{"id":"ce50fc18-7d5c-4945-96be-1bfe875d5fb6","arxiv_id":"2412.02545","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A two-stage luminance-then-color neural network that achieves state-of-the-art shadow removal on ISTD+ and SRD datasets.","lead":"ShadowHack removes shadows from photos by first repairing the image's brightness and texture channel, then separately restoring colors with a second network. It reports small but consistent gains over previous shadow-removal models on the ISTD+ and SRD benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inference-time mask pathway is unspecified: user-specified masks change outputs (Fig. 11) even though Sec. 4 and Fig. 4 never feed a mask into LRNet or CRNet, so the reported SRD gains could stem from hidden mask refinement rather than the proposed decoupling.","rationale":"The paper's central claim is that ShadowHack achieves state-of-the-art shadow removal. For that claim to hold, the benchmark comparisons must be apples-to-apples and the method must be specified well enough to reproduce. The mask pathway is the least secure part of that chain: the architecture in Sec. 4 makes no use of masks, yet Sec. 5.4 demonstrates both mask-conditioned outputs and a mask-refinement network. If the refined masks are used for ShadowHack's SRD evaluation while baselines use raw DHAN masks, the reported RMSE improvement of 0.43 and shadow-region PSNR lead of 0.66 over Homoformer could be an artifact of better mask inputs, not of the proposed luminance-color decomposition. This is a concrete, testable concern that does not require questioning the authors' integrity. The reader's weakest assumption about luminance as a proxy for material identity is plausible but less load-bearing: Table 5 shows the cross-attention mechanism helps empirically, and even without it the model remains competitive, so a failure of that assumption would degrade some color accuracy but would not directly undermine the headline benchmark comparisons. The mask issue, by contrast, bears directly on whether the reported numbers are valid comparisons. A conditional verdict remains appropriate because the concern can be resolved by code release and an explicit statement of the mask path in the paper; it does not by itself prove the results are wrong.","tokens_in":13874,"tokens_out":9268,"duration_ms":99901,"concrete_test":"Obtain the released code/checkpoints and rerun the SRD benchmark under a controlled mask protocol: (i) ShadowHack with mask refinement, (ii) ShadowHack with raw DHAN masks, and (iii) top baselines (e.g., Homoformer, RASM) using the same refined masks. If ShadowHack's RMSE advantage over the best baseline shrinks to below about 0.1 or reverses when all methods share identical mask inputs, the SOTA claim is not established. Also inspect the forward pass to confirm whether a mask tensor is concatenated into LRNet or CRNet; if it is absent, Fig. 11's dependence on user-specified masks is unexplained.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim rests on an underspecified mask pathway. In Sec. 4 and Fig. 4, LRNet and CRNet are described as operating only on luminance It and color Ic; no shadow mask is listed as an input. Yet Sec. 5.4 (Fig. 11) shows that user-specified masks change the output, so masks must condition the network at inference. The same section introduces a mask-refine U-Net 'to get rid of the performance degradation brought by inaccurate masks,' but the paper never states whether this refinement is applied in the Table 1 SRD/ISTD+ evaluations. Since SRD uses predicted DHAN masks ('We adopt the predicted masks provided by DHAN'), a hidden mask-refinement stage would give ShadowHack cleaner mask inputs than the published baselines, which used raw masks. The SRD gains (RMSE 2.90 vs 3.33; shadow-region PSNR 39.47 vs 38.81) could then be inflated by the mask preprocessing rather than by the luminance-color divide-and-conquer. This makes the claimed new SOTA unverifiable without code or an explicit statement of the inference-time mask path.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ShadowHack proposes a two-stage divide-and-conquer framework for single-image shadow removal. An invertible decoupling maps each RGB input into a luminance channel It and color channels Ic; LRNet restores brightness and texture in the luminance space using a Rectified Outreach Attention module with dilated overlapping windows and a differential attention-rectification mechanism, while CRNet regenerates color through cross-attention in which restored luminance features serve as query and key and multi-scale color features from a ConvNext-v2 color encoder serve as value, with training-time checkpoint ensembling to tolerate imperfect LRNet outputs. Experiments on ISTD+ and SRD report state-of-the-art numbers (PSNR 36.31 and RMSE 2.48 on ISTD+; PSNR 35.94 and RMSE 2.90 on SRD), along with ablations of the decoupling, the attention module, and the color network, and qualitative generalization results on UIUC and UCF. The paper states that code will be released.","tokens_in":14093,"tokens_out":14999,"duration_ms":135221,"significance":"The luminance-color decomposition is well motivated by the image-formation analysis in Sec. 3 (Eqs. 1-7), and the ablation study is a genuine strength: each design choice, including decoupling versus end-to-end RGB processing, outreach dilation, rectification, feature combinations in the attention module, cross-attention versus concatenation in CRNet, and checkpoint ensembling, is backed by a quantitative table (Tables 3-5) with internally consistent conclusions. The SRD result (ALL RMSE 2.90 versus 3.33 for the prior best) is a comparatively large improvement, the model is lightweight (23.3M parameters), and the re-evaluation of RASM from released output images is a fair comparison practice. The significance of the headline SOTA claim is, however, currently limited: the inference-time role of the shadow mask is never specified, so the SRD gains cannot yet be attributed to the proposed architecture rather than to unseen mask preprocessing, and the ISTD+ margin (0.15 dB) is reported without any error analysis. If the mask pathway is clarified and the gains hold up, this will be a solid and clean contribution.","major_comments":[{"comment":"The methodology section never specifies how the shadow mask enters the network. Eqs. (8)-(9) and Fig. 4 define the pipeline as D^{-1}(C(R(It), Ic)) with no mask term, yet Sec. 5.4 (Fig. 11) demonstrates that user-specified masks change the output at inference, and the same section introduces a mask-refine U-Net to 'get rid of the performance degradation brought by inaccurate masks.' The paper must state (a) the exact point at which the mask conditions LRNet or CRNet, (b) whether the mask-refine network was applied in the Table 1 evaluations, especially for SRD where masks come from DHAN predictions, and (c) the mask-refine network's training data, supervision, and loss. Without this information, the reported SRD improvements (ALL RMSE 2.90 vs 3.33; shadow-region PSNR 39.47 vs 38.81 over Homoformer) could stem from mask refinement rather than from the luminance-color divide-and-conquer, and the headline SOTA claim is not verifiable from the manuscript.","section":"Sec. 4 / Fig. 4; Sec. 5.4"},{"comment":"The ISTD+ state-of-the-art claim rests on a 0.15 dB PSNR margin (36.31 vs 36.16 for RASM) and a 0.05 RMSE margin (2.48 vs 2.53), reported without error bars, repeated-run statistics, or any significance analysis. The claim also depends on region aggregation: ShadowHack's shadow-region metrics on ISTD+ (PSNR 40.56, RMSE 4.46) are slightly worse than RASM's (40.73, 4.41), so the advantage is concentrated in non-shadow regions. Given the unspecified mask pathway raised above, I ask for per-image error distributions (or equivalent statistical support) and a discussion of the shadow-region deficit before the paper can fairly claim to surpass the prior art.","section":"Sec. 5.2 / Table 1"},{"comment":"The color regeneration mechanism rests on the assumption that luminance similarity is a reliable proxy for material identity: CRNet's cross-attention computes query and key from restored luminance features while color features serve as values, and the text states the aim is to 'reference regions with similar/identical material properties.' If a shadow region and a lit region have similar brightness but different true albedo, the attention will copy incorrect colors. The paper neither analyzes this failure mode nor tests the assumption beyond the aggregate Table 5 ablation; the checkpoint-ensemble paragraph in the same section implicitly concedes that LRNet's test-time outputs are imperfect, which heightens the concern. I ask for an analysis of this failure mode (for example, attention visualizations or a targeted study on same-luminance/different-albedo pairs), or an explicit statement of the limitation.","section":"Sec. 4.2 / Table 5"}],"minor_comments":[{"comment":"The ablation in Table 3 reports an ALL RMSE of 2.46 for 'Ours' on ISTD+, while Table 1 reports 2.48 for ShadowHack under the same protocol; please reconcile these numbers.","section":"Table 3 vs Table 1"},{"comment":"The RGB-input baseline in Table 3 attains PSNR 36.16 on ISTD+, exactly RASM's ALL PSNR in Table 1; please clarify whether this baseline is an independently trained model or a re-implementation or reuse of RASM.","section":"Table 3"},{"comment":"AdamW has no 'momentum' hyperparameter; the values (0.9, 0.999) are the beta_1 and beta_2 parameters of the optimizer, and the text should be corrected accordingly.","section":"Sec. 5.1"},{"comment":"The checkpoint-ensemble description is ambiguous: the sentence 'the luminance network often yields better performance on the training set' does not state the reference point of the comparison, and the number and selection scheme of the early-stage checkpoints are not given; please specify how many checkpoints are used and how they are sampled.","section":"Sec. 4.2"},{"comment":"The abstract states that 'Our code will be made publicly available at here,' with a placeholder instead of a URL; please complete the link or remove the placeholder.","section":"Abstract"},{"comment":"The phrase 'To kindle a poorly lit image' appears to be a word-choice error; 'kindle' should likely be 'brighten' or 'enhance.'","section":"Sec. 3"},{"comment":"The notation lambda_j^i with i, j in {1, 2} is hard to parse; please give the four scalar parameters distinct, readable names.","section":"Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The reader's verdict and the stress-test note align with my own reading: the mask-pathway gap between Sec. 4 and Sec. 5.4 is the load-bearing issue, and the revision must resolve it before the SOTA claims can be verified. I would encourage the editor to request the promised code or a precise textual specification of the inference-time mask handling, since the current manuscript cannot rule out that the SRD gains come from mask refinement. A small additional verification request: Table 3's RGB-input baseline PSNR (36.16) coincides exactly with RASM's ALL PSNR in Table 1; confirming that these are independent evaluations would remove any ambiguity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid shadow-removal paper with a real architectural idea and unusually complete ablations, but the headline SOTA numbers are less certain than the text implies because the inference-time mask pathway is never specified. If you are going to rely on these numbers, ask for code or a precise description of the pipeline.\n\nWhat's new: the luminance-color decoupling in YCbCr, with LRNet handling brightness/texture and CRNet doing color regeneration via cross-attention from luminance features to color features. The rectified outreach attention is a reasonable tweak on window attention, and the differential rectification is well motivated. The ablations in Tables 3-5 are genuinely informative; they isolate the contributions of decoupling, outreach, rectification, and cross-attention cleanly.\n\nThe SRD results are the more convincing part of the empirical claim: RMSE 2.90 vs 3.33 for the next best, a solid margin. On ISTD+, the all-image PSNR gain is 0.15 dB over RASM, which is within normal run-to-run variation for this kind of model. The absence of error bars or repeated runs is a real limitation.\n\nNow the soft spot, and it is a real one. The main architecture in Sec. 4 feeds only luminance and color into LRNet and CRNet; no shadow mask is shown as an input. Yet Sec. 5.4 shows that user-specified masks change the output and introduces a mask-refine U-Net. The paper never states whether that mask-refine U-Net is applied during the Table 1 evaluations. Since SRD masks are predicted by DHAN and could be noisy, a hidden mask-refinement stage would give ShadowHack cleaner mask inputs than the baselines, and part of the reported gain could come from mask preprocessing rather than the proposed decoupling. This is a verifiability problem, not necessarily a flaw: the authors may be using the raw masks in Table 1, but they need to say so clearly or release code.\n\nThe luminance-as-material-proxy assumption in CRNet is a legitimate theoretical concern, but the checkpoint ensemble shows the authors are aware of the fragility, and the ablations support the design. Minor point: the physics analysis in Sec. 3 is heuristic and not load-bearing.\n\nThis paper deserves a serious referee. The method is novel enough, the ablations are strong, and the main ambiguity is addressable with a clear statement or code. A reviewer should focus on the exact inference pipeline for the benchmark numbers and the robustness of the SOTA claim.","headline":"A solid two-stage shadow removal with strong ablations, but the headline SOTA claim rests on an unspecified mask pathway that needs clarification or code.","tokens_in":14639,"tokens_out":2384,"would_cite":false,"duration_ms":22302,"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":"ShadowHack claims that shadow removal is best done by first restoring luminance and then regenerating color, and it reports state-of-the-art PSNR and RMSE on the ISTD+ and SRD benchmarks.","keywords":["shadow removal","luminance restoration","color regeneration","rectified outreach attention","cross-attention","YCbCr decomposition","image restoration","transformer"],"falsifier":"Construct or find an image in which a shadowed object and a lit object have the same measured luminance but very different true color, such as a dark-red object in shadow and a bright-blue object in light. If ShadowHack turns the shadowed object blue, the luminance-as-material assumption is violated; if it keeps the object red, the network is using information beyond luminance similarity. A quantitative version would measure CRNet's output color error on synthetic shadow patches whose matched lit counterparts have identical luminance but different chroma.","tokens_in":13632,"feed_emoji":"🌗","tokens_out":7697,"duration_ms":68848,"temperature":0.7,"pith_summary":"This paper tries to establish that shadow removal can be solved more effectively by decomposing the problem into two sequential subproblems: luminance recovery and color regeneration. The authors argue that shadows degrade images in three entangled ways—reduced brightness, texture deterioration, and color distortion—and that a divide-and-conquer pipeline handles each more cleanly than a single end-to-end model. On standard benchmarks, the proposed two-stage network, ShadowHack, reports the best all-image PSNR and RMSE, including an RMSE improvement of 0.43 on SRD over the prior best. A sympathetic reader would care because the result suggests that explicit decomposition into luminance and chrominance is a viable design principle for shadow-removal networks, not just a pre-processing trick.","feed_headline":"Splitting light from color wins shadow removal benchmarks","feed_subtitle":"A two-stage network restores luminance first, then copies colors from lit regions, cutting RMSE by 0.43 on SRD.","key_machinery":"The load-bearing machinery is the YCbCr decoupling plus two specialized networks. LRNet is a U-shaped transformer whose deeper stages replace local blocks with a Rectified Outreach Attention (ROA) module: queries come from regular window partitions, keys and values come from dilated outreach windows, and the final attention is a rectified difference between an attention map computed on luminance-plus-color features and one computed on color-only features, which suppresses noise and negative correlation in shadow areas. CRNet is a U-shaped network augmented with a pretrained color encoder; cross-attention modules use restored luminance features as queries and keys and multi-scale color features as values to inject colors into the skip connections. A checkpoint-ensembling trick used only during training makes CRNet robust to imperfect luminance outputs without adding inference cost.","core_discovery":"The central discovery is that separating shadow removal into luminance restoration and color regeneration—rather than estimating a single shadow-free RGB image—lets each sub-network work from cleaner inputs and produces more accurate results. The pipeline converts RGB to YCbCr, restores the Y channel with LRNet, then regenerates the Cb and Cr channels in CRNet by using restored luminance features as queries and keys and color features as values, so colors in shadow regions are copied from non-shadow patches with similar luminance. The paper reports an all-image PSNR of 36.31 and RMSE of 2.48 on ISTD+, and 35.94 and 2.90 on SRD, surpassing the previous best results on most metrics. Ablation experiments show that the decomposition itself, the outreach windows, the differential rectification, and the cross-attention color injection each contribute to the gains.","pith_inferences":["The paper does not test what happens when luminance similarity fails to indicate material identity; an image with two objects of equal brightness but different true colors in shadow and lit regions would probe this boundary.","The same YCbCr divide-and-conquer design may transfer to low-light enhancement or shadow detection, where the luminance channel carries most of the illumination problem; applying LRNet alone to underexposed images would be a direct test.","The checkpoint-ensembling trick suggests CRNet is sensitive to training-time luminance errors; feeding deliberately corrupted luminance maps at test time would quantify that sensitivity.","If the public code reproduces the reported numbers, the strongest next comparison is against diffusion-based shadow removal at matched parameter counts, since ShadowHack claims a large RMSE improvement with a smaller, non-diffusion model."],"forward_implications":["If the reported numbers hold, end-to-end RGB shadow-removal models are not the only strong option; decomposing into luminance and color is a competitive design that can be combined with future backbones.","The SRD RMSE gain of 0.43 suggests color regeneration is where decoupling pays off most, so methods that ignore color bias will keep lagging on color-error metrics.","The Rectified Outreach Attention module could be dropped into other window-based restoration transformers whenever a region mask is available.","Because CRNet copies colors from non-shadow patches by luminance similarity, its accuracy should degrade gracefully as shadow masks become coarser, provided the mask-refine network keeps working.","The two-stage pipeline makes shadow removal user-controllable: users can supply or edit the shadow mask, and only the luminance correction and the corresponding color-copy region change."],"supporting_citations":[{"why":"Prior state-of-the-art baseline whose released output images are re-evaluated; ShadowHack improves on its metrics on both ISTD+ and SRD.","marker":"[29]"},{"why":"Strong transformer baseline with published results; it is the second-best method on SRD that ShadowHack outperforms by 0.43 RMSE.","marker":"[48]"},{"why":"Transformer shadow-removal model introducing shadow/non-shadow interaction; the proposed attention mechanism extends this idea.","marker":"[13]"},{"why":"Diffusion-based shadow-removal baseline that ShadowHack outperforms with a lighter non-diffusion architecture.","marker":"[14]"},{"why":"Provides the ISTD+ dataset with shadow, shadow-free, and mask triplets used for training and testing.","marker":"[25]"},{"why":"Provides the SRD dataset with shadow images and masks used for the SRD benchmark.","marker":"[39]"},{"why":"Supplies the intrinsic-image and Retinex decomposition model that motivates separating luminance from color.","marker":"[1]"},{"why":"Supplies the differential-attention mechanism that the Rectified Outreach Attention module adapts for shadow regions.","marker":"[51]"},{"why":"Provides the pretrained ConvNext-v2 ato model used as the multi-scale color encoder inside CRNet.","marker":"[34]"}],"fun_headline_variants":["ShadowHack splits luminance and color for top shadow removal","Two-step shadow removal: restore light, then recolor wins","ShadowHack: separate luminance and color to beat benchmarks","Luminance-color split tops shadow removal with 36.31 PSNR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The color regeneration network assumes that two patches with similar luminance are made of similar material, so it copies color from non-shadow patches whose brightness matches the shadow region; if that assumption fails, the wrong color is injected.","fun_headline_variants_meta":{"raw":{"variants":["ShadowHack splits luminance and color for top shadow removal","Two-step shadow removal: restore light, then recolor wins","ShadowHack: separate luminance and color to beat benchmarks","Luminance-color split tops shadow removal with 36.31 PSNR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1160,"prompt_tokens":863,"completion_tokens":297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":224}},"tokens_in":479,"tokens_out":297,"duration_ms":3300,"temperature":1.0,"reasoning_tokens":224,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:20:24.411001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or find an image in which a shadowed object and a lit object have the same measured luminance but very different true color, such as a dark-red object in shadow and a bright-blue object in light. If ShadowHack turns the shadowed object blue, the luminance-as-material assumption is violated; if it keeps the object red, the network is using information beyond luminance similarity. A quantitative version would measure CRNet's output color error on synthetic shadow patches whose matched lit counterparts have identical luminance but different chroma.","supporting_citations":[{"cited_title":"Regional atten- tion for shadow removal","cited_arxiv_id":null,"evidence_quote":"Prior state-of-the-art baseline whose released output images are re-evaluated; ShadowHack improves on its metrics on both ISTD+ and SRD."},{"cited_title":"Homoformer: Homogenized trans- former for image shadow removal","cited_arxiv_id":null,"evidence_quote":"Strong transformer baseline with published results; it is the second-best method on SRD that ShadowHack outperforms by 0.43 RMSE."},{"cited_title":"Shadowformer: Global context helps shadow removal","cited_arxiv_id":null,"evidence_quote":"Transformer shadow-removal model introducing shadow/non-shadow interaction; the proposed attention mechanism extends this idea."},{"cited_title":"Shadowdiffusion: When degradation prior meets diffusion model for shadow removal","cited_arxiv_id":null,"evidence_quote":"Diffusion-based shadow-removal baseline that ShadowHack outperforms with a lighter non-diffusion architecture."},{"cited_title":"Shadow removal via shadow image decomposition","cited_arxiv_id":null,"evidence_quote":"Provides the ISTD+ dataset with shadow, shadow-free, and mask triplets used for training and testing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SRD dataset with shadow images and masks used for the SRD benchmark."},{"cited_title":"Recovering intrinsic scene characteristics","cited_arxiv_id":null,"evidence_quote":"Supplies the intrinsic-image and Retinex decomposition model that motivates separating luminance from color."},{"cited_title":"A convnet for the 2020s","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained ConvNext-v2 ato model used as the multi-scale color encoder inside CRNet."}],"review_version":1}