{"id":"ec84fb9d-fe15-479e-b3c5-79fe4454a36c","arxiv_id":"2504.14092","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"ReHiT combines a Retinex decomposition with illumination-guided histogram attention to achieve competitive mask-free shadow removal at roughly one-seventeenth the parameters of the previous state of the art.","lead":"This paper presents ReHiT, a lightweight neural network for removing shadows from photos without needing a pre-computed shadow mask. It ranks 7th in the NTIRE 2025 shadow removal challenge while using only a fraction of the parameters of the top entries.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experimental section is internally contradictory, so the central efficiency-quality claim for ReHiT cannot be verified from the manuscript as submitted.","rationale":"I agree with the REJECT outcome but not exactly with the reader's stated weakest assumption. The Retinex estimator is indeed undisclosed and Eq. 3's inverse-map assumption is strong, but the manuscript has a more immediate problem: the experimental section is a collage of the authors' older ShadowRefiner text and the new ReHiT results. This makes Table 1's 'Ours' row and the NTIRE 2025 ranking impossible to interpret unambiguously. The central claim is not that the architecture is mathematically wrong; it is that the reported efficiency-quality trade-off is real. That claim depends entirely on experiments, and those experiments are currently presented in a self-contradictory way. The concrete check I propose—running the released code to reproduce the WSRD+ full-model row and locating the Retinex estimator—would settle whether the ReHiT numbers are real. If they reproduce, a clean rewrite of Section 4 could make the paper assessable; until then, the REJECT verdict stands.","tokens_in":15081,"tokens_out":6812,"duration_ms":63611,"concrete_test":"Run the released repository (github.com/dongw22/oath) on WSRD+ with the stated training protocol (crop 384x384, batch size 4, Adam with LR from 1e-4 to 6.25e-6, L1 + MS-SSIM + structure loss + ECMamba constraints), reproduce the Table 4 'Full Model' row (PSNR 26.15, SSIM 0.826, LPIPS 0.0860), and locate the Retinex estimator E used in Eq. 3; if the reproduced metrics differ by more than 0.3 dB PSNR or the estimator is absent from the code, the central efficiency-quality claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—competitive shadow-removal quality at 17.5M parameters and 66.4 GFLOPs—rests entirely on the ReHiT-specific rows in Tables 1–4 and the NTIRE 2025 ranking table. In the submitted manuscript, Section 4 contains two overlapping experimental narratives: a large verbatim block from the authors' earlier ShadowRefiner paper ('As documented in Tab. 1, our ShadowRefiner demonstrates superior performances...', a table captioned 'NTIRE 2024 Perceptual Track', and figure captions that say 'our ShadowRefiner') and a separate ReHiT narrative with conflicting table numbering (the first Table 2 is a params/FLOPs comparison, a later Table 2 is the NTIRE 2024 ranking; the first Table 3 is a ShadowRefiner ablation headed 'ShadowRefiner (Ours)', a later Table 3 is the NTIRE 2025 ranking). No statement specifies which training protocol produced Table 1's 'Ours' row, how the NTIRE 2025 entry 'Oath (ours)' relates to the 17.5M/66.4G model, or which Retinex estimator E in Eq. 3 was actually used. Because the evidence for the central claim is self-contradictory, the reported efficiency-quality trade-off is currently unverifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ReHiT, a mask-free shadow removal network that combines a dual-branch Retinex decomposition with an illumination-guided hybrid CNN-Transformer architecture. The method is evaluated on ISTD, ISTD+, WSRD+, and the NTIRE 2025 Shadow Removal Challenge, where the authors report competitive quality at 17.5M parameters and 66.4 GFLOPs, achieving 7th place in the fidelity track.","tokens_in":15247,"tokens_out":5207,"duration_ms":38800,"significance":"If substantiated, the reported efficiency-quality trade-off would be practically useful, and the authors do ship code. The design, however, reuses substantial components from the authors' previous works (ECMamba [13] and ShadowRefiner [11]) and from Sun et al. [42], so the novelty is incremental. More importantly, the experimental evidence as presented is internally contradictory and missing key technical details, so the central claim cannot currently be assessed.","major_comments":[{"comment":"The experimental section is internally inconsistent: it contains a large verbatim block from the authors' earlier ShadowRefiner paper, including text claiming 'our ShadowRefiner demonstrates superior performances,' a table captioned 'NTIRE 2024 Perceptual Track,' and an ablation table headed 'ShadowRefiner (Ours),' interleaved with the ReHiT narrative. As a result, Table 2 appears twice (once as model complexity, once as NTIRE 2024 ranking) and Table 3 appears twice (once as a ShadowRefiner ablation, once as the NTIRE 2025 ranking). No statement identifies which rows correspond to ReHiT, so the 'Ours (ReHiT)' numbers in Table 1 cannot be unambiguously attributed to the described model.","section":"Section 4 (after Table 1)"},{"comment":"The entire dual-branch construction depends on a Retinex estimator E that produces (bar-R, bar-L, F_i) satisfying bar-L ⊙ L_GT ≈ 1 and bar-R ⊙ R_GT ≈ 1. This estimator is never described in the manuscript; the only reference is [13], the authors' own ECMamba paper. The reader cannot determine the estimator's architecture, training data, or whether it is shared between the two branches, yet it is the basis for the reflectance/illumination separation and for the illumination guidance in IG-HSA. Without this component, the method is not reproducible and the ablation 'w/o dual-branch pipeline' in Table 4 is uninterpretable.","section":"Section 3.1, Eq. (3)"},{"comment":"All reported numbers are from a single training run with no error bars, multiple seeds, or significance tests. The claimed advantages over ShadowRefiner are 0.06-0.13 dB PSNR, while ReHiT is slightly worse on SSIM and LPIPS in two of the three datasets, and the ablations in Table 4 show performance drops of only 0.18-0.41 dB PSNR. These deltas are within typical run-to-run variation for image restoration networks; as presented, the results do not support the claims that ReHiT is competitive or that each component contributes meaningfully.","section":"Section 4.2, Table 1 and Section 4.4, Table 4"},{"comment":"The NTIRE 2025 entry is listed as 'Oath (ours)' in Table 3, while the paper's method is named ReHiT; the relationship between these names is never explained. Furthermore, the paper claims 'fastest inference speeds among top-ranked entries,' but no runtime or latency measurements appear anywhere in the manuscript; Table 3 reports only parameter counts. The efficiency claim is therefore unsupported by the presented data.","section":"Section 4.3, Table 3"},{"comment":"It is never stated which training data and protocol produced the Table 1 results versus the NTIRE 2025 entry. Section 4.1 describes training on a single RTX 3090Ti GPU with specific augmentations and losses but does not mention the NTIRE dataset, while Section 4.3 states the model was trained solely on the NTIRE 2025 challenge set. The connection between the model evaluated in Table 1 and the model in Table 3 is missing, so the central efficiency-quality claim cannot be verified from the manuscript.","section":"Sections 4.1 and 4.3"}],"minor_comments":[{"comment":"The abstract contains a duplicated phrase: 'multi-scale semantic fusion, multi-scale semantic fusion.'","section":"Abstract"},{"comment":"The captions for these figures refer to 'our ShadowRefiner' instead of 'our ReHiT,' which is inconsistent with the paper's method name.","section":"Figure 3 and Figure 4 captions"},{"comment":"These table captions are clearly leftovers from the authors' previous ShadowRefiner paper and should be removed or corrected.","section":"Table 2 (NTIRE 2024) and Table 3 (ShadowRefiner ablation)"},{"comment":"The paper mentions 'additional constraints [13]' as part of the loss but never specifies what these constraints are; they should be defined or precisely referenced.","section":"Section 4.1, loss function"},{"comment":"The code repository is linked as github.com/dongw22/oath while the model is named ReHiT; the relationship between these names should be clarified.","section":"Code link"}],"recommendation":"major_revision","confidential_remarks":"The manuscript in its current form appears to be a hastily assembled mixture of the authors' prior ShadowRefiner paper and a new ReHiT method. The core technical components are inherited from the authors' own ECMamba [13] and from Sun et al. [42], so the novelty is limited. If the authors can produce a clean version with a fully described Retinex estimator, consistent table numbering, and multi-run statistics, the paper could be considered again, but as submitted it does not meet the journal's standards for verifiability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take: the contribution here is real, but the manuscript is not. The new piece is the illumination-guided histogram self-attention (IG-HSA), which modulates histogram attention with Retinex illumination features. That's a sensible idea, and the reported efficiency numbers (17.5M params, 66.4 GFLOPs, versus 293.8M/274.7G for ShadowRefiner) would be a meaningful result for edge deployment if they hold up. But hold up is the problem: Section 4 is a mess. Large chunks are copied verbatim from the authors' own ShadowRefiner paper, including a table that says 'ShadowRefiner (Ours)', figure captions that say 'our ShadowRefiner', and an NTIRE 2024 ranking table. Table numbering is duplicated (two Table 2s, two Table 3s), the NTIRE 2025 entry is called 'Oath' while the model is ReHiT, and the abstract repeats 'multi-scale semantic fusion.' This isn't a style problem — it means the reader cannot tell which numbers actually belong to ReHiT.\n\nWhat is genuinely good: the method section is clearly written and the architecture is a reasonable combination. The authors credit Sun et al. for histogram attention and Yu et al. for DRDB/SAM; the dual-branch Retinex formulation is taken from their own ECMamba. The NTIRE 2025 challenge placement (7th with the smallest parameter count among the top entries) is a concrete, externally checkable result.\n\nWhere it falls down: the Retinex estimator E in Eq. 3 is never described here, only cited to ECMamba, and the whole dual-branch framework depends on it. The ablations in Table 4 are 0.1–0.4 dB, which is within run-to-run noise; no error bars, no multiple seeds, no model selection details. The PSNR gains over ShadowRefiner are marginal (0.06–0.13 dB), so the efficiency gain is the real selling point, not the accuracy.\n\nWho this is for: someone working on lightweight image restoration might get a useful architectural idea, but they should not rely on the reported numbers until the experiments are redone and reported cleanly. As submitted, I'd desk reject and invite a clean resubmission with Section 4 rebuilt and the Retinex estimator specified. The underlying idea deserves another look.","headline":"A plausible efficiency story and one genuinely new attention mechanism, buried under an experimental section copy-pasted from the authors' own ShadowRefiner paper — unverifiable as submitted.","tokens_in":15891,"tokens_out":5498,"would_cite":false,"duration_ms":46023,"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":"A compact Retinex-guided transformer removes shadows without masks and matches far larger models.","keywords":["shadow removal","mask-free","Retinex theory","histogram self-attention","CNN-Transformer hybrid","image restoration","illumination estimation","efficient deep learning"],"falsifier":"Take the trained model and set the estimator to a constant identity ($\\bar L=\\bar R=1$) so that $R'=L'=I^{Sh}$; if the PSNR on WSRD+ stays near the reported 26.15 dB rather than falling to the 25.86 dB of the 'w/o dual-branch' ablation, the Retinex guidance is not the source of the gain. The paper's own ablation table supplies the exact number this variant must beat to confirm the mechanism.","tokens_in":14740,"feed_emoji":"🌗","tokens_out":13005,"duration_ms":103441,"temperature":0.7,"pith_summary":"ReHiT is a mask-free shadow removal network whose central claim is that shadow removal is best organized as two separate restoration problems: one for reflectance (intrinsic object color) and one for illumination (lighting). Guided by Retinex theory, the network estimates approximate inverse maps that split the shadowed image into these two components, then restores each with a hybrid CNN-Transformer encoder-decoder whose histogram-based attention is modulated by the estimated illumination. The paper argues this design reaches the quality of much larger mask-free methods, reporting PSNR/SSIM/LPIPS comparable to ShadowRefiner on ISTD, ISTD+, and WSRD+ while using 17.5M parameters and 66.4 GFLOPs, roughly 17x fewer parameters and 4x fewer FLOPs than that baseline. Because it needs no shadow mask and was trained only on the NTIRE 2025 Shadow Removal Challenge training set, the result speaks to real-world shadow removal under imperfect annotations and limited compute.","feed_headline":"No masks, no huge model: shadow removal at 17.5M parameters","feed_subtitle":"A Retinex-guided dual-branch transformer matches a 294M-parameter model using 4x fewer FLOPs.","key_machinery":"The load-bearing component is the Illumination-Guided Histogram Transformer Block (IG-HTB), specifically its illumination-guided histogram self-attention (IG-HSA). Standard self-attention uses a fixed attention range; IG-HSA instead sorts spatial elements into histogram bins with equal pixel counts and computes attention within and across bins, giving the model a flexible attention span for shadows of varying size and softness. The estimated illumination map is injected to modulate the attention, so the block can concentrate on regions of abrupt lighting change. Around this block, each IG-HCT encoder/decoder level also contains a Dilated Residual Dense Block (DRDB) for residual dense features and a Semantic-aligned Scale-Aware Module (SAM) for multi-scale fusion. The dual-branch Retinex decomposition of Eq. 3 is what turns shadow removal into two sub-problems; IG-HSA is what makes each sub-problem tractable with a small network.","core_discovery":"The paper's central claim is that a Retinex-guided dual-branch pipeline can make a compact CNN-Transformer competitive with large pre-trained mask-free shadow removers. The shadowed image is modeled as $I^{Sh}=(R_{GT}+\\hat R)\\odot(L_{GT}+\\hat L)$; assuming a Retinex estimator yields approximate inverses $\\bar L$ and $\\bar R$ satisfying $\\bar L\\odot L_{GT}=1$ and $\\bar R\\odot R_{GT}=1$, multiplying the input by these inverses produces near-clean reflectance $R'=I^{Sh}\\odot\\bar L$ and illumination $L'=I^{Sh}\\odot\\bar R$. Two instances of the Illumination-Guided Hybrid CNN-Transformer (IG-HCT) predict the residual degradations in $R'$ and $L'$, and the refined outputs are multiplied to obtain $I_{out}=R_{out}\\odot L_{out}$. On benchmarks, ReHiT reports top-tier mask-free numbers: 28.81 dB PSNR on ISTD, 31.16 dB on ISTD+, and 26.15 dB on WSRD+, and its WSRD+ ablation attributes part of the gain to the dual-branch design and to the illumination-guided histogram attention.","pith_inferences":["A testable extension is to sweep the number of histogram bins in IG-HSA: if flexible attention span is the real cause of the gain, there should be an optimal bin count that tracks the spatial irregularity of shadows, with performance degrading on both sides.","The same reflectance-illumination split could transfer to low-light enhancement, dehazing, and exposure correction, since those tasks also fit the Retinex decomposition; the paper's citation of ECMamba already points in that direction.","The unspecified Retinex estimator is the main hidden variable: re-running the pipeline with the estimator frozen versus trained jointly, or replaced by a simple off-the-shelf decomposition, would reveal how much of the reported performance belongs to the estimator and how much to the IG-HCT modules.","Because the same IG-HCT design serves as both $M_R$ and $M_L$, the illumination branch's output could be tested as a zero-shot shadow locator: regions where $L'$ deviates most from their surroundings might localize shadows without any mask supervision."],"forward_implications":["If the results hold, mask-free shadow removal no longer needs a separate shadow-detection stage or manual masks, removing a major barrier to deployment on arbitrary real-world images.","The efficiency numbers (17.5M parameters, 66.4 GFLOPs) imply shadow removal can run on a single commodity GPU or edge device while retaining near-top quality, instead of requiring the multi-hundred-GFLOP budgets of large ConvNeXt-based competitors.","On WSRD+, where precise masks are unavailable and only estimated masks exist, ReHiT's 26.15 dB PSNR exceeds the mask-based ShadowFormer's 25.44 dB, suggesting mask-free methods can beat mask-dependent ones precisely where masks are hardest to obtain.","The WSRD+ ablation shows the dual-branch Retinex pipeline contributes roughly 0.29 dB PSNR and the IG-HTB roughly 0.41 dB; if these gains replicate, each architectural choice is load-bearing rather than cosmetic.","Training only on the NTIRE 2025 challenge set shows competitive shadow removal can be obtained without extra training data, large-scale pre-training, or multi-stage training pipelines."],"supporting_citations":[{"why":"Supplies the Retinex estimator that produces the approximate inverse maps $\\bar L$, $\\bar R$, and the dual-branch perturbation model in Eq. 1-3.","marker":"[13]"},{"why":"Defines Retinex theory, the reflectance-illumination decomposition on which the whole two-branch pipeline rests.","marker":"[28]"},{"why":"ShadowRefiner, the main mask-free baseline whose quality and parameter/FLOP counts ReHiT is compared against.","marker":"[11]"},{"why":"Introduces the histogram self-attention mechanism that IG-HSA adapts by injecting illumination guidance.","marker":"[42]"},{"why":"Supplies the Dilated Residual Dense Block and Semantic-aligned Scale-Aware Module used in every IG-HCT block.","marker":"[59]"},{"why":"NTIRE 2025 shadow removal challenge report that records ReHiT's 7th-place fidelity rank and the efficiency comparison among top entries.","marker":"[46]"},{"why":"Provides the ISTD dataset with shadow/shadow-free pairs used to evaluate the method.","marker":"[49]"},{"why":"Provides the ISTD+ dataset and the shadow image decomposition idea that motivates separating reflectance from illumination.","marker":"[29]"},{"why":"Provides the WSRD+ dataset and validation split used for the main comparisons and all ablations.","marker":"[43]"},{"why":"Supplies estimated shadow masks used to run mask-based methods on WSRD+, the setting where ReHiT reports surpassing them.","marker":"[8]"}],"fun_headline_variants":["Mask-free shadow removal with a 17.5M-parameter Retinex transformer","No masks, 17.5M params: Retinex-guided transformer rivals 294M models","Retinex-guided histogram transformer: mask-free shadows at 17.5M","Shadow removal without masks: compact Retinex transformer matches big models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole framework depends on an unshown preprocessing step that separates each image into reflectance and illumination layers; the paper borrows this estimator from earlier work and never evaluates it on its own, so an inaccurate estimator would collapse the two branches into one and remove the claimed benefit of the lighting/color separation.","fun_headline_variants_meta":{"raw":{"variants":["Mask-free shadow removal with a 17.5M-parameter Retinex transformer","No masks, 17.5M params: Retinex-guided transformer rivals 294M models","Retinex-guided histogram transformer: mask-free shadows at 17.5M","Shadow removal without masks: compact Retinex transformer matches big models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000901,"raw_usage":{"total_tokens":3920,"prompt_tokens":1028,"completion_tokens":2892,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":2803}},"tokens_in":644,"tokens_out":2892,"duration_ms":19704,"temperature":1.0,"reasoning_tokens":2803,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:56:30.792204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained model and set the estimator to a constant identity ($\\bar L=\\bar R=1$) so that $R'=L'=I^{Sh}$; if the PSNR on WSRD+ stays near the reported 26.15 dB rather than falling to the 25.86 dB of the 'w/o dual-branch' ablation, the Retinex guidance is not the source of the gain. The paper's own ablation table supplies the exact number this variant must beat to confirm the mechanism.","supporting_citations":[{"cited_title":"Ecmamba: Consolidating selective state space model with retinex guidance for efficient multiple exposure correc- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the Retinex estimator that produces the approximate inverse maps $\\bar L$, $\\bar R$, and the dual-branch perturbation model in Eq. 1-3."},{"cited_title":"Shadowrefiner: Towards mask-free shadow removal via fast fourier transformer","cited_arxiv_id":null,"evidence_quote":"ShadowRefiner, the main mask-free baseline whose quality and parameter/FLOP counts ReHiT is compared against."},{"cited_title":"Restoring images in adverse weather condi- tions via histogram transformer","cited_arxiv_id":null,"evidence_quote":"Introduces the histogram self-attention mechanism that IG-HSA adapts by injecting illumination guidance."},{"cited_title":"Towards efficient and scale-robust ultra- high-definition image demoir´eing","cited_arxiv_id":null,"evidence_quote":"Supplies the Dilated Residual Dense Block and Semantic-aligned Scale-Aware Module used in every IG-HCT block."},{"cited_title":"NTIRE 2025 image shadow removal challenge report","cited_arxiv_id":null,"evidence_quote":"NTIRE 2025 shadow removal challenge report that records ReHiT's 7th-place fidelity rank and the efficiency comparison among top entries."},{"cited_title":"Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal","cited_arxiv_id":null,"evidence_quote":"Provides the ISTD dataset with shadow/shadow-free pairs used to evaluate the method."},{"cited_title":"Shadow removal via shadow image decomposition","cited_arxiv_id":null,"evidence_quote":"Provides the ISTD+ dataset and the shadow image decomposition idea that motivates separating reflectance from illumination."},{"cited_title":"Wsrd: A novel benchmark for high resolution image shadow removal","cited_arxiv_id":null,"evidence_quote":"Provides the WSRD+ dataset and validation split used for the main comparisons and all ablations."},{"cited_title":"Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan","cited_arxiv_id":null,"evidence_quote":"Supplies estimated shadow masks used to run mask-based methods on WSRD+, the setting where ReHiT reports surpassing them."}],"review_version":1}