis—to constrain the latent ideal image J. At infe","pith_inferences":["A direct mechanism test: retrain the same loss committee on truly haze-free clean images (digitally cleared or rendered without atmosphere). If Equation (4) is the real cause of the gap, the gain over plain training should shrink toward zero.","The KL divergence between predicted J and Ic aligns distributions but not physical plausibility; a stronger generator prior, such as an adversarial or diffusion model, could both reduce the reported training collapse and improve NIQE, where FCDM remains ahead.","Because the derivation assumes the same scene depth z for Ic and Ih, the method inherits a hidden dependence on the pretrained depth estimator's accuracy; depth errors would propagate directly into Lcc and Ldc.","The same two-layer scattering argument could extend to other degradations with imperfectly collected training data, such as rain, fog with non-uniform scattering, or low-light enhancement."],"forward_implications":["Real-world dehazing can be improved without real paired hazy/haze-free data, by correcting the clean side of synthetic training pairs.","The same loss-committee training recipe improves several regression-based backbones beyond DEA-Net, including FFANet, MSBDN, and DehazeFormer, so the correction is architecture-agnostic.","Depth becomes an explicit training signal: adding depth distillation stabilizes inferred transmission and improves dehazing in deep scene regions.","Downstream segmentation and detection benefit from the dehazed output, supporting the practical value of the method as a preprocessing step."],"supporting_citations":[{"why":"Supplies the atmospheric scattering model used throughout and the starting point for the derived composite relation.","marker":"[7]"},{"why":"Provides the RESIDE synthetic training data (ITS/OTS) and real-world test data (RTTS) used in the experiments.","marker":"[29]"},{"why":"DEA-Net is the backbone network on which the loss committee is added and the main direct comparison.","marker":"[34]"},{"why":"FCDM is the strongest synthetic-to-real baseline, setting the training-data precedent and the performance level the method must beat.","marker":"[13]"},{"why":"PSD is a prior synthetic-to-real dehazing method guided by physical priors, used as a comparison and context for the new approach.","marker":"[9]"},{"why":"DepthAnything provides the pretrained depth predictions used in the depth consistency loss Ldc.","marker":"[32]"},{"why":"FFANet serves as a comparison method and as one of the alternative backbones tested with the proposed training recipe.","marker":"[11]"},{"why":"MSBDN serves as a comparison method and as another backbone validated in the generalization experiments.","marker":"[10]"},{"why":"DehazeFormer represents the ViT-based dehazing baseline and is also tested with the proposed training method.","marker":"[35]"}],"fun_headline_variants":["Clean images are hazy: unified model dehazes real scenes","Residual haze in 'clean' data: the real culprit behind dehazing gap","Treat clean images as hazy to fix synthetic-to-real dehazing","Composite scattering model unifies domains, lifts real PSNR","Clean images carry residual haze: key to real-world dehazing"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes every 'clean' training image is itself a hazy observation of an ideal image, with the residual haze following the same atmospheric scattering model; if that residual comes from sensor noise, color shifts, or non-scattering degradation, the derived composite relation and its losses are mis-specified.","fun_headline_variants_meta":{"raw":{"variants":["Clean images are hazy: unified model dehazes real scenes","Residual haze in 'clean' data: the real culprit behind dehazing gap","Treat clean images as hazy to fix synthetic-to-real dehazing","Composite scattering model unifies domains, lifts real PSNR","Clean images carry residual haze: key to real-world dehazing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000571,"raw_usage":{"total_tokens":2520,"prompt_tokens":711,"completion_tokens":1809,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":1718}},"tokens_in":455,"tokens_out":1809,"duration_ms":17236,"temperature":1.0,"reasoning_tokens":1718,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:21:54.174302+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Estimate residual haze parameters (atmospheric light and transmission) directly on RESIDE clean images with a standard dehazing prior, then re-haze them under Equation (4) and measure reconstruction error against the actual synthetic hazy images; if the error is large, the two-layer model is mis-specified. A cleaner controlled test: retrain on clean images that are genuinely haze-free (digitally cleared or rendered with no atmosphere); under the paper's explanation, the real-world gain over plain training should substantially disappear.","supporting_citations":[{"cited_title":"K., & Narasimhan, S","cited_arxiv_id":null,"evidence_quote":"Supplies the atmospheric scattering model used throughout and the starting point for the derived composite relation."},{"cited_title":"Benchmarking single image dehazing and beyond","cited_arxiv_id":null,"evidence_quote":"Provides the RESIDE synthetic training data (ITS/OTS) and real-world test data (RTTS) used in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DEA-Net is the backbone network on which the loss committee is added and the main direct comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FCDM is the strongest synthetic-to-real baseline, setting the training-data precedent and the performance level the method must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DepthAnything provides the pretrained depth predictions used in the depth consistency loss Ldc."},{"cited_title":"(2020, April)","cited_arxiv_id":null,"evidence_quote":"FFANet serves as a comparison method and as one of the alternative backbones tested with the proposed training recipe."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MSBDN serves as a comparison method and as another backbone validated in the generalization experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DehazeFormer represents the ViT-based dehazing baseline and is also tested with the proposed training method."}],"review_version":1}