{"id":"46faf227-eb59-4dcb-8c2a-1962dc60609e","arxiv_id":"2507.19184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A continual learning extension of AGLC-GAN that adds selective kernel fusion, cycle-contrastive loss, and EWC to handle unpaired dehazing, desnowing, and deraining in one model.","lead":"This paper combines three known deep learning components, selective kernel fusion, contrastive loss, and elastic weight consolidation, into a single CycleGAN-style model that performs dehazing, desnowing, and deraining sequentially without paired training data. The result is a modest improvement in PSNR and SSIM over the authors' previous AGLC-GAN on several benchmark datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never reports the sequentially trained 'unified' model's PSNR/SSIM on all three benchmarks; it validates components only as single-task models and reports only forgetting metrics for EWC, so the central continual-learning claim is not directly testable.","rationale":"I read the paper in good faith. The architecture and single-task results are plausible: the model is built on AGLC-GAN with SK fusion, cycle-contrastive loss, and EWC, and Tables I-IV show consistent gains over the self-cited AGLC-GAN baseline. However, the strongest claim in the abstract is about a continual-learning-based unified model, and that claim is not directly evaluated anywhere in the paper. Tables I-IV and VI-VIII are single-task training and testing. Section V-D is the only continual-learning experiment, and it reports only forgetting values and lambda sweeps. Forgetting alone cannot establish that the unified model restores images well; a model can forget little and still be useless if the EWC penalty freezes parameters before the later tasks are learned, or if the final performance on earlier tasks is far below the single-task numbers. The paper never gives the final PSNR/SSIM on all three test sets after sequential training, so the reader cannot verify the central claim. This is a concrete, fixable omission rather than a contradiction, but it is load-bearing: if the unified model performs substantially worse than single-task models, the paper's main contribution collapses. The reader's weakest assumption about test-set tuning of lambda and task order is real, but it is secondary to the absence of the central experiment. I also note the ablation inconsistencies where SK-Fusion alone beats the full model on some metrics; these small discrepancies reinforce the need for the missing final unified-model evaluation. I recommend UNVERDICTED because the current manuscript does not provide the evidence needed to assess the paper's primary claim, though conditional acceptance could be appropriate if the authors supply the missing sequential-training results and code.","tokens_in":15935,"tokens_out":4379,"duration_ms":44811,"concrete_test":"Add a table reporting PSNR/SSIM of the DA-AGLC-GAN trained sequentially with EWC (task order deHaze->deSnow->deRain, lambda=750, lambda1=700, lambda2=800) on the test splits of RESIDE-ITS/OTS, SRRS, and Rain100H, alongside (i) the same architecture trained independently per task, (ii) a no-EWC sequential fine-tuning baseline, and (iii) the best prior unpaired/paired method from Tables I-IV. Run each configuration with at least three seeds and report mean +/- std. If the unified model's scores are within ~0.5 dB PSNR of the single-task models on all three tasks and exceed the no-EWC baseline, the continual-learning claim is supported; otherwise it is not.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The paper's stated contribution is a continual-learning-based unified model, not merely three independently trained restoration networks. Sections V-A to V-C establish that the architecture is a strong single-task dehazer/desnower/derainer, but the sequential-training experiment in Section V-D is the only place the unified model appears, and it reports only forgetting values (Eq. 18, Table V) and lambda sweeps (Figs. 7-8). No table or figure reports the final PSNR/SSIM of the sequentially trained DA-AGLC-GAN on the test splits of RESIDE-ITS/OTS, SRRS, or Rain100H. Forgetting is a relative metric: a model can have low forgetting and still be poor at all tasks if the EWC penalty prevents adaptation, or it can have modest forgetting while losing the large gains shown in Tables I-IV. Without absolute final performance on all three tasks after the deHaze->deSnow->deRain sequence, the central claim that EWC yields a competitive unified model is unsupported. This is compounded by the choice of lambda=750 and lambda1<lambda2 being made on the same benchmark test sets, with no error bars or held-out validation. A further internal inconsistency appears in the ablations: in Table VII, AGLC-GAN + SK-Fusion has SSIM 0.9437 vs. DA-AGLC-GAN 0.9432 on SRRS, and in Table VIII, AGLC-GAN + SK-Fusion has PSNR 32.37 vs. DA-AGLC-GAN 32.33 on Rain100H, so the full model does not consistently outperform its own components.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DA-AGLC-GAN, an unpaired CycleGAN-based restoration model intended to handle dehazing, desnowing, and deraining in a single framework. The architecture extends the authors' AGLC-GAN with Selective Kernel Fusion layers, a cycle-contrastive loss, and Elastic Weight Consolidation for sequential task learning. The paper reports single-task results on RESIDE ITS/OTS, SRRS, and Rain100H, compares them with paired and unpaired baselines, and presents an EWC-λ analysis plus component ablations. The central claim is that the proposed model outperforms state-of-the-art unpaired methods and mitigates catastrophic forgetting when tasks are learned sequentially.","tokens_in":16363,"tokens_out":7514,"duration_ms":64227,"significance":"If the claimed results are reproducible, the work would be a useful contribution to unpaired multi-weather image restoration, and the ablation study gives a clear picture of each component's standalone effect. The comparison includes a wide range of paired and unpaired baselines, and the authors are transparent about the EWC parameter analysis. However, the central continual-learning claim is not directly demonstrated: the sequentially trained model's absolute PSNR/SSIM on the three benchmarks is never reported, and the EWC hyperparameters are tuned on the same test sets used for the headline numbers. The absence of error bars, an internal contradiction between Tables IV and VIII, and an ablation outcome in which the full model loses to one of its components on two metrics all make the quantitative conclusions difficult to assess.","major_comments":[{"comment":"The continual-learning experiment reports only forgetting values and λ sweeps; no table or figure gives the final PSNR/SSIM of the sequentially trained DA-AGLC-GAN on the RESIDE ITS/OTS, SRRS, or Rain100H test splits after the deHaze→deSnow→deRain sequence. Since forgetting (Eq. 18) is a relative metric, a model can have low forgetting while remaining poor on all tasks if the EWC penalty prevents adaptation. Please report the absolute final metrics for all three tasks, alongside the single-task results from Tables I–IV, so the unified model's performance can be directly compared.","section":"V-D, Table V, Eq. (18)"},{"comment":"The EWC regularization strength λ is chosen by sweeping 0–1000 and examining PSNR/SSIM and forgetting on the same benchmark test sets used for the headline results, and the λ1/λ2 comparison in Table V is evaluated on the same test sets. This means the 'optimal' configuration is partly a fitting outcome on the test benchmarks. Please use a held-out validation split or an inner cross-validation, and report variation over multiple seeds or runs.","section":"V-D, Figs. 7–8, Table V"},{"comment":"The ablation tables do not consistently support the claim that DA-AGLC-GAN outperforms its components: on SRRS, AGLC-GAN + SK-Fusion has SSIM 0.9437 versus DA-AGLC-GAN's 0.9432 (Table VII), and on Rain100H, AGLC-GAN + SK-Fusion has PSNR 32.37 versus DA-AGLC-GAN's 32.33 (Table VIII). The statement that DA-AGLC-GAN 'consistently outperforms' the component variants is therefore contradicted; please discuss this outcome and its implications for the fusion design.","section":"V-E, Tables VII–VIII"},{"comment":"The same model and dataset, DA-AGLC-GAN on Rain100H, are reported with PSNR 32.23 in Table IV and 32.33 in Table VIII. This internal inconsistency needs to be resolved, and it raises doubt about the reliability of other reported numbers; please verify all benchmark entries against the exact evaluation protocol.","section":"Tables IV and VIII"},{"comment":"The cycle-contrastive loss is called 'novel' in the abstract and contributions, but the formulation in Eq. (8), the temperature τ=0.07, and the number of negatives N=64 match CCLGAN [14], which is cited only as motivation. If the loss is identical, please credit CCLGAN and revise the novelty claim; if it differs, specify the differences explicitly.","section":"III-C, Eq. (8)"},{"comment":"All reported PSNR/SSIM values appear to be single runs with no error bars or statistical tests. The improvements over the AGLC-GAN baseline are modest in several cases (e.g., 31.69→32.31 on RESIDE ITS, 36.71→37.13 on OTS, 31.79→32.23 on Rain100H), so without variance estimates it is unclear whether these differences are significant. Please report multiple seeds or at least state the run-to-run variability.","section":"V, Tables I–VIII"}],"minor_comments":[{"comment":"The caption of Figure 6 says 'Results of the proposed model on ITS/OTS dataset' but the figure shows rain removal on Rain100H; please correct the dataset reference.","section":"Fig. 6"},{"comment":"Section IV-B reports 60,000 training iterations for all datasets, while Section V-A states 5 epochs with 13k samples for ITS and 1 epoch with 60k samples for OTS; please reconcile these training budgets.","section":"IV-B and V-A"},{"comment":"There is a typo, 'dependance', in the abstract; it should be 'dependence'.","section":"Abstract"},{"comment":"The text says Figure 8 shows final PSNR/SSIM on Task 3 (deRain), but the figure caption says it shows deSnow (Task 2) and deRain (Task 3); please make the description consistent.","section":"Fig. 8 and V-D"},{"comment":"The header notation 'F-PSNR (ITS T1 →2)' is unclear: forgetting is defined in Eq. (18) as the drop on Task 1 after Tasks 2 and 3, but the table appears to use different pairings. Please define all column abbreviations explicitly.","section":"Table V"},{"comment":"The abstract claims improvements in 'perceptual quality', but the paper reports only PSNR and SSIM; please either include a perceptual metric such as LPIPS or NIQE, or soften the claim.","section":"Abstract and V"},{"comment":"Reference [16] is titled 'Cycle-dehaze' but is cited in Section II-B as a CycleGAN baseline, and [27] is the CycleGAN paper; please check that each citation points to the intended work.","section":"References [16] and [27]"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's novelty claims depend heavily on the contrastive loss, which appears to be a direct reuse of CCLGAN's formulation, and the improvements over the authors' own AGLC-GAN baseline are small enough that the missing error bars are a real concern. The missing absolute metrics for the sequentially trained model are fixable within the manuscript's scope, so I recommend major revision rather than rejection, but the authors should be asked to clarify prior art and to provide a validation-based selection procedure for the EWC hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a decent engineering effort but a weak scientific one. The authors take their own AGLC-GAN backbone, add SK fusion layers from DRGNet, a cycle-contrastive loss from CCLGAN (same temperature, same negative sampling), and EWC for continual learning. That is not a novel contribution; it is a recombination. The credit it does earn is in framing—unpaired multi-weather restoration via continual learning is a sensible problem—and in being fairly explicit about where each component comes from.\n\nThe single-task results are positive but marginal: PSNR gains over AGLC-GAN are 0.4–0.9 dB across the four benchmarks. There are no error bars, no significance tests, and the hyperparameters (EWC λ, contrastive weight, temperature) are tuned on the same test sets used for the headline numbers. That alone would make me cautious about over-reading the improvements.\n\nThe bigger problem is that the central claim—a continual-learning unified model—is not actually tested. Sections V-A to V-C validate the architecture as three separate single-task models. Section V-D, the only continual-learning experiment, reports forgetting metrics and λ sweeps but never gives the final PSNR/SSIM of the sequentially trained model on all three benchmarks. Forgetting is a relative metric; a model with low forgetting can still be poor at everything if the EWC penalty blocks adaptation. As written, the unified model could be strictly worse than the single-task versions and the paper would not show it.\n\nThere are also internal inconsistencies that look sloppy rather than fatal: Table IV reports DA-AGLC-GAN on Rain100H at 32.23 dB, while Table VIII reports 32.33 dB for the same model. And in the ablations, AGLC-GAN + SK-Fusion beats the full DA-AGLC-GAN on SRRS SSIM (0.9437 vs 0.9432) and Rain100H PSNR (32.37 vs 32.33), so the full combination is not consistently better than one of its own components.\n\nBottom line: the problem is meaningful and the flaws are fixable in principle, but the current manuscript does not support its headline claim. I would not cite it in its current form, and I would not bring it to a reading group as a model of rigor. That said, I would not desk-reject it either—a serious referee could force the authors to report the sequential model's absolute performance and add proper validation, and if those numbers hold, it becomes a useful incremental data point.","headline":"A plausible recombination of known components whose central continual-learning claim is not actually demonstrated by the reported experiments; worth referee time only if the authors can supply the missing absolute-performance numbers.","tokens_in":16876,"tokens_out":2211,"would_cite":false,"duration_ms":22011,"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":"A single unpaired model can learn to dehaze, desnow, and derain without catastrophic forgetting.","keywords":["image restoration","unpaired learning","continual learning","catastrophic forgetting","cycle-consistent GAN","selective kernel fusion","cycle-contrastive loss","adverse weather removal"],"falsifier":"Train the model with λ chosen on a held-out validation split and with the task order reversed (derain → desnow → dehaze); if the PSNR advantage over AGLC-GAN on RESIDE, SRRS, and Rain100H disappears or forgetting on the first task grows beyond the reported roughly 3–5 dB, the headline results depend on the benchmark-tuned λ=750 schedule rather than on the architecture itself.","tokens_in":15775,"feed_emoji":"🌦️","tokens_out":8830,"duration_ms":78446,"temperature":0.7,"pith_summary":"This paper argues that a single unpaired image-restoration model can handle three different weather degradations—haze, snow, and rain—learned one after another, without the usual catastrophic forgetting. The authors build on a cycle-consistent GAN for unpaired dehazing and add three mechanisms: selective-kernel fusion layers that adaptively combine multi-scale feature maps, a cycle-contrastive loss that keeps content aligned across domain translation, and elastic weight consolidation (EWC) that protects parameters important to previously learned tasks. On standard benchmarks—RESIDE for dehazing, SRRS for desnowing, and Rain100H for deraining—they report PSNR and SSIM gains over existing unpaired methods, and higher PSNR than several paired methods. The practical payoff the paper seeks is a unified restoration model suitable for applications like autonomous driving, where weather conditions vary and paired training data is scarce.","feed_headline":"One model learns haze, snow, and rain removal","feed_subtitle":"Elastic weight consolidation stops the network from forgetting earlier tasks as it trains on new weather.","key_machinery":"The load-bearing machinery is a set of three additions to the AGLC-GAN backbone. Selective Kernel (SK) Fusion layers concatenate a skip-connection feature map and a main-path feature map, compute attention weights via global average pooling followed by a small MLP, and output a weighted sum a1x1 + a2x2, so the network can emphasize the scale most informative for a given degradation. The Cycle-Contrastive Loss samples 64 spatial patches from features extracted at four generator depths, treats the patch at the same location in the reconstructed image as the positive and patches at other locations as negatives, and applies a temperature-scaled cosine-similarity softmax; this pulls corresponding content together while pushing unrelated content apart. Elastic Weight Consolidation adds a Fisher-information-weighted quadratic penalty on parameter deviations from the previously learned solution, which is what lets the model train task 2 and task 3 without erasing task 1. Each component is ablated separately in the paper, and each contributes a PSNR/SSIM improvement on at least one of the three benchmarks.","core_discovery":"The paper's central claim is that DA-AGLC-GAN—a cycle-consistent adversarial network formed by adding selective-kernel fusion layers, a cycle-contrastive loss, and elastic weight consolidation to AGLC-GAN—restores hazy, snowy, and rainy images better than the listed unpaired methods, and reports higher PSNR than several paired methods, while the EWC schedule keeps the model from forgetting dehazing when it later learns desnowing and deraining. The single-task experiments report 32.31 dB PSNR / 0.9697 SSIM on RESIDE ITS, 37.13 dB / 0.9793 on RESIDE OTS, 35.53 dB / 0.9432 on SRRS, and 32.23 dB / 0.8434 on Rain100H, each above the AGLC-GAN baseline. The continual-learning experiments sweep the EWC strength λ from 0 to 1000, settle on λ=750, and report forgetting of the earlier task below the no-regularization levels. The claim is therefore twofold: the architectural additions improve per-task restoration quality, and EWC makes the three-task sequence learnable in one model.","pith_inferences":["A testable extension is to apply the same unpaired continual-learning recipe to other restoration domains, such as low-light enhancement, underwater color correction, or super-resolution, since neither the selective-kernel fusion nor the cycle-contrastive loss is weather-specific.","Because the optimal λ was selected on the same benchmarks that produce the headline numbers, the safest reading is that the architecture improves per-task quality while the specific forgetting numbers depend on that tuning; an out-of-sample validation split would settle how much.","The paper compares against only a handful of unpaired baselines on the snow and rain benchmarks, so a natural next check is how DA-AGLC-GAN fares against more recent unpaired all-in-one restoration models rather than mainly against the CycleGAN family.","If the forgetting asymmetry (consolidate later tasks more strongly) is a general property, task ordering itself becomes a design choice: starting with the hardest or most critical degradation and increasing consolidation strength over time could improve final performance beyond the three-task sequence reported here."],"forward_implications":["A single set of weights can serve dehazing, desnowing, and deraining, so a deployed system would not need to detect the weather condition or switch models.","Because the method trains on unpaired images, it can in principle be transferred to real-world degraded images where clean/degraded pairs do not exist.","The EWC configuration with a stronger penalty on the second consolidation step (λ2=800 vs λ1=700) yields both better final deraining metrics and lower forgetting of the dehazing task, pointing to an asymmetry in how strongly later tasks should be consolidated.","On the reported benchmarks, the model exceeds every listed unpaired baseline, showing that unpaired training does not have to concede quality on these restoration tasks.","The ablations show each added component—SK fusion and cycle-contrastive loss—improves at least one benchmark over the AGLC-GAN baseline, so the reported gains do not rest on a single ingredient."],"supporting_citations":[{"why":"supplies the AGLC-GAN cycle-consistent backbone and the global-local attention baseline that the proposed model extends and compares against.","marker":"[11]"},{"why":"provides the Fisher-information weight-consolidation penalty used to mitigate catastrophic forgetting across the three tasks.","marker":"[44]"},{"why":"provides the cycle-contrastive loss formulation and the temperature value 0.07 used for feature alignment.","marker":"[14]"},{"why":"introduces selective kernel networks, the multi-scale attention mechanism behind the SK Fusion layers.","marker":"[40]"},{"why":"motivates the recursive-gated fusion design that the paper adapts for fusing skip-connection and main-path features.","marker":"[41]"},{"why":"provides the RESIDE ITS and OTS datasets used for dehazing training and evaluation.","marker":"[45]"},{"why":"provides the SRRS dataset used to train and test the desnowing task.","marker":"[46]"},{"why":"provides the Rain100H dataset used to train and test the deraining task.","marker":"[47]"},{"why":"supplies the CycleGAN adversarial and cycle-consistency losses that form the training backbone.","marker":"[27]"}],"fun_headline_variants":["One model erases fog, snow, and rain without forgetting","Weather-agnostic restoration: one model, zero forgetting","Dehaze, desnow, derain: a single continual learner","No catastrophic forgetting in all-weather image cleanup","Unified model tackles haze, snow, rain via continual learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the performance edge is not an artifact of tuning the regularization strength (λ=750) and the task order on the same test benchmarks that produced the headline results.","fun_headline_variants_meta":{"raw":{"variants":["One model erases fog, snow, and rain without forgetting","Weather-agnostic restoration: one model, zero forgetting","Dehaze, desnow, derain: a single continual learner","No catastrophic forgetting in all-weather image cleanup","Unified model tackles haze, snow, rain via continual learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1713,"prompt_tokens":979,"completion_tokens":734,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":650}},"tokens_in":595,"tokens_out":734,"duration_ms":7059,"temperature":1.0,"reasoning_tokens":650,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:58:10.041446+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model with λ chosen on a held-out validation split and with the task order reversed (derain → desnow → dehaze); if the PSNR advantage over AGLC-GAN on RESIDE, SRRS, and Rain100H disappears or forgetting on the first task grows beyond the reported roughly 3–5 dB, the headline results depend on the benchmark-tuned λ=750 schedule rather than on the architecture itself.","supporting_citations":[{"cited_title":"Aglc-gan: Attention-based global- local cycle-consistent generative adversarial networks for unpaired single image dehazing,","cited_arxiv_id":null,"evidence_quote":"supplies the AGLC-GAN cycle-consistent backbone and the global-local attention baseline that the proposed model extends and compares against."},{"cited_title":"Overcoming catastrophic forgetting in neural networks,","cited_arxiv_id":null,"evidence_quote":"provides the Fisher-information weight-consolidation penalty used to mitigate catastrophic forgetting across the three tasks."},{"cited_title":"Cycle Contrastive Adversarial Learning for Unsupervised image Deraining","cited_arxiv_id":"2407.11750","evidence_quote":"provides the cycle-contrastive loss formulation and the temperature value 0.07 used for feature alignment."},{"cited_title":"Selective kernel networks,","cited_arxiv_id":null,"evidence_quote":"introduces selective kernel networks, the multi-scale attention mechanism behind the SK Fusion layers."},{"cited_title":"Efficient dehazing with recursive gated convolution in u-net: A novel approach for image dehazing,","cited_arxiv_id":null,"evidence_quote":"motivates the recursive-gated fusion design that the paper adapts for fusing skip-connection and main-path features."},{"cited_title":"Joint rain detection and removal from a single image,","cited_arxiv_id":null,"evidence_quote":"provides the Rain100H dataset used to train and test the deraining task."},{"cited_title":"Unpaired image-to-image translation using cycle-consistent adversarial networks,","cited_arxiv_id":null,"evidence_quote":"supplies the CycleGAN adversarial and cycle-consistency losses that form the training backbone."}],"review_version":2}