{"id":"8f70a624-571a-4584-a9bc-4571223b62a5","arxiv_id":"2501.15043","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A prompt-aware network, PACSRNet, removes the shadow of a user-specified subject using a dot, line, or subject mask, and the authors introduce a new dataset for this task.","lead":"Researchers built a network that removes a chosen object's shadow from a photo after the user points at it with a dot, line, or mask, instead of requiring the shadow region to be marked. They also created a new synthetic dataset of 11,900 images to train and test this prompt-based approach.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prompt-aware claim is validated only on auto-generated dot/line prompts derived from the same subject masks that define the ground truth; without human-prompt or real-image tests, user-selectable shadow removal may not transfer.","rationale":"Reader's verdict is sound. I considered two other candidates: (i) the abstract says 'eliminates the need for shadow annotations,' although training uses shadow masks; this is rhetorical overstatement but does not invalidate the method. (ii) PCSRD is synthesized from DESOBAv2, so real-image behavior is unproven; this is real but secondary because the paper's claimed contribution is a new task/benchmark, and a synthetic benchmark can be a legitimate first step. The prompt-distribution issue is more load-bearing because the entire novelty is prompt-aware controllability: if real user prompts differ from the automatically generated ones, the headline capability simply has not been demonstrated, regardless of the image domain. The proposed test—a human study with off-distribution controls—would settle this directly. Because this is an addressable evaluation gap, not a contradiction in the method, the appropriate verdict remains CONDITIONAL; I therefore set verdict_should_be to UNCHANGED relative to the reader's conditional.","tokens_in":12817,"tokens_out":7363,"duration_ms":74365,"concrete_test":"Run a human-prompt evaluation on a PCSRD test subset: recruit 10-20 users, show each shadow image, ask them to click a dot or draw a line on the object whose shadow should be removed, and feed the resulting prompts to the released PACSRNet. Compare PSNR/SSIM and shadow-mask IoU with the auto-generated prompt results (Tables 1 and 3). Add a control set of hand-designed off-distribution prompts (dot placed on the shadow rather than the subject, dot at the boundary, line spanning two subjects) on the same images. Pre-specify a threshold: if the human/off-distribution prompts lower all-region PSNR by more than 1 dB or shadow-mask IoU by more than 0.05 relative to auto prompts, the user-prompt claim is not established. If code/dataset are not released for this test, that itself blocks verification of the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the dot and line prompts in PCSRD are a faithful proxy for real user prompts. Section 4.1 states that dot/line prompts are automatically generated from the subject mask 'by a dynamic programming strategy'; the same subject mask is one of the prompt types and is tied to the ground-truth shadow mask. Thus at test time the model only sees prompts sampled from the exact distribution defined by the training labels. There is no human study, no user-drawn strokes, and no out-of-distribution prompt (e.g., a click on the shadow, a click outside the subject, or a line crossing multiple subjects). The ISTD+ results in Table 1 use only the shadow-removal module with the ground-truth shadow mask (PACSRNet†), so they do not test prompt-aware behavior on real images. Additionally, PCSRD is built from DESOBAv2, so the full prompt-aware pipeline is evaluated only on synthesized shadow images. The claim that a user can specify an object's shadow with a dot/line is therefore supported only within a narrow, generator-controlled distribution; failure to transfer to human prompts or real images would not contradict the architecture's internal consistency, but it would undermine the paper's central practical claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new task, prompt-aware controllable shadow removal, in which a user specifies a subject via a dot, a line, or a subject mask and the model removes only the shadow of that subject without requiring a shadow mask at inference. The proposed PACSRNet consists of a prompt-aware module that predicts the shadow mask for the specified subject and a shadow removal module that uses this mask and prompt-guided features to restore shadowed regions. A new dataset, PCSRD, is constructed from DESOBAv2 by automatically generating dot and line prompts from subject masks. Experiments are reported on PCSRD and ISTD+, comparing against five shadow removal baselines, along with ablations of the main components.","tokens_in":13008,"tokens_out":4743,"duration_ms":43940,"significance":"If the central claim is validated, the paradigm could make shadow removal more practical and interactive by replacing precise shadow-mask annotations with simple user prompts. The dataset is the first of its kind for this task and may facilitate follow-up work. The architecture itself is plausible: the spatial-frequency interaction and dense-sparse attention blocks are sensible, and the ablation study in Table 2 supports their individual contributions. However, the current evidence for the key prompt-awareness claim is weak. The prompts used in evaluation are automatically generated from the same subject masks that define the ground truth, the full prompt-based model has not been tested on real images, and the full model's quantitative performance is below several strong mask-based baselines. These issues currently limit the support for the paper's central practical claim.","major_comments":[{"comment":"The evaluation of prompt-awareness is circular. Dot and line prompts in PCSRD are generated from the subject mask by a dynamic programming strategy (Section 4.1), and the same subject mask is the ground truth for the mask prediction metrics reported in Table 3. At test time the model only sees prompts sampled from the exact distribution that the training labels define, so the reported IoU/BCE values do not indicate how the model will behave with real user prompts such as off-center clicks, partial line strokes, or clicks placed on the shadow itself. Please add a human-prompt experiment (even a small user study) or, at minimum, a held-out set of manually annotated dot/line prompts, and report robustness to varying prompt placements.","section":"Section 4.1, Table 3"},{"comment":"The quantitative comparison does not support the stated 'superiority' of the full prompt-based model. On PCSRD, PACSRNet with dot, line, and subject mask prompts achieves all-region PSNR of 40.96, 41.06, and 41.59 dB, respectively, which is lower than ShadowFormer (42.30 dB), HomoFormer (42.22 dB), and Inpaint4Shadow (41.97 dB); on shadow regions the gap is larger (e.g., 43.38 dB for the dot prompt versus 45.32 dB for ShadowFormer). The claim in Section 4.2 that the method is 'comparable to ShadowDiffusion' selects the weakest mask-based baseline and ignores the stronger ones. Please provide a more balanced interpretation of the results, add error bars or statistical significance tests, and explicitly discuss the trade-off between prompt-based control and removal quality.","section":"Table 1"},{"comment":"The full prompt-aware pipeline is never evaluated on real images. The ISTD+ experiments in Table 1 use only PACSRNet†, i.e., the shadow removal module with the ground-truth shadow mask as input, not the complete system with dot/line/subject-mask prompts. The PCSRD dataset itself is built from DESOBAv2 composite shadow images, so the full system is evaluated only on synthetic data. Since the paper's central practical claim is that a user can specify a subject's shadow without any shadow mask on real scenes, the absence of a real-image test of the full pipeline leaves the main claim unsupported. Please test the full model on a real shadow removal benchmark with manually provided prompts, or clearly state this as a limitation and temper the corresponding claims.","section":"Section 4.1, Testing Dataset"},{"comment":"The description of the dot/line prompt generation is not reproducible. The paper states only that dot and line prompts are generated 'by a dynamic programming strategy' from the subject mask, without specifying how the dot location is chosen, how the line is drawn, whether multiple prompts per subject are sampled, and what the resulting distribution looks like. Because the prompt distribution is the defining feature of the new dataset and directly shapes the task, please provide the full algorithm and a visualization of the generated prompts relative to the subject masks.","section":"Section 4.1, Dataset Generation"}],"minor_comments":[{"comment":"There are several typos: 'Cntrollable' in the abstract, 'the our PACSRNet' in the caption of Figure 2, 'Shdow' in the caption of Figure 3, and 'PSCSRNet' in the caption of Table 3.","section":"Abstract and various"},{"comment":"The paper uses both 'dot' and 'point' for the same prompt type (e.g., 'w/ Dot' in tables versus 'point' in Section 4.2). Please use one consistent term.","section":"Throughout"},{"comment":"The input representation of the prompt is not specified. Please state how a dot or line is encoded (e.g., a sparse binary map, a coordinate list) and how it is concatenated or fused with the image features in the prompt-aware module.","section":"Section 3.2"},{"comment":"No error bars or statistical significance measures are reported; all metrics appear to be from a single run. Given that the ablations report margins as small as 0.11 dB (Table 2), please report means and standard deviations over multiple seeds or runs.","section":"Tables 1-3"},{"comment":"Instance shadow detection methods (Wang et al., 2020, 2021, 2022) are mentioned in related work but are not compared or discussed as baselines for the prompt-aware mask prediction. Since these methods also predict a specific object's shadow, a comparison (or an explicit explanation of why they are not applicable) would strengthen the evaluation.","section":"Section 2.2 and Section 4.2"},{"comment":"The loss weight λ is set to 3 with no sensitivity analysis. Please report the effect of varying λ on the validation performance, as this is the only free parameter in the loss.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a strong practical claim ('user prompts' enabling controllable shadow removal) that is supported only by auto-generated prompts derived from the same masks used as ground truth. This is a central issue that goes beyond presentation. The manuscript needs either a genuine user-prompt evaluation or a substantial reframing of the claims. I would also note that the dataset, while new, is built on the synthetic DESOBAv2 and its novelty is incremental; the release of the dataset is a positive step, but the paper does not mention code release, which may be worth requesting for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes a genuinely new task—point at an object and remove its shadow without supplying a shadow mask—but the evaluation of that core claim is partly circular, because the dot and line prompts are generated from the same subject masks that define the ground truth. The task itself is worth knowing about; the evidence for the 'user prompt' claim is not solid yet.\n\nWhat's new and good: first prompt-based controllable shadow removal, a network (PACSRNet) that integrates a prompt-aware mask predictor with a shadow removal module, and a new dataset (PCSRD) with dot, line, and subject-mask prompts. The shadow-region results on PCSRD are strong (45.56 PSNR for the mask-based variant vs 44-45 for baselines), the ablations are conventional, and the paper is clearly written.\n\nThe soft spots: the prompts in PCSRD are auto-generated from the subject mask, and the same subject mask defines the ground-truth shadow mask used for evaluation in Table 3. So the model is tested on prompts drawn from the exact distribution defined by the training labels. There is no human study, no user-drawn strokes, no prompts that miss the subject or click on the shadow, and the ISTD+ experiments only test the shadow removal module with GT masks, not the full prompt-aware pipeline on real images. The claim that the approach 'eliminates the need for shadow annotations' is also overstated, since training uses shadow masks as supervision.\n\nThese are addressable gaps rather than fatal flaws. The architecture is coherent and the new dataset is a useful resource, though it is not yet released. A referee should ask for human-prompt evaluation, out-of-distribution prompts, real-image tests with the full model, and community access to the data and code.\n\nOverall: the paper is a legitimate first step on a new task, but the central user-prompt claim needs experimental support beyond the auto-generated distribution. It deserves peer review, not a desk reject, but heavy revision should be expected.\n\nBest,","headline":"Prompt-aware shadow removal is a new and sensible task, but the evaluation is partly circular—auto-generated prompts come from the same masks that define ground truth—so the user-prompt claim is not yet supported.","tokens_in":13577,"tokens_out":4263,"would_cite":false,"duration_ms":34856,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Prompt-aware controllable shadow removal lets a user erase one chosen subject's shadow with a dot, line, or subject mask, with no shadow annotation at test time.","keywords":["prompt-aware shadow removal","controllable image editing","shadow detection","user prompt","spatial-frequency interaction","dense-sparse attention","shadow removal dataset","PACSRNet"],"falsifier":"Collect prompts from human users on images outside PCSRD's generation procedure and compare the predicted shadow mask against manual shadow annotations; if dot or line IoU falls far below the reported 0.816 or 0.825, the prompt-to-subject mapping does not transfer to natural prompts.","tokens_in":12555,"feed_emoji":"🎯","tokens_out":4653,"duration_ms":40376,"temperature":0.7,"pith_summary":"The paper proposes a new paradigm for shadow removal: instead of removing all shadows or requiring a precise shadow mask, the user indicates the subject whose shadow should disappear with a dot, line, or subject mask. It builds PACSRNet, a two-module network in which a prompt-aware module predicts the shadow mask of the specified subject, and a shadow removal module restores that region using both the predicted mask and prompt-aware features. The paper also contributes PCSRD, a dataset of 11,900 shadow/non-shadow image pairs with three types of prompts, and reports that PACSRNet matches or exceeds mask-based state-of-the-art methods. A sympathetic reader would care because this repositions shadow removal as an interactive, subject-selective image-editing operation rather than a batch restoration step.","feed_headline":"Point at a shadow and this model removes just that one","feed_subtitle":"A dot or line can select one subject's shadow to remove without drawing a precise shadow mask.","key_machinery":"The load-bearing mechanism is the prompt-to-shadow-mask pathway plus the prompt-guided restoration pathway. A prompt-aware module maps an image $x$ and a prompt $c$ to a predicted mask $\\hat{m}$, and a shadow removal module maps $(x, \\hat{m})$ to the restored image $\\hat{y}$. Inside the prompt-aware module, a spatial-frequency interaction block applies the discrete wavelet transform to capture the frequency difference between shadow and non-shadow regions, while the shadow removal module uses a dense-sparse local attention block that balances full dense attention with a screened sparse attention to avoid pulling in irrelevant content.","core_discovery":"On its own terms, the central discovery is a new task formulation: shadow removal should be conditioned on a user's indication of the subject whose shadow is to be erased, so the required input is a subject prompt instead of an accurate shadow mask. PACSRNet realizes this by first predicting the shadow mask of the chosen subject and then restoring the region; the paper reports that this prompt-conditioned removal is comparable to or better than mask-conditioned state-of-the-art methods, and that the predicted masks are accurate under all three prompt types.","pith_inferences":["A direct extension is iterative prompting: allowing a user to add, move, or refine a dot or line after seeing the initial prediction, which would test whether the prompt-aware module supports correction rather than only one-shot selection.","The same two-module structure could transfer to other spatially local degradations such as glare or reflection removal, where the user points at the object owning the artifact instead of annotating the artifact region.","If the paradigm holds, shadow removal in photo-editing tools could shift from a pre-processing batch operation to an interactive step where the user need not know where shadows are, only which subject matters."],"forward_implications":["A user can remove the shadow of a selected subject at test time with no shadow annotation, using only a dot, line, or subject mask.","The full PACSRNet with dot, line, and subject-mask prompts reaches 40.956, 41.056, and 41.592 dB PSNR on PCSRD, above the mask-required ShadowDiffusion baseline at 40.823 dB.","The shadow removal module alone improves shadow-region PSNR on ISTD+ by 0.63 dB over the strongest baselines.","Prompt type is robust: predicted shadow-mask IoU stays high across dot, line, and subject masks at 0.816, 0.825, and 0.872.","PCSRD supplies 11,900 multi-subject samples with three prompt types for training and benchmarking prompt-based removal."],"supporting_citations":[{"why":"Provides the ISTD+ dataset used to validate the shadow removal module against baselines.","marker":"[Le and Samaras, 2019]"},{"why":"Provides DESOBAv2 multi-subject scenes from which the PCSRD dataset is sampled.","marker":"[Liu et al., 2024a]"},{"why":"ShadowFormer is a mask-based state-of-the-art baseline and a design antecedent for global context modeling.","marker":"[Guo et al., 2023a]"},{"why":"ShadowDiffusion is the mask-required baseline that PACSRNet must match or beat on PCSRD and ISTD+.","marker":"[Guo et al., 2023b]"},{"why":"HomoFormer is both a baseline and the random-shuffle attention approach that the dense-sparse local attention block responds to.","marker":"[Xiao et al., 2024]"},{"why":"Supplies the discrete wavelet transform used by the spatial-frequency interaction block for shadow-boundary perception.","marker":"[Mallat, 1989]"},{"why":"BMNet is a bijective-mapping baseline compared on both PCSRD and ISTD+.","marker":"[Zhu et al., 2022]"},{"why":"Inpaint4Shadow is an inpainting-based baseline used in the PCSRD and ISTD+ comparisons.","marker":"[Li et al., 2023]"}],"fun_headline_variants":["Point at a subject, its shadow is removed","Erase just one shadow with a dot or line","Prompt-based shadow removal without shadow masks","Select a subject, remove its shadow only","Shadow removal on demand: pick your subject"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The demonstration of prompt-driven control rests on dot and line prompts generated from subject masks by a fixed dynamic-programming routine being representative of what real users would actually draw.","fun_headline_variants_meta":{"raw":{"variants":["Point at a subject, its shadow is removed","Erase just one shadow with a dot or line","Prompt-based shadow removal without shadow masks","Select a subject, remove its shadow only","Shadow removal on demand: pick your subject"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000362,"raw_usage":{"total_tokens":1917,"prompt_tokens":872,"completion_tokens":1045,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":977}},"tokens_in":488,"tokens_out":1045,"duration_ms":9141,"temperature":1.0,"reasoning_tokens":977,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:40:45.040192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect prompts from human users on images outside PCSRD's generation procedure and compare the predicted shadow mask against manual shadow annotations; if dot or line IoU falls far below the reported 0.816 or 0.825, the prompt-to-subject mapping does not transfer to natural prompts.","supporting_citations":[{"cited_title":"Shadow removal via shadow image decomposition","cited_arxiv_id":null,"evidence_quote":"Provides the ISTD+ dataset used to validate the shadow removal module against baselines."},{"cited_title":"A theory for multiresolution signal decomposition: the wavelet representation","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete wavelet transform used by the spatial-frequency interaction block for shadow-boundary perception."},{"cited_title":"Bijective mapping network for shadow removal","cited_arxiv_id":null,"evidence_quote":"BMNet is a bijective-mapping baseline compared on both PCSRD and ISTD+."}],"review_version":1}