{"id":"e736f52a-03fa-4cab-ad86-d0cf54c6e03a","arxiv_id":"2507.16472","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DenseSR uses depth, normal, and DINO priors plus a split smoothing/detail decoder to remove shadows from single images, reporting SOTA on five benchmarks.","lead":"The authors propose DenseSR, a shadow removal network that blends depth, surface normal, and semantic features from pretrained models with a decoder that separates smoothing from boundary sharpening. The method reports top results on several benchmarks, though without released code or statistical significance tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Table 1 does not support its SOTA claim: StableShadowDiffusion reports higher PSNR than DenseSR on ISTD+ and SRD, and DenseSR's gains over OmniSR are small and unreplicated.","rationale":"The reader's weakest_assumption focused on noisy geometric and semantic priors and the fixed 60-degree FOV assumption. Those are legitimate but secondary: the paper's ablations show the priors contribute, and small errors in a prior would not necessarily invalidate an empirical performance claim. The more load-bearing issue is the one the reader gestured at under \"small margins over strong baselines\": the paper's own Table 1 contains a baseline, StableShadowDiffusion, that beats DenseSR in PSNR on two of the four datasets, so the unqualified \"state-of-the-art\" claim is internally inconsistent with the reported evidence. This is not a disagreement with external consensus; it is a mismatch between the evidence and the conclusion drawn from it. The CONDITIONAL verdict remains appropriate, because the architecture, ablations, and evaluation are otherwise presented reasonably and the method may still be a solid contribution once the claims are qualified and the margins are checked for statistical significance. I would add explicit conditions: correct or qualify the SOTA claim, report run-to-run statistics for the close comparisons, and release code or predictions for independent verification. I do not see grounds to reject the paper outright, since the central mechanism and ablations are plausible and the reported numbers, while not uniformly SOTA, are competitive on every benchmark.","tokens_in":18347,"tokens_out":5484,"duration_ms":52245,"concrete_test":"Reconstruct the mask-free comparison from Table 1 and compute per-dataset and average PSNR/SSIM for DenseSR, OmniSR, and StableShadowDiffusion; if StableShadowDiffusion is higher on ISTD+ and SRD, then \"state-of-the-art\" must be reworded to \"competitive\" or explicitly scoped to SSIM and visual quality. Independently, retrain DenseSR and OmniSR with three seeds on ISTD+ and report mean +/- std; if the DenseSR advantage of 0.64 dB is within one standard deviation, the claimed improvement is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of state-of-the-art shadow removal rests on Table 1, but Table 1 itself contradicts an unqualified SOTA reading. In the mask-free rows, StableShadowDiffusion [60] achieves 35.19 PSNR on ISTD+ versus DenseSR's 33.98, and 33.63 PSNR on SRD versus DenseSR's 33.45; on WSRD+, DenseSR leads by only 0.02 dB (26.28 vs 26.26). DenseSR does beat OmniSR on all four datasets, but the margins are small (0.19, 0.64, 0.58, and 0.21 dB) and are reported without error bars or multiple seeds, so they may lie within run-to-run variation. The text itself says \"competitive PSNR and SSIM scores,\" which is accurate, but the abstract and conclusion claim \"state-of-the-art performance on multiple benchmarks,\" which the reported numbers do not establish. A second internal inconsistency appears in Section 5.2: the mask-guided variant is said to be \"surpassed only by ShadowDiffusion + GM and ShadowFormer + GM,\" yet the table shows both of those methods have higher PSNR than DenseSR + GM (32.33 and 32.21 vs 32.14), so DenseSR + GM is actually third, not second. These issues directly affect the main contribution claim and should be corrected or statistically supported before the SOTA claim is accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DenseSR, a mask-free single-image shadow removal network. It augments a U-Net-style Swin Transformer encoder with depth, surface-normal, and DINO-V2 semantic priors via a Scene-Integrated Module (SIM), and inserts a Dense Fusion Block (DFB) in the decoder. The DFB splits restoration into an Adaptive Content Smoothing Module (ACSM) for low-frequency consistency and a Texture-Boundary Recuperation Module (TBRM) for high-frequency detail, fusing the two paths with CARAFE. The model is trained with Charbonnier loss and evaluated on ISTD, ISTD+, SRD, WSRD+, and INS. The paper claims state-of-the-art performance in mask-free shadow removal.","tokens_in":18708,"tokens_out":5817,"duration_ms":58333,"significance":"If the reported numbers are reliable, DenseSR advances a timely direction: combining frozen geometric and semantic priors with adaptive feature fusion for shadow removal. The component ablations in Table 3 give evidence that depth, normals, DINO features, and the DFB each help, and the method is substantially faster than diffusion baselines. However, the manuscript's headline claim is currently overstated: Table 1 shows that StableShadowDiffusion beats DenseSR on ISTD+ and SRD, and the margins over OmniSR are small and unreplicated. Because no error bars, multiple seeds, or released code are provided, the SOTA claim is not yet established. The work is best assessed as a solid architectural contribution whose headline claim needs to be either statistically supported or scaled back.","major_comments":[{"comment":"The unqualified 'state-of-the-art' claim in the abstract and conclusion is not supported by Table 1. In the mask-free rows, StableShadowDiffusion [60] outperforms DenseSR by 1.21 dB on ISTD+ (35.19 vs. 33.98) and by 0.18 dB on SRD (33.63 vs. 33.45); on WSRD+ the lead over StableShadowDiffusion is 0.02 dB, and the gains over OmniSR are 0.19, 0.64, 0.58, and 0.21 dB across the four datasets. Since no error bars or repeated-seed results are given, these margins may lie within run-to-run variation. The authors should either soften the SOTA wording (as the 'competitive' phrasing in Sec. 5.2 already does) or provide a proper statistical comparison.","section":"Sec. 5.2, Table 1"},{"comment":"The text says DenseSR+GM is 'second-best' on ISTD, 'surpassed only by ShadowDiffusion + GM and ShadowFormer + GM,' but Table 1 lists DenseSR+GM at 32.14 PSNR, ShadowFormer+GM at 32.21, and ShadowDiffusion+GM at 32.33; it is therefore third, not second. This internal inconsistency should be corrected.","section":"Sec. 5.2, +GM paragraph"},{"comment":"The normal map is derived from depth using a fixed 60-degree field of view (Eq. 8 in the supplementary). The benchmark datasets contain images from different cameras, and the paper gives no justification for this value or any sensitivity analysis. Because Table 3 shows that removing the normal prior costs 1.32 dB on INS testing and 0.48 dB on WSRD+, an incorrect FOV could materially bias an input that the method depends on. The authors should validate the FOV choice or evaluate robustness to this hyperparameter.","section":"Supplementary Eq. (8), Sec. 4.1.1"},{"comment":"The individual ablations of ACSM and TBRM are small: removing ACSM costs 0.12 dB (INS testing) and 0.17 dB (WSRD+), and removing TBRM costs 0.15 dB and 0.21 dB. Without error bars, multiple seeds, or paired significance tests, these differences are not distinguishable from stochastic variation. The claim that both modules demonstrably contribute needs statistical support.","section":"Table 3"}],"minor_comments":[{"comment":"The abstract and conclusion use 'state-of-the-art' while Sec. 5.2 correctly says 'competitive PSNR and SSIM scores'; unify the wording after addressing the corresponding major comment.","section":"Abstract and Sec. 6"},{"comment":"The abstract says 'Our code can be available' but no release is indicated; please state clearly whether and where the code will be released.","section":"Abstract"},{"comment":"The method of reference [60] is called 'StableShadowRefiner' in one place and 'StableShadowDiffusion' elsewhere; use one consistent name.","section":"Sec. 5.2 and Figure 6"},{"comment":"The caption says the best results are highlighted as 1st, 2nd, and 3rd, but the table as rendered has no visible highlighting; clarify or add the formatting.","section":"Table 1"},{"comment":"The text mentions evaluation on INS alongside ISTD, ISTD+, SRD, and WSRD+, but INS appears only in Table 2; add a cross-reference so the reader can locate the INS numbers.","section":"Sec. 5.2"}],"recommendation":"major_revision","confidential_remarks":"The main weakness is the lack of statistical support for the headline claim. If the authors can provide repeated-seed results or error bars, or revise the SOTA claims to match the reported numbers, the paper could become suitable for publication. The self-citations to DRCT and PromptHSI are not excessive, but the comparison section should more clearly credit OmniSR and StableShadowDiffusion, since their numbers are competitive with the proposed method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental architecture paper for single-image shadow removal. The new Dense Fusion Block is a legitimate idea, and the prior-guided attention is well motivated. But the SOTA claim doesn't survive contact with its own Table 1, so treat the headline with caution.\n\nThe paper does several things well. Splitting decoder features into low-frequency content smoothing (ACSM) and high-frequency detail recuperation (TBRM) before fusing with CARAFE is a clean design. Modulating window self-attention with depth, normal, and DINO similarity maps is a natural extension of OmniSR's prior-guided attention, and the specific combination is new. The evaluation is broad: ISTD, ISTD+, SRD, WSRD+, INS, plus real captured images, with ablations for each prior and each DFB submodule, and inference time compared. That is more than many shadow removal papers provide.\n\nThe soft spots are real. The abstract and conclusion claim state-of-the-art on multiple benchmarks, but Table 1 shows StableShadowDiffusion beats DenseSR on ISTD+ (35.19 vs 33.98) and SRD (33.63 vs 33.45). DenseSR leads on ISTD, WSRD+, and INS, but that is not the same as SOTA everywhere. The margins over OmniSR are 0.19 to 0.64 dB with no error bars or multiple seeds, so the advantage could be run-to-run noise. Also, Section 5.2 says the mask-guided variant is 'second-best' and 'surpassed only by ShadowDiffusion+GM and ShadowFormer+GM,' but the table lists both ahead, making it third; that is an internal contradiction. The individual ablations for ACSM and TBRM are small (about 0.1-0.2 PSNR), which weakens the 'crucial' language. Code is promised on GitHub but not yet available. Minor: normal maps are computed with an assumed 60-degree FOV (supplementary Eq. 8), which may not match the benchmark images.\n\nNone of this kills the paper. The architecture is coherent and the experiments are honest in the specifics; the problem is the overclaiming and lack of statistical support. A good referee will ask for code, error bars, and a toned-down conclusion. With those in place, this would be a solid conference paper.\n\nRecommendation: send it to peer review. The work is substantive, the flaws are calibrational, and the DFB idea is worth examining.","headline":"Solid incremental shadow-removal architecture, but the SOTA claim is contradicted by its own Table 1; worth peer review after fixes.","tokens_in":19256,"tokens_out":3689,"would_cite":true,"duration_ms":35708,"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":"Single-image shadow removal is treated as a dense prediction problem, with the paper claiming that combining deep geometric-semantic scene priors with an adaptive dense-fusion decoder achieves state-of-the-art mask-free shadow removal.","keywords":["single-image shadow removal","dense prediction","scene priors","geometric-semantic guidance","adaptive feature fusion","content smoothing","boundary restoration","mask-free restoration"],"falsifier":"Run the paper's ablations with the normal-estimation field of view changed to, say, 45 and 90 degrees, or with the depth and DINO features deliberately corrupted while keeping the network fixed: if PSNR/SSIM on ISTD+ or INS does not drop beyond the reported margins, the scene-prior pathway is not load-bearing. A simpler check is an independent per-image significance test on the Table 3 differences: if removing depth or normals never changes results outside the evaluation noise floor, the geometric-prior claim collapses.","tokens_in":18146,"feed_emoji":"🌗","tokens_out":8198,"duration_ms":80050,"temperature":0.7,"pith_summary":"Single-image shadow removal is treated as a dense prediction problem, and the paper's central claim is that combining deep geometric-semantic scene priors with an adaptive dense-fusion decoder produces state-of-the-art mask-free shadow removal. The paper argues that shadows non-uniformly degrade multi-scale features, so standard fusion blurs boundaries and leaves intra-shadow inconsistency; DenseSR instead separates content smoothing from detail recuperation before recombination. The model fuses frozen depth, surface-normal, and DINO semantic features through attention modulation to resolve ambiguity between shadows and intrinsically dark surfaces, then restores the image with the Dense Fusion Block. On ISTD, ISTD+, SRD, WSRD+, and INS benchmarks it reports the best or near-best PSNR/SSIM among mask-free methods, and remains competitive with methods that use ground-truth masks. If right, the work establishes that implicit shadow localization via scene priors plus component-aware fusion is a practical alternative to mask-guided and diffusion-based restoration.","feed_headline":"Scene priors plus dense fusion top shadow-removal benchmarks","feed_subtitle":"A mask-free network fuses depth, normals, and DINO features, then restores texture and boundary detail to beat prior methods.","key_machinery":"The Dense Fusion Block (DFB) is the central mechanism: at each decoder upsampling stage it splits the incoming features into low-frequency and high-frequency components, applies adaptive content-aware kernels (ACSM with softmax-normalized low-pass kernels for intra-shadow consistency, and TBRM with inverted low-pass kernels as high-pass filters for texture and boundary recovery), and recombines them via CARAFE reassembly and element-wise addition. The companion mechanism is the Scene-Integrated Module (SIM), which makes window self-attention explicitly depend on scene priors by computing pairwise semantic similarity from DINO feature dot products and geometric consistency from planar distances derived with depth and normals, and multiplying those maps into the QK^T scores before softmax.","core_discovery":"The paper's discovery is that the main bottleneck in modern shadow removal is not just scene understanding but the restoration-stage feature fusion itself: shadows degrade deep features non-uniformly, so the decoder has to process content and detail components separately before combining them. DenseSR's Dense Fusion Block predicts spatially-variant low-pass kernels to smooth the base appearance inside shadows for consistency and spatially-variant high-pass kernels to re-inject textures and sharpen boundaries, then fuses the two streams after content-aware reassembly. Scene-Integrated Modules in the encoder and bottleneck modulate window attention with DINO-derived semantic similarity maps and depth/normal-derived geometric consistency maps, which lets the network implicitly localize shadows without a mask. With this design, the reported numbers on ISTD, ISTD+, SRD, WSRD+, and INS are the best or second-best among mask-free methods, and the ablations show each prior and each DFB sub-module contributes to the final score.","pith_inferences":["Because the priors are frozen and the attention modulation is generic, DenseSR's architecture could transfer to other spatially non-uniform restoration tasks such as low-light enhancement, reflection removal, or dehazing; the paper does not test this.","The normal map depends on a fixed 60-degree field-of-view assumption, so on images with different camera intrinsics the normals will be biased; testing with calibrated intrinsics or omitting normals would reveal whether this prior is robust or dataset-specific.","A per-image breakdown of the reported averages would test whether the gains concentrate on the indoor, indirect-lighting images the paper emphasizes, or spread uniformly across all benchmarks.","Fine-tuning or adaptively weighting the depth and DINO priors instead of freezing them could push performance further, at greater training cost; the paper leaves this unexplored."],"forward_implications":["Mask-free shadow removal can reach or approach mask-guided quality on standard benchmarks, so real applications no longer need an accurate shadow mask at inference time.","Frozen geometric and semantic priors reduce shadow ambiguity enough to substitute for explicit shadow detection in the restoration pipeline.","Splitting restoration into smoothed base content and reinjected high-frequency detail improves both intra-shadow consistency and boundary sharpness relative to standard feature fusion, as the ablations and optimization curves show.","DenseSR runs in about 125 ms on a 640x480 image, faster than diffusion-based shadow removal, while matching or beating those methods on quality.","The decoder's frequency spectra show stronger high-frequency energy at each stage when DFB is used, which the paper connects to sharper boundaries and finer textures."],"supporting_citations":[{"why":"supplies the depth map from which the normal map is derived and which is concatenated into the RGBD input.","marker":"[61]"},{"why":"supplies the multi-scale semantic feature maps used to compute similarity maps in the SIM attention modulation.","marker":"[40]"},{"why":"is the closest prior mask-free method using indirect-lighting awareness and the main baseline to beat on INS and real images.","marker":"[59]"},{"why":"is the Transformer-based baseline whose global-context design DenseSR extends and compares against on all benchmarks.","marker":"[12]"},{"why":"is the diffusion-based baseline that DenseSR compares against for quality and runtime.","marker":"[14]"},{"why":"provides the content-aware reassembly kernels used to fuse the ACSM and TBRM streams in the Dense Fusion Block.","marker":"[53]"},{"why":"provides the surface-normal estimation used to convert depth into the geometric normal prior.","marker":"[15]"},{"why":"provides the evaluation data and code used for WSRD+ comparisons in the NTIRE 2024 challenge.","marker":"[52]"}],"fun_headline_variants":["Mask-free shadow removal with dense fusion tops benchmarks","Dense fusion of scene priors sharpens shadow removal","Shadow removal reframed as dense prediction","No masks needed: dense fusion restores shadowed scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frozen depth and DINO features, plus the assumed 60-degree field of view for normal estimation, give trustworthy geometric and semantic cues for shadow disambiguation; if those priors are noisy or the camera intrinsics differ, the attention modulation can mislead restoration.","fun_headline_variants_meta":{"raw":{"variants":["Mask-free shadow removal with dense fusion tops benchmarks","Dense fusion of scene priors sharpens shadow removal","Shadow removal reframed as dense prediction","No masks needed: dense fusion restores shadowed scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000834,"raw_usage":{"total_tokens":3652,"prompt_tokens":968,"completion_tokens":2684,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2622}},"tokens_in":584,"tokens_out":2684,"duration_ms":20114,"temperature":1.0,"reasoning_tokens":2622,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:08:01.690993+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the paper's ablations with the normal-estimation field of view changed to, say, 45 and 90 degrees, or with the depth and DINO features deliberately corrupted while keeping the network fixed: if PSNR/SSIM on ISTD+ or INS does not drop beyond the reported margins, the scene-prior pathway is not load-bearing. A simpler check is an independent per-image significance test on the Table 3 differences: if removing depth or normals never changes results outside the evaluation noise floor, the geometric-prior claim collapses.","supporting_citations":[{"cited_title":"OmniSR: Shadow Removal under Direct and Indirect Lighting","cited_arxiv_id":"2410.01719","evidence_quote":"is the closest prior mask-free method using indirect-lighting awareness and the main baseline to beat on INS and real images."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the Transformer-based baseline whose global-context design DenseSR extends and compares against on all benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the diffusion-based baseline that DenseSR compares against for quality and runtime."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the surface-normal estimation used to convert depth into the geometric normal prior."}],"review_version":1}