{"id":"6e1d66e5-3c65-4ec7-aafb-60e3ed87282e","arxiv_id":"2505.08190","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A diffusion-inpainting raindrop removal system is proposed, but its own experiments report poor masks, unstable reconstructions, and no quantitative results.","lead":"This paper proposes DropWiper, a two-step raindrop removal pipeline that first detects raindrops and then fills the detected regions with a diffusion-based inpainting model. The paper's own results show poor mask quality on real images, unstable reconstruction, and no quantitative evaluation, so the method is not demonstrated to work.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim unsupported: the paper's own experiments report poor masks and no working background reconstruction for the actual test inputs, and Eq. 8 assigns the known-pixel term to the mask instead of its complement.","rationale":"The reader's verdict is REJECT with high confidence, and I reach the same conclusion after an independent pass. My strongest concern is the same one the reader identified: the paper presents no working end-to-end raindrop removal result on the target dataset. The evidence for this is internal and unambiguous. Section IV-C3 explicitly labels the detector masks as poor and identifies dataset domain shift as the cause. Section V explicitly states that effective background reconstruction was not performed because the masks were inadequate. There are no quantitative results anywhere in the paper, no comparison against the Raindrop benchmark, and no figure showing a raindrop-removed image that is then compared to the clean ground truth. The residual-mask alternative cannot rescue the central claim because it uses the clean image (B) to build the mask, so it is not a single-image method and cannot be used at test time when B is unknown. I additionally checked the diffusion update equation, because it is the technical core of the reconstruction module. As written, Eq. 8 multiplies the known-pixel term by m and the unknown-pixel term by (1-m), which is the inverse of the roles defined earlier in the same section (m is the missing/raindrop mask; (1-m) is the known region). This is a real inconsistency, not a stylistic issue: if executed as written, the update keeps known pixels from the generated sample and replaces only the known region with the diffused ground-truth sample, which would destroy the unmasked background and leave the raindrop regions untouched. The paper does not provide the implementation needed to determine whether Eq. 8 is a typo or the code is correct, so the equation stands as a plain technical error. The reader's framing therefore matches mine: the argument is not supported and contains a likely incorrect central equation. REJECT is the right verdict, not CONDITIONAL, because the conditions for acceptance are not missing details; the setup as described cannot establish the claimed method, and the paper itself concedes the key failure. I do not attribute this to bad faith, and the negative domain-shift finding could be useful to future work, but the central scientific claim is not demonstrated.","tokens_in":8234,"tokens_out":2145,"duration_ms":19246,"concrete_test":"Run the released DropWiper inference code end-to-end on the Raindrop test set (set of 249 images): use the trained detector to generate masks, feed them to the DDPM inpainting module, and compute PSNR/SSIM against the provided clean images, reporting a mean and per-image distribution. If the pipeline is genuinely functional, these metrics should be computable and should beat or at least be comparable to the input degraded image; if the reported 'poor masks' and 'unstable reconstructions' make it impossible to produce outputs, or if Eq. 8 is implemented exactly as written, the central claim is refuted. On a separate holdout split of synthetic Cityscapes raindrops (in-distribution for the detector), the same quantitative pipeline should be run as a control to separate domain-shift failure from a fundamentally broken inpainting equation.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that DropWiper removes raindrops from a single image via two modules: a raindrop mask generator and a DDPM-based background restorer. For this claim to hold, the pipeline must, at minimum, run end-to-end on Raindrop inputs and produce restored images that can be compared against ground truth. The paper's own Section IV-C3 reports that the detector's masks on the Raindrop dataset are 'poor quality' and attributes this to domain shift from Cityscapes. Section V then states that because of the lack of a proper mask, 'we are not able to perform an effective background reconstruction' and that forcing reconstruction on poor masks leads to unstable reconstructions. No quantitative metric (PSNR/SSIM) is reported for any removal result. The only alternate mask source, the residual branch, explicitly requires the clean image B (Section III-A1, Eq. 4–7), so it cannot support an unsupervised single-image method. A second, independent technical flaw is that Eq. 8 defines xt-1 = m * x_known + (1-m) * x_unknown, which inverts the paper's own earlier definitions: m was defined as missing/raindrop pixels and (1-m) as known pixels (Section III-B). With the stated assignment, the known-pixel term is multiplied by the mask and the unknown term by its complement. If re-derived with consistent roles, the harness would be the standard RePaint-style update; as written, the equation would corrupt known regions and would not restore masked ones. Whichever correction is intended, the equation as given cannot implement the claimed restoration. Together, these make the headline claim not merely under-evaluated but internally contradicted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DropWiper, a two-step pipeline for removing raindrops from a single image: a lightweight CNN raindrop detector produces a pseudo-mask, and a DDPM-based inpainting module (following the RePaint-style blending of forward and reverse diffusion) reconstructs the background in masked regions. The detector is trained on synthetic raindrops rendered on Cityscapes via a refraction model, and the DDPM is trained on clean Raindrop-dataset images. The experiments compare two mask-generation approaches (residual masks from thresholding the difference between a raindrop image and its clean counterpart, and detector masks) and report that detector masks on the Raindrop dataset are of poor quality and that effective background reconstruction is not achieved.","tokens_in":8619,"tokens_out":3099,"duration_ms":30931,"significance":"The intended contribution is a diffusion-based alternative to GAN-based raindrop removal, combining a small detector with a pre-trained DDPM. This idea is not without interest, but the manuscript does not deliver a working method: no quantitative evaluation is reported, the only mask branch that the paper itself considers 'more solid' requires the ground-truth clean image, and the paper's own conclusion states that effective background reconstruction was not achieved. As a result, the claimed novelty is not demonstrated, and the significance of the work as presented is low.","major_comments":[{"comment":"The paper's central claim of raindrop removal is directly contradicted by its own conclusions. Section IV-C3 states that the detector masks on the Raindrop dataset are 'poor quality' and attributes this to domain shift, and Section V states 'we are not able to perform an effective background reconstruction' and that forcing reconstruction on poor masks leads to 'unstable reconstructions'. The abstract and introduction claim a novel technique for raindrop removal from a single image, but the experiments section offers no end-to-end restoration result, no PSNR/SSIM or any other quantitative metric, and no figure showing an actual raindrop-removed output compared with ground truth. The paper as written does not support its title claim.","section":"Section V and Section IV-C3"},{"comment":"The residual mask branch uses the ground-truth clean image B to compute R = A - B (or variants). This is a load-bearing limitation: for a single-image, unsupervised method, B is not available at test time. The residual masks therefore cannot serve as an unsupervised mask generator for the method; they are usable only in a paired-data setting, which contradicts the stated task and also undermines the conclusion in Section V that residual masks are 'more solid' than detector masks, since that comparison is not an independent test of a deployable pipeline.","section":"Section III-A1, Eqs. (4)-(7)"},{"comment":"Equation (8) inverts the mask definitions stated just above it. The text defines m ⊙ x as 'the missing pixels' and (1 - m) ⊙ x as 'the known pixels', but Eq. (8) computes x_{t-1} = m ⊙ x_known + (1 - m) ⊙ x_unknown. With the stated definitions, the known-pixel term is multiplied by the raindrop mask and the unknown term by its complement, which is the opposite of the standard RePaint-style harmonic update. As written, this equation would corrupt known pixels and would not restore masked regions. This is not a mere presentation typo: it is the core reconstruction equation of the proposed method, and the manuscript's own definitions render it internally inconsistent.","section":"Section III-B, Eq. (8)"},{"comment":"No quantitative evaluation is reported anywhere in the experiments. The only results shown are visual: residual masks (Fig. 6), synthetic raindrop samples (Fig. 7), training/validation losses (Fig. 8), detector masks judged to be poor (Fig. 9), and DDPM samples of clean images (Fig. 10). Figure 10 does not show a single raindrop-removal result. Without any quantitative comparison against the existing GAN-based baselines cited in Section II-A, or even against the ground-truth clean images the paper says it plans to use, the claimed advantage over prior work is unsubstantiated.","section":"Section IV-C and Figure 10"}],"minor_comments":[{"comment":"The text says 'The notations H and W represent the image's height and weight respectively'; 'weight' should be 'width'.","section":"Section III-A1"},{"comment":"The notation in the background-reconstruction section is confusing: x is introduced as the ground-truth image, but Eq. (8) and the surrounding text use x_{t-1} both as the partially denoised latent and as the image being constructed, and the sentence about 'alter the known pixels in (1-m) ⊙ x_{t-1}' is hard to follow. The relationship between the forward-diffusion sample x_known and the reverse-diffusion sample x_unknown should be stated more explicitly.","section":"Section III-B"},{"comment":"The method is called 'conditional diffusion' but no conditioning mechanism beyond the mask-blending is described; the reconstruction module is essentially the RePaint inpainting procedure of [5]. The paper would be more accurate to describe this as an application of RePaint rather than a new conditional diffusion model.","section":"Title and Section II-B"},{"comment":"The phrase 'residual mask generators provide more solid masks than the raindrop detection model' is misleading because the residual masks are computed with access to the ground-truth clean image, whereas the detector is evaluated on unseen raindrop images. The comparison is therefore not apples-to-apples.","section":"Section V"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a negative-result paper that combines RePaint-style DDPM inpainting with a learned raindrop mask generator trained on synthetic refraction-model raindrops. The combination is new in a narrow sense, but the paper does not show it works, and one of the central equations appears to have the mask and its complement swapped. I'd desk-reject it, not because negative results are unwelcome, but because the method as written can't be evaluated and the authors' own conclusion is that it doesn't work.\n\nWhat deserves credit: the failure analysis is honest and specific. Section IV-C3 explicitly says the masks on Raindrop are poor and attributes it to domain shift; Section V states they are not able to perform effective background reconstruction. That kind of transparency is rare, and the negative transfer finding (Cityscapes-trained detector fails on Raindrop) could be useful to people building synthetic-to-real raindrop detectors. The architecture choices are reasonable: lightweight CNN for masks, pre-trained DDPM for inpainting, and the synthetic refraction model follows Hao et al. The code is linked.\n\nWhere it falls apart: there are no quantitative metrics anywhere. The only mask source that produces 'more solid' masks is the residual branch, which subtracts the ground-truth clean image (Eq. 4). That branch is not an unsupervised single-image method; it uses the target. The synthetic detector branch is independent, but it fails. So the pipeline never runs end-to-end with a working mask on the actual test set. Second, Eq. 8 is wrong as written. The paper defines m as missing pixels and (1-m) as known pixels, but Eq. 8 multiplies x_known by m and x_unknown by (1-m). If you follow the text, the equation corrupts the known region and leaves the unknown region untouched. If the roles were meant to be swapped, the paper should say so. Either way, the stated equation cannot implement the claimed restoration.\n\nThe citation pattern is fine; related work is covered. The mask-quality discussion is more qualitative than quantitative, but the main problem is not the writing—it's that the central claim is unsupported by the experiments.\n\nWho is this for? Someone specifically interested in negative transfer for synthetic raindrop detectors might get a small signal. As a paper claiming a working restoration method, it doesn't hold up. I'd not send it to peer review; a serious referee would have to rewrite the paper's experiments from scratch. If the authors address Eq. 8, add PSNR/SSIM on the residual-mask branch, and reframe as a negative result, it could be a workshop note or benchmark report.\n\nFor now, my recommendation: reject/desk-reject, with the honest feedback above.","headline":"Honest negative result, but the method never works and Eq. 8 is backwards; desk-reject.","tokens_in":9092,"tokens_out":2026,"would_cite":false,"duration_ms":18644,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DropWiper proposes to remove raindrops from a single image by generating pseudo-masks and filling them with a diffusion inpainter; its own experiments show the masks are too poor for effective background reconstruction.","keywords":["raindrop removal","diffusion model","image inpainting","DDPM","refraction model","pseudo-mask","single-image restoration","synthetic raindrops"],"falsifier":"Manually annotate raindrop masks on a sample of the Raindrop test set and run the same DDPM inpainter with those exact masks, comparing outputs to the ground-truth clean images. If reconstruction quality is still poor, Equation 8's update is the weak link; if quality is good, the detector's synthetic-to-real transfer is the weak link.","tokens_in":7968,"feed_emoji":"💧","tokens_out":10750,"duration_ms":88927,"temperature":0.7,"pith_summary":"This paper proposes DropWiper, a two-step raindrop-removal pipeline that first generates a pseudo-mask of the raindrop regions and then fills those regions with a Denoising Diffusion Probabilistic Model (DDPM) used for image inpainting. The intended contribution is to swap the usual GAN-based background restoration for diffusion-based inpainting, on the idea that a pre-trained DDPM can reconstruct the background from the masked areas. A sympathetic reading of the proposal is that, given a usable mask, Equation 8's combination of forward- and reverse-diffusion steps should restore the clean image. The paper's own experiments do not confirm this: Section IV-C3 describes the detector's masks as poor and not representing the true raindrop areas, and Section V states that effective background reconstruction was not achieved. The value of the paper lies in laying out the architecture and in reporting exactly where the pipeline breaks.","feed_headline":"Diffusion-based raindrop remover fails at the mask step","feed_subtitle":"Synthetic-trained masks don't match real raindrops, so the diffusion inpainter has nothing proper to fill.","key_machinery":"The load-bearing mechanism is the masked diffusion inpainting update, Equation 8, which merges a forward-noised version of the known pixels with the model's reverse-diffusion prediction for the masked pixels: $$x_{t-1} = m \\odot $x^{{\\text{known}}$}_{t-1} + (1-m) \\odot $x^{{\\text{unknown}}$}_{t-1}.$$ This is the standard masked-inpainting trick from diffusion-based inpainting, letting a DDPM trained only on clean images fill arbitrary masked regions. The second essential component is the pseudo-mask itself: the paper experiments with residual masks (computed from the difference between the raindrop image and the clean image) and with a lightweight CNN detector trained on synthetic raindrops. The argument depends on the mask being accurate enough that the inpainter only has to fill true raindrop areas; when the mask fails, the whole pipeline fails.","core_discovery":"The central claim, on the paper's own terms, is that raindrop removal from a single image can be reformulated as a conditional inpainting problem: detect the raindrops, then use a DDPM to regenerate the detected regions while preserving the known background. The mechanism is the masked inpainting update of Equation 8, which at each reverse step replaces the known pixels with their forward-diffused values and leaves the masked pixels to the denoiser. The paper further claims that the detector can be trained without real ground-truth masks by synthesizing raindrops with a non-parametric refraction model on Cityscapes images. In reporting, the authors find the detector overfits to synthetic data and produces masks that fail on the Raindrop dataset, and consequently that the background reconstruction step cannot be evaluated as intended. The paper's own conclusion is that the two-step architecture as implemented does not achieve effective raindrop removal.","pith_inferences":["The paper's negative result suggests the bottleneck is the mask, not the inpainter: a better detector (fine-tuned on real raindrop images, or trained with domain adaptation) might make the same Equation 8 reconstruction effective.","The residual-mask branch requires the clean image at test time, so it is not a genuine single-image method; the paper's comparison between residual and learned masks therefore does not directly test the unsupervised scenario it claims to address.","A natural next step would be to evaluate the DDPM inpainter with ground-truth raindrop masks (manually annotated) on the Raindrop test set, isolating inpainter quality from mask quality.","The synthetic-to-real failure is a useful data point for the broader deraining literature, where synthetic training data is common: the refraction model renders physically plausible drops, but the domain gap to real photographs remains large enough to break the downstream pipeline."],"forward_implications":["If the pipeline worked as claimed, raindrop removal would reduce to choosing a mask and letting a clean-image-trained diffusion model fill it, with no need for a GAN discriminator or paired clean images at inference time.","The method would tolerate imperfect masks because the DDPM can reconstruct regions even when the masked areas are larger than the true raindrops.","Training the detector on synthetic raindrops would remove the need for manually annotated raindrop masks, which the Raindrop dataset does not provide.","The same masking trick could be reused for any single-image restoration task where the corruption can be localized, since the DDPM itself is task-agnostic."],"supporting_citations":[{"why":"Supplies the masked-inpainting update of Equation 8, the mechanism that lets a DDPM fill masked regions while preserving known pixels.","marker":"[5]"},{"why":"Provides the pre-trained DDPM model used as the background reconstructor.","marker":"[4]"},{"why":"Supplies the Raindrop dataset, the benchmark and the clean-image training set for the DDPM.","marker":"[2]"},{"why":"Provides the refraction-model implementation and synthetic raindrop generation used to train the raindrop detector.","marker":"[27]"},{"why":"Supplies the Cityscapes images and camera parameters on which the synthetic raindrops are rendered.","marker":"[6]"},{"why":"Defines the DDPM forward and reverse processes that the whole method builds on.","marker":"[15]"}],"fun_headline_variants":["Diffusion raindrop remover fails at mask generalization","Synthetic-trained masks sink diffusion raindrop removal","Raindrop inpainting fails on real masks after synthetic training","DDPM raindrop remover tripped by overfit detector masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that synthetic raindrops, rendered on Cityscapes by simulating how a water droplet refracts the background, are similar enough to real raindrops that a detector trained on them will produce masks usable for inpainting on the Raindrop dataset.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion raindrop remover fails at mask generalization","Synthetic-trained masks sink diffusion raindrop removal","Raindrop inpainting fails on real masks after synthetic training","DDPM raindrop remover tripped by overfit detector masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1375,"prompt_tokens":837,"completion_tokens":538,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":468}},"tokens_in":453,"tokens_out":538,"duration_ms":4614,"temperature":1.0,"reasoning_tokens":468,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:00:55.609537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Manually annotate raindrop masks on a sample of the Raindrop test set and run the same DDPM inpainter with those exact masks, comparing outputs to the ground-truth clean images. If reconstruction quality is still poor, Equation 8's update is the weak link; if quality is good, the detector's synthetic-to-real transfer is the weak link.","supporting_citations":[{"cited_title":"RePaint: Inpainting using Denoising Diffusion Probabilistic Models","cited_arxiv_id":null,"evidence_quote":"Supplies the masked-inpainting update of Equation 8, the mechanism that lets a DDPM fill masked regions while preserving known pixels."},{"cited_title":"Learning From Synthetic Photorealistic Raindrop for Single Image Raindrop Removal","cited_arxiv_id":null,"evidence_quote":"Provides the refraction-model implementation and synthetic raindrop generation used to train the raindrop detector."}],"review_version":1}