{"id":"ccf608fd-0f03-4a96-9a66-23cef930e4d1","arxiv_id":"2507.01422","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DocShaDiffusion removes shadows from document images by running a mask-guided denoising diffusion in latent space, and contributes a synthetic color-shadow dataset and state-of-the-art benchmark numbers.","lead":"This paper presents DocShaDiffusion, a latent-space diffusion model that estimates a shadow soft-mask and uses it to concentrate noise addition and denoising on shadowed regions of document images. The authors also construct a synthetic color-shadow document dataset, SDCSRD, and report higher PSNR/SSIM than several baselines on Adobe, FSDSRD, RDD, and their own dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. 5-6 are internally inverted as written: with x=shadowed and µ=shadow-free, reverse sampling from a noisy shadow image cannot produce a shadow-free output, and the mask term (1-I_m)dw suppresses noise exactly where shadows are darkest.","rationale":"The paper's central claim is state-of-the-art document shadow removal, with color shadows as a highlighted novelty and SMGDM as the core mechanism. For that claim to hold, the forward/reverse SDE pair must actually map a noisy shadowed input to a shadow-free output. As written, Eq.5 has the drift and mask weighting pointing the wrong way: the forward process moves toward the clean image while the noise is suppressed in shadows. This is an internal inconsistency, not a disagreement with consensus, and it is directly checkable from the equations. No code or dataset is released, so the experimental tables cannot show which SDE was actually run; the reported Adobe and FSDSRD numbers are competitive, and the ablation is plausible, which is why I credit the possibility that the implementation follows Refusion and the paper text is misstated. If the implementation matches the text, the central construction fails; if it matches Refusion, the paper needs a corrected formulation plus the Eq.1 color fix and a fairer evaluation baseline. Those requirements are exactly what the reader's CONDITIONAL verdict already asks for, so my concern strengthens the condition but does not move the verdict category. The reader's weakest assumption (scalar Ck in Eq.1) is a separate, valid issue that undercuts the color-shadow motivation; I treat it as secondary to the SDE direction problem.","tokens_in":13581,"tokens_out":13451,"duration_ms":161296,"concrete_test":"Re-derive the time-reversal of Eq.5 under the printed semantics (x=shadowed, µ=shadow-free) and simulate one reverse trajectory starting from X_T = I_s + (1-I_m)ε with I_m=1 inside shadows. If the output at t=0 is not close to I_sf, the SMGDM direction is confirmed wrong. As a control, repeat the derivation with the standard Refusion assignment (µ=shadowed observation); if that control yields I_sf, then Eqs.5-6 in the paper are inverted and must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is the SMGDM SDE, not the color scalar in Eq.1. Section IV.B gives Eq.5 as the forward IR-SDE, with x signifying the shadowed image and µ the target shadow-free image. Under those semantics the drift θt(µ-x) pulls the state toward the clean image, so X_T from the forward process is essentially clean-plus-noise; the reverse Eq.6 is the time-reversal of that map and would send clean-plus-noise back toward shadowed images. The pipeline in Fig.4 and Section IV.A instead forms X_T by adding mask-modulated noise to the shadow input and then uses reverse denoising to obtain a shadow-free X_0, which requires the opposite assignment (µ = shadowed observation), as in Refusion. Independently, Eq.5 uses -σt(1-I_m)dw; because I_m=1 marks the darkest shadow (Section III.A), this makes the noise coefficient vanish in shadow regions and largest in non-shadow regions, contradicting the stated aim of adding noise specially to shadow regions. If Eqs.5-6 are taken literally, the core module cannot perform the claimed restoration. Since code and data are not released, the tables cannot resolve which convention was actually implemented. The reader's Eq.1 scalar-Ck concern is real but affects the color-shadow claim; the SDE issue affects the entire method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DocShaDiffusion, a latent-space diffusion model for document image shadow removal. The method consists of a shadow soft-mask generation module (SSGM) that produces a soft mask from the luminance ordering of the input, and a shadow mask-aware guided diffusion module (SMGDM) built on the IR-SDE framework. The authors also introduce a synthetic dataset, SDCSRD, generated by a linear blending formula, and a shadow-robust perceptual feature loss based on VGG-19 features. Experiments are reported on Adobe, FSDSRD, SDCSRD, and RDD, with the central claim that DocShaDiffusion outperforms the state of the art, especially on color shadows. The paper states that code and dataset will be released.","tokens_in":13842,"tokens_out":4428,"duration_ms":53514,"significance":"If correct, the paper would make two contributions: a diffusion-based architecture tailored to document shadow removal, and a synthetic color-shadow dataset that could serve as training data for other methods. The ablation study shows a large positive effect of the proposed components (PSNR increases from 23.20 to 32.54 when all components are included), and the reported numbers on Adobe and FSDSRD are competitive with existing methods. However, the significance assessment is weakened by the absence of code/data, the lack of statistical validation, and, more importantly, by internal inconsistencies in the mathematical formulation of the core diffusion module and in the color-shadow dataset construction. The claimed SOTA result on the authors' own synthetic dataset is also not established under a fair comparison protocol.","major_comments":[{"comment":"The forward SDE in Eq. (5) is internally inconsistent with the stated restoration goal. The text defines x as the shadowed image and mu as the shadow-free target. Under that assignment, the drift theta_t(mu - x) pushes the state toward the clean image, so X_T at the end of the forward process is approximately clean-plus-noise. The reverse SDE in Eq. (6) is the time reversal of that map and would send clean-plus-noise back toward shadowed images, not the other way around. The pipeline described in Section IV.A and Fig. 4 instead constructs X_T by adding mask-modulated noise to the shadowed input and then denoises to obtain a shadow-free image, which requires the opposite convention, namely mu equal to the shadowed observation, as in the cited Refusion work. Independently, the noise coefficient in Eq. (5), -sigma_t(1 - I_m)dw, vanishes where I_m = 1, i.e., in the darkest shadow regions, whereas the text in Section IV.B and the contribution bullet in the Introduction state that noise is added 'specially to shadow regions.' The two statements cannot both be true. This issue is load-bearing: as written, the core SMGDM module does not implement the described restoration process, and without released code or data the experimental tables cannot resolve which convention was actually used.","section":"Section IV.B, Eqs. (5)-(6)"},{"comment":"The dataset construction formula in Eq. (1) cannot generate color shadows as claimed. The text states that both a and C_k are scalar values, and I_sf, I_m are images in RGB color space. A scalar C_k multiplied by a scalar-valued mask (after the stated histogram equalization) scales all three color channels identically, producing grayscale shadow intensity shifts but no chromatic shift. Thus SDCSRD, as specified, is a gray-shadow dataset, not a color-shadow dataset, directly contradicting the paper's central motivation that existing methods 'ignore color shadows' and that SDCSRD 'simulates the distribution of realistic color shadows.' To produce colored shadows, C_k would need to be a three-dimensional vector (or otherwise channel-dependent). The authors should either correct Eq. (1) to include per-channel color factors or revise the claim that the dataset addresses color shadows.","section":"Section III.A, Eq. (1)"},{"comment":"The claim of outperforming the state of the art is not supported by the evaluation on SDCSRD. In Table II, DocShaDiffusion is trained on SDCSRD, while all baseline methods are evaluated with their original weights or provided results, as stated in Section V.B ('The results of other models are provided by MS-GAN, DCShadow-Net, CBENet'). The very large margins on SDCSRD (e.g., PSNR 41.66 vs. CBENet's 24.58) therefore largely reflect a training-set advantage, not model superiority. The averaging in Table IV mixes heterogeneous conditions: CBENet is trained on RDD while the proposed method is trained on SDCSRD. A valid SOTA claim requires a fair protocol, e.g., retraining all baselines on SDCSRD or testing all methods under the same cross-dataset conditions, or at minimum clearly labeling the tables as 'trained on SDCSRD' vs. 'not trained on SDCSRD.' Without this, the central 'superiority over state-of-the-art' assertion is not established.","section":"Section V.B, Tables II and IV"},{"comment":"The paper reports no error bars, repeated runs, or significance tests for any of the quantitative claims, including the statements in Section V.C that NAFBlock, SMGDM, and L_fea 'play important role' and 'prove beneficial.' Additionally, the feature-loss weights w_i in Eq. (8) are never specified, so the loss in Eq. (9) is incompletely defined. Given the single-run nature of the tables and the unspecified weights, the 'significantly outperforms' language in Section V.B and the conclusion is not backed by statistical evidence.","section":"Section V.A and V.C"}],"minor_comments":[{"comment":"There are numerous typographical errors and inconsistencies: 'datset' for dataset, 'tansformer' for transformer, 'P SN R' spacing artifacts in Tables I-III, and 'the k-th slice (k = 0, 1, ... , 4)' in Section IV.C, where k should range over 0..4 consistently.","section":"Throughout"},{"comment":"The notation is unclear: the loss uses ∥e_t − ϵ∥, where e_t is described as the predicted noise map, but e_t is not defined in the text; presumably it should be ϵθ, the neural network's noise prediction. Please clarify.","section":"Section IV.C, Eq. (7)"},{"comment":"The dataset 'HS' is referenced in the caption of Fig. 3 and in Section III.B, but it is never introduced or defined in the paper. Please identify which real dataset is meant.","section":"Fig. 3 and Section III.B"},{"comment":"The venue/year for 'Our Method' is listed as 'CVPR/2024,' which is not appropriate for a submitted manuscript and does not match the paper's arXiv date of July 2025. This should be corrected to avoid confusing the reader.","section":"Tables I and II"},{"comment":"The sentence 'The experimental results indicate that executing the denoising process in pixel space would necessitate excessive computational resources' reports a result that is never shown; either provide the supporting experiment or remove the unsupported claim.","section":"Section IV.B"}],"recommendation":"major_revision","confidential_remarks":"The core architectural idea is plausible and the ablation indicates that the proposed modules help, but the SDE formulation, as written, is internally inconsistent with the pipeline description, and the color-shadow dataset construction has the same kind of inconsistency. The paper also evaluates SOTA claims on its own synthetic dataset without retraining baselines and without any statistical validation. These issues are fixable in principle, but they affect the central claims and therefore require major revision with, ideally, code/data release to resolve the ambiguity about what was actually implemented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's load-bearing diffusion module (SMGDM) does not work as written. The forward SDE in Eq.5 has drift θ_t(µ−x), which pulls the state toward the shadow-free image µ, so the forward process ends near the clean image; the reverse in Eq.6 correspondingly sends clean-plus-noise back toward shadows. That is the wrong direction for the claimed restoration, where X_T is the shadowed image plus noise and X_0 is shadow-free. The noise term −σ_t(1−I_m)dw also suppresses noise where the mask is darkest, exactly opposite to the stated design intent. This is not a presentation slip; it is the core of the method. The paper also states that both a and C_k in Eq.1 are scalars, but a scalar C_k cannot introduce chromatic shifts; either the text is wrong about the color-shadow claim or the equation is. Both issues need fixing before the experiments can be interpreted.\n\nThat said, there is useful material here. The application of latent-space diffusion to document shadow removal is a plausible direction, and the proposed SDCSRD dataset, promised to be public, could be a genuinely useful resource for the subfield. The idea of conditioning noise injection with a shadow soft-mask is sensible, and the reported numbers on Adobe and FSDSRD are competitive. The ablations show the components add value.\n\nThe surrounding problems are minor by comparison but real: feature weights w_i are unspecified, there are no error bars or significance tests, and the baselines on SDCSRD were not retrained, making the SOTA claim on that dataset self-serving. Code and data are not released, so we cannot tell whether the SDE issue is a typo or a real implementation detail.\n\nNo reviewable paper should take the SOTA claim at face value. But the problem is real, the dataset is potentially valuable, and the issues are addressable. Send it to review only if the authors are willing to fix the SDE, release the artifacts, and provide honest comparisons. As it stands, I would not cite it.","headline":"The forward SDE contradicts the method's own pipeline, so this needs major revision despite a potentially useful dataset.","tokens_in":14406,"tokens_out":4840,"would_cite":false,"duration_ms":52863,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that latent-space diffusion with shadow-mask guidance removes document shadows, including colored ones, better than existing methods.","keywords":["document image shadow removal","latent diffusion model","shadow soft-mask","color shadows","synthetic dataset","mask-guided diffusion","shadow-robust perceptual loss"],"falsifier":"Take any SDCSRD shadow image for which the ground truth is known and measure the color of a shadowed patch before and after removing the shadow. If the synthetic shadow only scales RGB values by a scalar factor, the hue angle in the shadowed patch equals the hue in the ground-truth patch; a dataset that claims color shadows should show a shift in hue or saturation. A second check compares SSGM masks with manually annotated shadow boundaries on images with colored shadows on colored paper; low IoU on such images would mean the mask guidance fails exactly in the regime the paper targets.","tokens_in":13355,"feed_emoji":"📄","tokens_out":4873,"duration_ms":57734,"temperature":0.7,"pith_summary":"This paper aims to make document image shadow removal work for colored shadows, not just the constant-color shadows that earlier methods target. It proposes DocShaDiffusion, a diffusion model that operates in latent space and uses a predicted shadow soft-mask to concentrate noise and denoising on shadowed regions, together with a shadow-robust perceptual loss. The authors also build SDCSRD, a 17,624-image synthetic dataset designed to mimic real color-shadow distributions, and report that their method outperforms prior state-of-the-art methods on the Adobe, FSDSRD, and SDCSRD benchmarks, with particularly large gains on their own dataset. The practical goal is a usable document shadow remover plus a training set that other methods can adopt.","feed_headline":"Latent diffusion removes document shadows, including colored ones","feed_subtitle":"A mask-guided model tops three benchmarks and ships a 17,624-image color-shadow training set.","key_machinery":"The load-bearing object is the shadow soft-mask $I^m$ produced by SSGM (Algorithm 1), a normalized map with values in $[0,1]$, set to 0 outside shadows and near 1 in the darkest shadow interior, obtained by sorting filtered grayscale pixels and thresholding with the mean of the darkest fraction. Inserted into the forward SDE $dx = \\theta_t(\\mu - x)\\,dt - \\sigma_t(1 - I^m)\\,dw$, the mask is intended to concentrate random perturbation and denoising effort on shadow regions while leaving non-shadow areas largely untouched. The same mask guides a latent-space denoising network built from NAFBlocks with time embeddings, and the loss adds a VGG-19-based perceptual term that is designed to be less sensitive to shadow regions.","core_discovery":"On its own terms, the paper establishes that document shadow removal can be cast as a mask-guided latent diffusion problem. The shadow soft-mask generation module (SSGM) derives a soft mask from a single shadow image through grayscale conversion, dilation, median filtering, and luminance sorting, while the shadow mask-aware guided diffusion module (SMGDM) uses that mask to control the noise term in an IR-SDE forward process and to supervise the reverse denoising. Training couples the diffusion objective with a shadow-robust perceptual feature loss computed from VGG-19 features. According to the reported experiments, the resulting system outperforms existing heuristic and neural document shadow removal methods on Adobe, FSDSRD, and SDCSRD, with the largest margins on the synthetic color-shadow dataset it introduces.","pith_inferences":["If Eq. 1 is read literally with scalar $C_k$, the synthetic shadows are only darker versions of the same hue, so the dataset would not actually contain chromatic color shadows; the color-shadow story requires $C_k$ to be a color vector or channel-wise blending, which the text does not state.","The very large performance gap on SDCSRD compared with the more modest gains on RDD suggests that much of the reported advantage comes from matching the synthetic degradation model; a decisive test would be evaluation on real captured colored shadows on colored paper.","The same SSGM-plus-SMGDM recipe could plausibly transfer to other localized document degradations such as stains, watermarks, or demoiréing, since the mask mechanism is agnostic to the specific degradation type.","A direct reading of Eq. 5 gives noise coefficient $1-I^m$, which suppresses noise where the mask is 1, the opposite of the text's claim that noise is amplified in shadow regions; reconciling this sign would clarify what the mask is actually doing."],"forward_implications":["Document shadow removal can be formulated as a mask-guided latent diffusion problem, so shadow regions receive targeted noise and denoising while background detail is preserved.","The synthetic color-shadow dataset SDCSRD, generated with random shadow weight and shadow color parameters, gives other methods a large-scale training resource that existing synthetic datasets such as FSDSRD do not provide.","Because only shadow regions need to be denoised, the approach is expected to train faster and to avoid unnecessary changes to non-shadowed text and background.","The reported numbers on Adobe, FSDSRD, and SDCSRD indicate the model transfers to real document images without retraining on those sets, although on the RDD dataset it does not surpass CBENet, which was trained on RDD.","The ablation results suggest that the NAFBlock encoder-decoder, the mask-aware diffusion guidance, and the perceptual loss each contribute to the final performance, with the architecture change giving the largest single jump in PSNR and LPIPS."],"supporting_citations":[{"why":"Supplies the Adobe dataset, the real benchmark used for testing and for the main comparison tables.","marker":"[16]"},{"why":"BEDSR-Net is presented as the first neural document shadow removal method, the baseline this work extends and compares against.","marker":"[29]"},{"why":"CBENet is the strongest neural baseline and the source of the real-world RDD dataset used for evaluation.","marker":"[30]"},{"why":"FSDSRD provides the existing synthetic dataset, shadow mask templates, and a comparison benchmark for SDCSRD.","marker":"[31]"},{"why":"ShadowDiffusion demonstrates diffusion-based shadow removal on natural scenes, which this paper adapts to document images.","marker":"[37]"},{"why":"Inoue et al. supply the shadow mask templates used to construct the SDCSRD dataset.","marker":"[49]"},{"why":"IR-SDE/Refusion is the foundational diffusion framework whose forward and reverse SDEs SMGDM builds on.","marker":"[54]"},{"why":"DCShadow-Net is a neural natural-image shadow removal method whose reported results serve as a comparison baseline.","marker":"[59]"}],"fun_headline_variants":["Mask-guided latent diffusion clears document shadows","Latent diffusion erases document shadows, colored ones too","DocShaDiffusion: latent diffusion for shadow-free documents","New dataset and diffusion model for document shadow removal","Shadow mask guides diffusion to remove document shadows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that synthetic shadows generated by Eq. 1 with a scalar $C_k$ adequately represent real colored shadows, and that the luminance-sorting heuristic in SSGM can locate those shadows. If $C_k$ is truly scalar, the synthetic images are darker versions of the same hue, so the dataset contains no genuine color shadows and the paper's core motivation is not exercised.","fun_headline_variants_meta":{"raw":{"variants":["Mask-guided latent diffusion clears document shadows","Latent diffusion erases document shadows, colored ones too","DocShaDiffusion: latent diffusion for shadow-free documents","New dataset and diffusion model for document shadow removal","Shadow mask guides diffusion to remove document shadows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1322,"prompt_tokens":904,"completion_tokens":418,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":346}},"tokens_in":520,"tokens_out":418,"duration_ms":5370,"temperature":1.0,"reasoning_tokens":346,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:51:53.203893+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any SDCSRD shadow image for which the ground truth is known and measure the color of a shadowed patch before and after removing the shadow. If the synthetic shadow only scales RGB values by a scalar factor, the hue angle in the shadowed patch equals the hue in the ground-truth patch; a dataset that claims color shadows should show a shift in hue or saturation. A second check compares SSGM masks with manually annotated shadow boundaries on images with colored shadows on colored paper; low IoU on such images would mean the mask guidance fails exactly in the regime the paper targets.","supporting_citations":[{"cited_title":"Removing shadows from images of documents,","cited_arxiv_id":null,"evidence_quote":"Supplies the Adobe dataset, the real benchmark used for testing and for the main comparison tables."},{"cited_title":"Bedsr-net: A deep shadow removal network from a single document image,","cited_arxiv_id":null,"evidence_quote":"BEDSR-Net is presented as the first neural document shadow removal method, the baseline this work extends and compares against."},{"cited_title":"Document image shadow removal guided by color-aware background,","cited_arxiv_id":null,"evidence_quote":"CBENet is the strongest neural baseline and the source of the real-world RDD dataset used for evaluation."},{"cited_title":"Document shadow removal with foreground detection learning from fully synthetic images,","cited_arxiv_id":null,"evidence_quote":"FSDSRD provides the existing synthetic dataset, shadow mask templates, and a comparison benchmark for SDCSRD."},{"cited_title":"Shadowdiffusion: When degradation prior meets diffusion model for shadow removal,","cited_arxiv_id":null,"evidence_quote":"ShadowDiffusion demonstrates diffusion-based shadow removal on natural scenes, which this paper adapts to document images."},{"cited_title":"Learning from synthetic shadows for shadow detection and removal,","cited_arxiv_id":null,"evidence_quote":"Inoue et al. supply the shadow mask templates used to construct the SDCSRD dataset."},{"cited_title":"Refusion: Enabling large-size realistic image restoration with latent- space diffusion models,","cited_arxiv_id":null,"evidence_quote":"IR-SDE/Refusion is the foundational diffusion framework whose forward and reverse SDEs SMGDM builds on."},{"cited_title":"Dc-shadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network,","cited_arxiv_id":null,"evidence_quote":"DCShadow-Net is a neural natural-image shadow removal method whose reported results serve as a comparison baseline."}],"review_version":1}