{"id":"490c3beb-e6ab-483a-a661-246f6da884ba","arxiv_id":"2607.26203","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"WildShadowRemover fine-tunes a pretrained video diffusion model with LoRA plus detail-injection and depth conditioning to produce temporally consistent shadow-free videos, trained on a new synthetic dataset.","lead":"The paper adapts a video diffusion model to remove shadows from videos, adding modules to preserve fine detail and using depth maps for guidance. It also builds a synthetic paired video dataset for training and evaluation, but its claims about real-world performance are supported only by qualitative examples.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'in-the-wild' claim rests on qualitative real-world examples; the paper's own Limitations concedes synthetic-data reliance, leaving synthetic-to-real generalization the key untested assumption.","rationale":"The reader's weakest_assumption is the same as the one I find most load-bearing: the synthetic-to-real transfer assumption. I agree with the reader's analysis. The strongest empirical support in the paper is Table 1, Table 2, and Table 4, all on synthetic benchmarks or public image datasets; real-world evidence is limited to Fig. 5 qualitative samples. The abstract's 'strong generalization across challenging in-the-wild scenarios' is exactly the claim that needs real-world quantitative validation. The paper's own Limitations section concedes synthetic-data reliance, so this is not an external standard being imposed. The proposed test is a concrete way to measure the domain gap: if the model and baselines are evaluated on real paired videos, the PSNR/SSIM/Ewarp numbers would either support or refute the generalization claim. I do not see an internal inconsistency in the method; the concern is about evidence sufficiency, not correctness of the architecture. Therefore the appropriate verdict remains CONDITIONAL/UNCHANGED — conditional because the claim can be verified by additional real-world evaluation; not REJECT because there is no positive evidence of failure.","tokens_in":11970,"tokens_out":5930,"duration_ms":63387,"concrete_test":"Construct a real paired video benchmark: place a fixed camera on a tripod overlooking a scene where a shadow moves across a static ground plane (e.g., a building or car shadow during the day), record continuously, and manually select aligned frame pairs with and without the shadow to serve as shadow/shadow-free references. Evaluate the trained model plus OmniSR/StableSR on these pairs, reporting PSNR, SSIM, and Ewarp. If the method's PSNR/SSIM gain is not significant or Ewarp is not lower, the 'strong generalization across challenging in-the-wild scenarios' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion is that a model trained on WildShadow (synthetic clips from 3D-FRONT/iCity/Infinigen) generalizes to unconstrained real videos. This is load-bearing because the title and abstract claim 'in-the-wild' and 'strong generalization.' Yet the only real-world evidence is Fig. 5: qualitative stills from DL3DV and UAV footage with no paired ground truth, no metrics, no user study, and no comparison of temporal consistency on real data. The paper's own Limitations state that the framework 'relies primarily on synthetic paired data for training' and that leveraging real-world videos remains open. Section 'Our Video Dataset' further states outdoor shadow-free targets remove 'only shadows cast by direct illumination,' while real videos contain penumbrae, interreflections, and ambient occlusion; the size of this domain gap is never measured. Table 2's cross-domain study is indoor-vs-outdoor synthetic, not synthetic-to-real. Consequently, the sentence 'strong generalization across challenging in-the-wild scenarios' is an extrapolation from synthetic benchmarks, not a demonstrated result. The missing PSTNet comparison compounds this: the only prior video shadow removal method is mentioned but never compared, so 'outperforms existing approaches' in the video setting is also unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WildShadowRemover, a video shadow removal framework built on a pretrained video diffusion model (Wan-Control). It adapts the DiT through LoRA fine-tuning, condenses the input shadow video and Depth Anything 3 depth maps, and augments the frozen VAE decoder with a shadow-mask-guided detail injection module that uses frequency-decomposed modulation to restore high-frequency details without re-injecting shadow remnants. The authors also introduce WildShadow, a large-scale synthetic video dataset rendered from 3D-FRONT, iCity, and Infinigen, with paired shadow/shadow-free clips, and a related image dataset (WildShadow-I) for cross-domain evaluation. Experiments compare against image shadow removal baselines on ISTD+, SRD, INS, and WildShadow-I, and include an ablation study, a synthetic video benchmark comparison (against OmniSR and a frame-by-frame variant), and qualitative examples on real DL3DV and UAV videos. The paper claims state-of-the-art performance and strong generalization to in-the-wild videos.","tokens_in":12329,"tokens_out":4387,"duration_ms":46764,"significance":"If the claims were fully supported, this would be among the first video-diffusion-based video shadow removal methods and would provide a useful large-scale synthetic benchmark for the community. The technical components are reasonable and mostly standard: flow matching, LoRA adaptation, depth conditioning, and a residual detail-injection module; the ablations in Table 3 give independent evidence for each component, and the paper does not rely on circular or parameter-fitting claims. The main weakness is that the central 'in-the-wild generalization' claim is supported only by qualitative examples on real videos, while all quantitative evidence is on synthetic benchmarks or image datasets. Additionally, the only prior video shadow removal method (PSTNet) is cited but never compared, leaving the 'state-of-the-art' video claim unsubstantiated. The method is not presented as SOTA on conventional image benchmarks, and the abstract overstates the results relative to Table 1. The dataset contribution is potentially valuable, though reproducibility would benefit from a clear release/access statement.","major_comments":[{"comment":"The core claim of strong generalization to in-the-wild videos is not supported by the evidence shown. Real-world evaluation on DL3DV and UAV videos (Fig. 5) is purely qualitative, with no ground truth, no metrics, no user study, and no temporal-consistency measurement. The only quantitative evidence (Tables 1, 2) comes from synthetic benchmarks, while the Limitations section explicitly states that the framework 'relies primarily on synthetic paired data for training' and that leveraging real-world videos remains open. Moreover, the dataset section says outdoor shadow-free targets remove only shadows cast by direct illumination, whereas real videos contain penumbrae, interreflections, and ambient occlusion; the size of this domain gap is never measured. Please add quantitative real-world evaluation (e.g., a paired real-video subset, no-reference or perceptual metrics, temporal-consistency","section":"Real-World Evaluation; Limitations"},{"comment":"PSTNet (Chen et al. 2024) is the only prior video shadow removal method cited, but it is never compared against. Table 4 compares only OmniSR (image-based, frame-by-frame) and the authors' own frame-by-frame variant. Without a PSTNet comparison under the same protocol, the claim of 'state-of-the-art performance' and 'outperforming existing approaches' in the video setting is unsubstantiated. Either include PSTNet in the video benchmark (if the source data are available, as the paper notes they are 'unavailable' — but then a claim of superiority should be explicitly qualified) or restrict the claim to the image-based baselines evaluated.","section":"Video Shadow Removal Evaluation; Related Work"},{"comment":"The abstract and conclusion claim that the method 'outperforms existing approaches', but Table 1 shows it is below several methods on ISTD+ (34.72 dB vs. 35.72 for HomoFormer and 35.46 for ShadowFormer) and below HomoFormer on SRD. The paper's own text ('Results on public and our benchmarks') says the method achieves 'competitive' performance on ISTD+/SRD. The claims should be aligned with the evidence: report the result as SOTA only on the proposed benchmark and for the video model, or acknowledge the public-benchmark trade-off explicitly.","section":"Abstract; Table 1"}],"minor_comments":[{"comment":"The kernel definition is incomplete: 'k = 1/25 1_{5×5}' should specify that the 5×5 filter contains all ones and state explicitly that the channel-wise convolution is applied independently per channel.","section":"Eq. (7)"},{"comment":"The 'Ours (frame-by-frame)' variant has a higher Ewarp (2.208) than OmniSR (2.032), while the video model achieves 1.775. Please add a sentence explaining whether the frame-by-frame variant is expected to be less temporally consistent and whether the difference is meaningful given the small benchmark size.","section":"Table 4"},{"comment":"The '–' entries (e.g., PhaSR on SRD, several methods on WildShadow-I) should be explicitly footnoted to indicate whether the result is unavailable or the method was not applicable, so that readers can distinguish 'not reported' from 'not evaluated'.","section":"Table 1"},{"comment":"The construction of WildShadow-I from 'INS, iCity, and Infinigen' is underspecified: it is unclear whether the images are sampled as single frames from the video clips or generated separately, and what the 'INS' split contains. Clarifying the sampling protocol and the exact training/test split would improve reproducibility.","section":"Our Video Dataset"},{"comment":"No code or dataset release URL is provided. Given that the paper introduces two new benchmarks (WildShadow and WildShadow-I), an availability statement is important for the community to reproduce and build upon the contribution.","section":"Implementation Details"},{"comment":"The reference format is uneven (e.g., 'Smair 2024' is cited as (Smair 2024) but the reference list uses a different key formatting), and there are minor typographical issues in the extracted abstract. A careful proofread is recommended.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central claim is broader than the evidence, but the core approach is sound and the dataset has potential value. The requested changes — real-world quantitative evaluation or scoped claims, and a comparison with PSTNet — are within the scope of a major revision. I would not recommend rejection at this stage, provided the authors either supply the missing evidence or carefully rewrite the claims to match what is actually demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is the first paper to adapt a pretrained video diffusion model to video shadow removal, and it introduces a large synthetic video dataset for the task. The core pipeline is sensible — LoRA fine-tune Wan-Control with depth conditioning, then a shadow-mask-guided detail injection module to recover texture. The ablations are clean and each component shows a measurable gain. The dataset (4,500+ clips from 3D-FRONT, iCity, Infinigen) is a genuine resource if they release it.\n\nThe soft spot is the one the stress-test flags, and it is load-bearing. The title and abstract claim \"in-the-wild\" and \"strong generalization across challenging in-the-wild scenarios,\" but the only real-world evidence is Fig. 5: qualitative stills from DL3DV and UAV footage with no ground truth, no metrics, no user study, no temporal-consistency measure. The paper's own Limitations section admits the framework \"relies primarily on synthetic paired data for training.\" The cross-domain study in Table 2 is indoor vs. outdoor synthetic, not synthetic vs. real. So the central generalization claim is an extrapolation, not a demonstrated result.\n\nThe abstract also says the method \"outperforms existing approaches,\" but Table 1 shows it is below ShadowFormer and HomoFormer on ISTD+ and below several methods on SRD. It wins on INS and on their own WildShadow-I, which is fine — but the blanket statement is too strong. More concerning, PSTNet, the only prior video shadow removal method, is mentioned but never compared. The video evaluation only benchmarks against OmniSR frame-by-frame and their own frame-by-frame variant. That leaves \"outperforms existing approaches\" in the video setting unsubstantiated.\n\nMinor issues: no code or data release, and the mask estimation uses a simple ratio with an epsilon heuristic — probably fine, but no sensitivity analysis. These are minor relative to the generalization gap.\n\nWho is this for? People working on shadow removal or video restoration will want to know about the approach and the dataset. It deserves a serious referee: the idea is new, the engineering is careful, and the paper is honest about some limitations. But the claims need to be pulled back to what is actually measured, and the authors should add a quantitative real-world evaluation or at least a user study, a PSTNet comparison, and release artifacts.\n\nRecommendation: send to peer review, but expect major revision to align the claims with the evidence.","headline":"First video-diffusion video shadow removal with a substantial synthetic dataset, but the in-the-wild claim outruns the evidence: real-world evaluation is qualitative only.","tokens_in":12753,"tokens_out":1651,"would_cite":true,"duration_ms":17937,"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":"Adapting a video diffusion model with LoRA and shadow-mask-guided detail injection yields temporally consistent shadow removal in unconstrained videos.","keywords":["video shadow removal","video diffusion models","LoRA fine-tuning","detail preservation","temporal consistency","synthetic dataset","shadow mask","in-the-wild generalization"],"falsifier":"Record the same real scene twice with a movable light or occluder so that one video has shadows and the other is shadow-free but otherwise identical; run the trained model and a strong frame-by-frame image method on the shadowed video, and compare both against the physically truthful shadow-free frames using PSNR/SSIM plus a temporal-consistency metric. If the video model does not beat the per-frame baseline on temporal consistency and quality, the central claim fails.","tokens_in":11891,"feed_emoji":"🎥","tokens_out":7682,"duration_ms":74649,"temperature":0.7,"pith_summary":"Existing video shadow removal methods work frame-by-frame and lose temporal coherence. This paper argues that a pretrained video diffusion model carries enough implicit knowledge of how shadows move and change with objects that only a light-weight LoRA fine-tuning step is needed to turn it into a video shadow remover. To stop the diffusion model's latent decoding from blurring away fine texture, the authors insert a shadow-mask-guided frequency-decomposed modulation into the frozen VAE decoder, which restores high-frequency detail where the mask signals shadow while keeping shadow patterns out. They also condition on monocular depth to disambiguate shadows under complex lighting, and train on WildShadow, a large new synthetic dataset spanning indoor, urban, and natural scenes. The paper's claim is that the resulting model beats per-frame image methods on reconstruction quality and temporal consistency and generalizes to real videos, with real-world evaluation shown qualitatively.","feed_headline":"Video diffusion model removes shadows and keeps frames stable","feed_subtitle":"LoRA fine-tuning plus depth guidance keeps textures and temporal coherence in unconstrained videos.","key_machinery":"The carrying mechanism is a two-stage pipeline built around a pretrained latent video diffusion model. Stage one is low-rank (LoRA) fine-tuning of the diffusion transformer, conditioned on the input shadow video and on monocular depth maps, trained with a flow-matching objective to predict a coarse shadow-free latent video. Stage two is a shadow-mask-guided detail injection module: a soft shadow mask M = 1 − X/(Y0 + ε) is estimated by comparing the input with the coarse prediction, and at each decoder stage the encoder and decoder features are concatenated, split into low- and high-frequency components using a fixed 5×5 mean filter, and independently modulated by mask-predicted affine parame","core_discovery":"On the paper's own terms, the central discovery is that a video foundation model's spatiotemporal prior, not a task-specific temporal architecture, is what solves video shadow removal. The method performs shadow removal in the latent space of a pretrained video diffusion model: a diffusion transformer conditioned on the shadow video and a depth video is fine-tuned with LoRA using a flow-matching objective, producing a coarse shadow-free latent that is decoded to a coarse video. Because the frozen decoder loses thin structures and textures, the authors augment each decoder stage with a detail injection module driven by a soft shadow mask computed as the ratio between input and coarse output.","pith_inferences":["The real-world 'in-the-wild' claim rests only on qualitative examples; a quantitative evaluation on real videos with physically captured ground truth would be the natural test and is not in the paper.","The frequency-decomposed modulation idea could be extracted as a general detail-restoration module for other conditional diffusion models where residual artifacts and texture are entangled.","Given that depth conditioning adds only a small improvement in the ablations, a cheaper or more robust geometry cue may capture the same benefit; this is an inexpensive experiment to run.","The dependence on synthetic data means the method's ceiling is set by how faithfully renderers model real illumination; self-supervised adaptation on real videos is the stated open problem."],"forward_implications":["If the method is right, video shadow removal no longer needs per-frame post-processing to stay temporally stable; the diffusion model supplies coherence across frames.","Clean shadow-free video can be piped into downstream tasks such as video editing, object tracking, intrinsic decomposition, and 3D content creation, which have all been hampered by shadow artifacts.","The new large paired synthetic dataset, with its benchmark protocol, gives the community a common training and evaluation resource for video shadow removal.","Because the same model handles single images, a single video diffusion-based remover could unify image and video shadow removal, matching specialized image methods on standard benchmarks."],"fun_headline_variants":["Diffusion model erases video shadows, keeps details crisp","LoRA-tuned diffusion strips shadows from video, preserves texture","Depth-guided diffusion removes shadows in unconstrained video","Video diffusion fine-tuned with LoRA erases shadows, keeps motion stable","Shadow removal for video: diffusion prior keeps temporal coherence"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the synthetic dataset, whose shadow-free targets are made by removing only direct-light shadows in rendered scenes, is representative enough of real in-the-wild shadow videos for the trained model to generalize — a premise the paper tests only qualitatively on real footage.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model erases video shadows, keeps details crisp","LoRA-tuned diffusion strips shadows from video, preserves texture","Depth-guided diffusion removes shadows in unconstrained video","Video diffusion fine-tuned with LoRA erases shadows, keeps motion stable","Shadow removal for video: diffusion prior keeps temporal coherence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001179,"raw_usage":{"total_tokens":4682,"prompt_tokens":695,"completion_tokens":3987,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":439,"completion_tokens_details":{"reasoning_tokens":3904}},"tokens_in":439,"tokens_out":3987,"duration_ms":26152,"temperature":1.0,"reasoning_tokens":3904,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:29:10.798313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the same real scene twice with a movable light or occluder so that one video has shadows and the other is shadow-free but otherwise identical; run the trained model and a strong frame-by-frame image method on the shadowed video, and compare both against the physically truthful shadow-free frames using PSNR/SSIM plus a temporal-consistency metric. If the video model does not beat the per-frame baseline on temporal consistency and quality, the central claim fails.","supporting_citations":[],"review_version":1}