{"id":"b9ea968a-48fb-401a-af0b-96a4941248fe","arxiv_id":"2412.03019","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An unsupervised, iterative layer-separation network with a feedback loop outperforms prior unpaired-image methods for single-image raindrop removal.","lead":"This paper trains a neural network to remove raindrops from single photos without needing matching clean versions of the same scenes. It splits each rainy image into a clean background, a raindrop layer, and a mask, using an iterative feedback loop, and reports improved scores on two standard raindrop benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim omits RainGAN, a directly comparable unsupervised raindrop removal method; without its reported results the claim that the model surpasses all unsupervised methods is unverified.","rationale":"The reader's verdict is CONDITIONAL with MODERATE confidence, and the internal issues they list (identity loss mis-specification, no code, arithmetic mismatch) are all real. However, the single most load-bearing threat to the paper's headline assertion is the omission of RainGAN from the experimental comparison. The paper's own literature review identifies RainGAN as an unsupervised raindrop removal method that also uses decomposition and composition, so it is not an obscure baseline; it is the most direct competitor. A state-of-the-art claim is only as strong as the set of methods it is compared against, and here the most relevant one is absent. This is a concrete, checkable gap rather than a speculative modeling concern. The alpha-blend model concern raised in the reader's weakest_assumption is valid as a limitation, but the benchmark results would not be invalidated by the model being approximate; the missing baseline, by contrast, can directly overturn the claimed performance advantage. I therefore recommend no change to the CONDITIONAL verdict: the paper should be accepted only after the RainGAN comparison is added (or the claim is narrowed to 'outperforms the four tested baselines'). No independent support (code, formal verification) exists in the manuscript to make the claim self-verifying.","tokens_in":14295,"tokens_out":10497,"duration_ms":92485,"concrete_test":"Locate RainGAN (Yan & Loke, WACVW 2022, doi:10.1109/WACVW54805.2022.00007) and extract its reported PSNR/SSIM on the NUS raindrop test_a/test_b and on RainDS_real. Compare directly with the paper's results: test_a 27.0562/0.8738, test_b 24.7124/0.8281, RainDS 21.7624/0.7472. If RainGAN exceeds any of these on the same split, the state-of-the-art claim is refuted; if RainGAN is not evaluated on these datasets, the paper should at minimum add a comparison after retraining RainGAN under the same protocol. This single check settles whether the missing baseline changes the headline conclusion.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed unsupervised layer-separation network achieves state-of-the-art raindrop removal among unpaired-image methods (Section 4.2: 'surpasses the unsupervised model by 1.5938 dB of PSNR on test_a and 1.4679 dB on test_b'; Section 4.3: 'outperforms all other methods with 21.7624 dB PSNR and 0.7472 SSIM on the RainDS dataset'). For this claim to hold, the comparison set must include all relevant unsupervised raindrop removal methods. The paper itself cites RainGAN (Yan et al. [40], WACVW 2022, 'RainGAN: Unsupervised raindrop removal via decomposition and composition') in the Introduction as an existing unsupervised method, yet the experimental comparison (Tables 1 and 2) includes only UA_GAN, AGGAN, CycleGAN, and the weakly supervised model [21]. RainGAN is arguably the closest baseline: it also performs unsupervised decomposition and composition, the same problem and paradigm as the proposed method. Omitting it from the quantitative evaluation leaves the state-of-the-art claim unverified. If RainGAN achieves higher PSNR/SSIM on NUS or RainDS, the central claim is false. No code is released, so the only way to assess the claim is to check the published numbers of the omitted competitor.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised single-image raindrop removal method that decomposes a rainy image into a clean background, a transparency mask, and a raindrop layer using the linear blend model of Eq. (1). The architecture combines a CycleGAN-style generator/discriminator setup with an iterative feedback mechanism that reuses the transparency mask from the previous iteration as additional input. The training losses are a weighted combination of GAN, cycle-consistency, identity, and sparsity terms. Experiments on the NUS and RainDS datasets report PSNR/SSIM gains over four unsupervised baselines, and an ablation study shows each loss and the iterative module contribute to the final performance. The paper's central claim is that this is a state-of-the-art unsupervised raindrop removal method, quantitatively surpassing the previous best by about 1.5 dB PSNR on NUS and achieving 21.76 dB on RainDS.","tokens_in":14555,"tokens_out":5245,"duration_ms":49097,"significance":"If the empirical results hold, the paper offers a practical unsupervised alternative to paired-data raindrop removal, with the feedback mechanism and sparsity-constrained decomposition being reasonable contributions. The ablation study is a useful step toward understanding the contribution of each loss. However, the significance is currently undercut by two load-bearing issues: (1) the identity loss in Eq. (6) is not an identity loss in any CycleGAN sense, because it compares the decomposed background of a rainy image to an unrelated clean image; and (2) the state-of-the-art claim is made without comparing against RainGAN, a directly applicable unsupervised raindrop removal method cited in the paper's own introduction. The degradation model of Eq. (1) is also a strong simplification that the paper does not critically assess.","major_comments":[{"comment":"The identity loss is mathematically and conceptually incorrect. The text states that 'sending B to G should still generate B' and therefore defines L_identity as E_B[||B_i - B||_1], where B_i is the generator output for a rainy input I and B is a random clean image from the training set. These are not corresponding images of the same scene, so the L1 penalty does not measure any identity property; it merely pulls the output toward arbitrary clean pixels. The ablation result in Table 3 for 'w/o L_identity' cannot be interpreted as validating an identity-preserving mechanism. This needs to be clarified: either the loss should be defined for the same input passed through the generator, or the paper should explain why an unpaired L1 term is a useful regularizer.","section":"§3.3, Eq. (6)"},{"comment":"The claim of outperforming all unsupervised methods is unverified because RainGAN (Yan et al. [40], WACVW 2022), an unsupervised raindrop removal method based on decomposition and composition, is cited in Section 1 but omitted from the experimental comparisons. Section 4 states that 'the unsupervised model [21] is the only one we can find in the literature specifically designed for raindrop removal,' which is contradicted by the authors' own reference to [40]. Without quantitative results for RainGAN on NUS and RainDS, the state-of-the-art assertion is unsupported. The paper should either add this baseline or explicitly justify its exclusion.","section":"§4.2 and Tables 1-2"},{"comment":"The method does not implement a cycle-consistency loss between two domains in the sense of CycleGAN. The 'cycle' here is a self-reconstruction: the generator decomposes the rainy input I and the reconstruction recombines the predicted layers to approximate I. Equation (5) is therefore a standard reconstruction loss, not a cycle-consistency loss. The terminology overstates the architectural novelty and may mislead readers about the relation to CycleGAN. The authors should describe this as a reconstruction loss and avoid the claim that they are using a cycle network.","section":"§3.3 and Fig. 2"},{"comment":"The degradation model in Eq. (1) assumes every raindrop effect can be expressed as a per-pixel linear blend of a clean background B and a raindrop layer R with a single transparency mask α. The paper itself notes in Section 1 that raindrops 'alter the focus of both foreground and background, leading to unexpected blurring,' and that they cause background distortion. Refraction and defocus are not obviously captured by one alpha blend per pixel. While the benchmarks show good results, this modeling limitation should be acknowledged as a potential barrier to generalization; the current text presents Eq. (1) as if it were an exact physical model without discussing its validity.","section":"§1 and Eq. (1)"},{"comment":"The loss weighting is specified inconsistently. Equation (7) introduces K_i = i - 1 as a per-iteration weight for the GAN loss, but Section 4.1 then states that 'the hyper-parameters β1∼4 are 2 × 1.5^{i−1}, 10, 5, and 1,' implying an iteration-dependent β1 that does not appear in Eq. (9) where β1 is a single scalar. This makes the exact training objective ambiguous and hampers reproducibility. Please clarify whether the per-iteration weights are K_i, β1, or both, and give the exact final expression.","section":"§4.1, Eqs. (7) and (9)"}],"minor_comments":[{"comment":"The abstract and Section 3.3 refer to a 'cycle network architecture' and 'cycle structure,' but the actual mechanism is a single-domain decomposition and reconstruction; consider using 'reconstruction' instead of 'cycle' to avoid confusion.","section":"Abstract and §3.3"},{"comment":"The text contains a typo: 'Combing' should be 'Combining' in the sentence introducing the total loss.","section":"§3.4"},{"comment":"There are two minor language errors: 'the qualification of the generated clean images decays' should likely be 'quality,' and 'we can also obverse' should be 'observe.'","section":"§4.5"},{"comment":"The order of the methods in the figure caption (Input, UA_GAN, AGGAN, CycleGAN, Ours, GT) does not match the order described in the text of §4.2; please align them.","section":"Fig. 6"},{"comment":"Reference [30] (Stollenga et al.) is missing the publication year and venue in the bibliography.","section":"References"},{"comment":"The paper states that the network structure is 'borrowed from CycleGAN' but does not provide the specific architecture of the generator (e.g., number of Dense ResNet blocks, channels). Adding these details would aid reproducibility.","section":"Implementation details"}],"recommendation":"major_revision","confidential_remarks":"The paper's central empirical claims are plausible, but the omission of RainGAN is a serious comparison gap that must be addressed before the state-of-the-art claim can be taken at face value. The identity loss issue also suggests a conceptual misunderstanding that may require rewriting part of the methodology. If the authors can correct these points and add the missing baseline, the paper may become suitable for publication. I do not see a fundamental circularity or irreparable flaw, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable unsupervised raindrop-removal paper with a new combination of ideas, but the headline claim that it beats all unsupervised methods is not supported by the evidence presented, because RainGAN is left out of the comparison.\n\nWhat's actually new: the paper combines layer separation in the style of Double-DIP with an iterative feedback generator, trained under a CycleGAN-style unpaired objective. Using feedback iterations for unsupervised raindrop removal is a genuine extension, and the ablation study shows each component (cycle loss, identity loss, sparsity loss, iterative network) helps. The PSNR gains over the included baselines check out against the table; the text's \"comparable SSIM on test_a\" is slightly off because the model is actually 0.0025 lower, but that's minor.\n\nThe load-bearing problem is the missing baseline. The paper cites RainGAN (Yan and Loke, WACVW 2022) in the introduction as an unsupervised raindrop-removal method, then in Section 4 claims the weakly supervised model [21] is \"the only one we can find in the literature specifically designed for raindrop removal.\" That is internally inconsistent. RainGAN does exactly this task—unsupervised decomposition and composition for raindrop removal—and is the closest competitor. Without its numbers on NUS or RainDS, the claim to surpass all unsupervised methods is unverified, and if RainGAN does better, the central claim is false.\n\nThere is also a formulation issue with the identity loss. Equation (6) compares the decomposed background B_i to real clean images from the dataset, not to G(B). As written, it is not a CycleGAN identity loss; it's a distribution-preserving term that may over-constrain the output. The \"cycle\" in Eq. (5) is a self-reconstruction of the same rainy image, not a domain cycle, so the architecture's connection to CycleGAN is looser than the paper suggests. These are fixable, and the method may still work as an autoencoder-style decomposition with a sparsity prior, but the text should be corrected.\n\nThe linear model I = (1-alpha)⊙B + alpha⊙R is a simplification; refraction and defocus are not obviously one alpha blend. The real-data results on NUS and RainDS partly mitigate this, but the model could struggle on more varied real images.\n\nNo code is provided, which makes independent verification harder.\n\nBottom line: this paper is worth a serious peer review. The method is novel, the ablation is honest, and the problem is practical. But the authors need to add RainGAN to the comparison, fix the identity loss description, and tone down the SOTA claim. I'd recommend major revision.","headline":"A plausible unsupervised raindrop-removal method with a solid ablation study, but its state-of-the-art claim is undercut by the omission of RainGAN, a directly comparable method cited in the paper itself.","tokens_in":15103,"tokens_out":3997,"would_cite":false,"duration_ms":35886,"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":"This paper establishes that an unsupervised cycle-consistent layer-separation network, guided by an iteratively refined transparency mask, outperforms all unpaired-image baselines for single-image raindrop removal on the NUS and RainDS…","keywords":["raindrop removal","unsupervised learning","layer separation","cycle-consistent GAN","feedback mechanism","iterative neural network","transparency mask","single image restoration"],"falsifier":"Synthesize a raindrop as a pure refractive displacement of the background (no intensity change), feed it to the trained model, and compare the restored background against the original; if the displaced edge remains as a ghost and PSNR drops far below the reported roughly 27 dB, the single-alpha blend assumption fails.","tokens_in":14081,"feed_emoji":"🌧️","tokens_out":9711,"duration_ms":79496,"temperature":0.7,"pith_summary":"This paper tackles single-image raindrop removal without requiring paired rainy and clean training images. It proposes a network that decomposes a rainy image into a clean background, a raindrop layer, and a transparency mask according to the linear blend model $I = (1-\\alpha)\\odot B + \\alpha\\odot R$. A cycle-consistent adversarial structure trained on unpaired image sets, combined with an iterative feedback loop that refines the mask across six shared-weight iterations, lets the model improve its separation progressively. On the NUS raindrop dataset and the RainDS real subset, the model reports the highest PSNR and SSIM among the unsupervised and weakly supervised baselines it compares against. The paper argues that unpaired training makes raindrop removal practical when paired real-world data is scarce.","feed_headline":"Raindrop removal without paired data tops two benchmarks","feed_subtitle":"Only needs unpaired rainy and clean images, yet beats all compared unsupervised methods on NUS and RainDS.","key_machinery":"The load-bearing identity is the image-decomposition model $I = (1-\\alpha)\\odot B + \\alpha\\odot R$, which assumes every raindrop effect is a per-pixel linear blend of a clean background $B$ and a raindrop layer $R$ with a single transparency mask $\\alpha$. Around this model, the machinery is a cycle-consistent GAN with one generator and two PatchGAN discriminators, trained with adversarial, cycle-consistency, identity, and sparsity losses. The distinctive mechanism is the iterative feedback network: $N$ sub-networks share weights, and the output mask $\\alpha^{i-1}$ from iteration $i-1$ is concatenated with the input rainy image to feed iteration $i$, so the mask is refined from coarse to fine. Each iteration computes its own losses, later iterations are weighted more heavily by $K_i = i-1$, and the clean background from the last iteration is the final output.","core_discovery":"The discovery is that unpaired, cycle-consistent layer separation can beat existing translation-based unsupervised methods for raindrop removal when the separation is driven by a transparency mask that is refined iteratively. The generator splits $I$ into $B$, $R$, and $\\alpha$; the recomposition $F(B,R,\\alpha) = (1-\\alpha)\\odot B + \\alpha\\odot R$ must reproduce $I$ (cycle loss), while $B$ is pushed toward the clean-image distribution by a PatchGAN discriminator and an identity loss, and $\\alpha$ is pushed toward zero (sparsity). Feeding the previous iteration's $\\alpha$ back as extra input to the next of $N$ shared-parameter iterations yields coarse-to-fine masks. On the NUS dataset the model reaches $27.0562$ dB PSNR and $0.8738$ SSIM on test_a and $24.7124$ dB and $0.8281$ on test_b, surpassing the weakly supervised baseline by $1.5938$ and $1.4679$ dB respectively; on RainDS it reaches $21.7624$ dB and $0.7472$ SSIM, the best among the compared unpaired methods.","pith_inferences":["The same layer-separation plus feedback recipe could plausibly transfer to other adherent-occlusion problems — dust, mud, frost, or water on a lens — since none of the losses is specific to raindrop physics.","Because the linear blend model cannot represent refraction, the reported gains over translation baselines may shrink on images with large, strongly refractive drops; a dedicated test set of close-up droplet images would reveal whether a physics-aware decomposition is needed.","The sparsity loss encodes the prior that raindrops occupy a small fraction of pixels; on images with dense, overlapping droplets this prior could suppress legitimate mask values and cap performance.","A direct comparison with fully supervised methods on the same test splits would quantify how much accuracy is sacrificed for unpaired training, and could motivate a semi-supervised hybrid that uses a few paired samples alongside the unpaired collections."],"forward_implications":["Unpaired training removes the need for laboriously captured paired raindrop and clean photos, so the method can be re-targeted to a new camera or environment using only independent clean and rainy collections.","The learned transparency mask localizes the raindrops explicitly, giving a detection map as a by-product of restoration.","The feedback loop allows a fixed-weight network to be run for any number of refinement iterations, trading inference compute against restoration quality without retraining.","The decomposition into background, raindrop layer, and mask supports editing beyond restoration, such as removing individual drops or re-synthesizing rain.","On the NUS and RainDS benchmarks tested, the method reports higher PSNR and SSIM than all four unpaired baselines, with the largest margins on the harder test splits."],"supporting_citations":[{"why":"Supplies the layer-separation principle (an image as a mixture of simple layers) that motivates decomposing a rainy image into background, raindrop layer, and transparency mask.","marker":"[11]"},{"why":"Provides the cycle-consistent adversarial framework and generator/discriminator design that lets the network train on unpaired rainy and clean image sets.","marker":"[49]"},{"why":"Introduces the NUS raindrop dataset, the attention-recurrent raindrop removal baseline, and the train/test splits and evaluation protocol that the paper follows.","marker":"[25]"},{"why":"Provides the RainDS dataset with paired raindrop and rain-free real images used as the second benchmark.","marker":"[26]"},{"why":"The weakly supervised raindrop removal model whose published PSNR/SSIM numbers the paper uses as the primary unsupervised baseline to surpass.","marker":"[21]"},{"why":"Establishes the feedback mechanism for refining low-level representations with high-level information, which the paper adapts into its iterative neural network for mask refinement.","marker":"[20]"},{"why":"An attention-guided unpaired image translation model (AGGAN) that serves as a comparison baseline on both datasets.","marker":"[31]"},{"why":"An attention-guided translation model (UA_GAN) used as a baseline in quantitative and qualitative raindrop removal comparisons.","marker":"[1]"}],"fun_headline_variants":["Unpaired raindrop removal beats unsupervised baselines","Cycle-consistent mask feedback removes raindrops without pairs","Transparency-mask iteration drives raindrop removal unpaired","Feedback loop refines transparency mask for raindrop removal","Unsupervised cycle network with mask feedback tops unpaired methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a raindrop's visual effect can be undone by blending the clean background with a raindrop layer using a single per-pixel transparency mask, even though the paper notes that raindrops refract and defocus light from both foreground and background.","fun_headline_variants_meta":{"raw":{"variants":["Unpaired raindrop removal beats unsupervised baselines","Cycle-consistent mask feedback removes raindrops without pairs","Transparency-mask iteration drives raindrop removal unpaired","Feedback loop refines transparency mask for raindrop removal","Unsupervised cycle network with mask feedback tops unpaired methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001768,"raw_usage":{"total_tokens":7001,"prompt_tokens":1000,"completion_tokens":6001,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":5919}},"tokens_in":616,"tokens_out":6001,"duration_ms":41776,"temperature":1.0,"reasoning_tokens":5919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:50:49.086334+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize a raindrop as a pure refractive displacement of the background (no intensity change), feed it to the trained model, and compare the restored background against the original; if the displaced edge remains as a ghost and PSNR drops far below the reported roughly 27 dB, the single-alpha blend assumption fails.","supporting_citations":[{"cited_title":"Double-DIP","cited_arxiv_id":null,"evidence_quote":"Supplies the layer-separation principle (an image as a mixture of simple layers) that motivates decomposing a rainy image into background, raindrop layer, and transparency mask."},{"cited_title":"Attentive generative adversarial network for raindrop removal from a single image, in: IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Introduces the NUS raindrop dataset, the attention-recurrent raindrop removal baseline, and the train/test splits and evaluation protocol that the paper follows."},{"cited_title":"Feedback network for image super-resolution, in: IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Establishes the feedback mechanism for refining low-level representations with high-level information, which the paper adapts into its iterative neural network for mask refinement."},{"cited_title":"Unsupervised attention-guided image-to-image translation, in: Proceedings of the 32nd International Conference on Neural Information Processing Systems, p","cited_arxiv_id":null,"evidence_quote":"An attention-guided translation model (UA_GAN) used as a baseline in quantitative and qualitative raindrop removal comparisons."}],"review_version":1}