acceptodds
Under review as a conference paper at ICLR 2027

ShadowEdit: Relation-Aware Modeling for Selective Shadow Removal

Abstract

Existing shadow removal formulations assume that spatially localized shadows correspond to a uniquely defined restoration target, leaving intervention intent unspecified in multi-shadow scenes, yet a single shadowed observation may correspond to multiple plausible interventions, rendering conventional all-shadow restoration formulations fundamentally under-specified. We recast selective shadow removal as a relation-indexed intervention problem, where caster–shadow–receiver relations define an explicit intervention representation for target modification and effect preservation. ShadowEdit separates intervention specification from appearance reconstruction and compiles relation-level actions into constrained image updates, constraining image changes to the compiled permission support while evaluating preservation against action-specific references. Under exact Kubric counterfactuals, ShadowEdit reduces overlap error by 26.5% relative to a full-information spatial control with identical relation/action inputs and output permissions. Mechanism-aligned analyses show that action binding specifies intervention identity, protection constraints preserve unselected effects, and permission grounding restricts unintended edits. On DESOBAv2, ShadowEdit reduces overlap error by 19.4% over an equal-information spatial control under matched supervision and output permissions and supports held-out combinations of seen relation primitives, while a separately trained relation-free restoration variant of ShadowEdit achieves the highest full-image PSNR among evaluated methods on ISTD+ and SRD. Code is available in the supplementary materials.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Shadow Removal, Shadow Editing, Relation-Aware Modeling, Intervention Ambuiguity

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_fctYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a relation-indexed representation for selective shadow removal. The task is an interesting topic in shadow manipulation, and the authors identify that the current gaps are: 1) single caster, multiple receivers, and 2) shared overlap pixels between shadows. The authors formulate the caster-shadow-receiver relations and corresponding remove/preserve instructions, and propose a method to learn the correct reconstruction conditioned on the given relations and instructions. The authors evaluated the proposed method on both Kubric synthetic scenes and real DESOBAv2 images.

Critical strengths
  1. The authors' definition of the task is clear. The formulation of caster-shadow-receiver relations is straightforward, with well-defined remove/preserve actions.

  2. The evaluation strategy is clear. The authors first test and validate the proposed method on fully controllable synthetic data using Kubric. Then the method is evaluated on real images from DESOBAv2 data.

Critical weaknesses
  1. The organization and presentation make the method section hard to follow. A constructed example from a Kubric scene would help explain Sec. 3.1 and Sec. 3.2. Fig. 2 also does not clearly show the data flow described in Secs. 3.2–3.5, especially the interaction between the appearance features and the conditioner and the predicted residual in (a). The largely sequential flow in (b) does not clearly reflect the interactions and dependencies described in the sections. Can the authors clarify the actual operations involved in the Action-CSR conditioner, its intermediate outputs, and its interactions with other components of the network?

  2. The fairness of the comparison on real images is questionable. The proposed method is trained on DESOBAv2, but the authors have not specified whether the compared models are also retrained on the same data. If, as reported, the compared models use their official checkpoints, then the comparison is not fair, since the images in other shadow removal datasets, like ISTD+ and SRD, differ a lot from the DESOBAv2 data. Consequently, the fairness of the visual comparison is also affected. Can the authors clarify what datasets the compared models were trained on and report results where all models (or some representative compared models) are trained on the same DESOBAv2 dataset?

  3. The authors' strategy is to verify the idea on fully controllable synthetic scenes. However, the experimental results do not show comparative visual evidence of the benefit of the proposed method on the two failure modes in Kubric scenes. Fig. 4 only shows an example of the proposed method on overlapping shadow pixels. Similarly, the authors do not provide clear comparative visual evidence in real scenes. The improvements in MAE scores do not show whether the improvements are visually meaningful. Can the authors provide qualitative results for the single-caster, multiple-receiver case and the shared-overlap case, including visual comparisons between existing methods, ablated versions of the proposed method, and the full proposed method on both synthetic and real scenes?

Questions for authors
  1. Please see my question in Weakness 1.

  2. Please see my question in Weakness 2.

  3. Please see my question in Weakness 3.

  4. How substantial are the reported improvements in MAE (in Table 2 and Table S12) for this task? The reported improvements are all below 0.1 on the original scale.

  5. What is on the right side of each panel in Fig. 1? Is it just an illustration of the relations and actions, or is it some intermediate outcome? The authors should clarify this, as the layout and positions of the casters and shadows do not match the images on the left side. This could lead to misunderstanding.

  6. I suggest that the authors add an illustrative figure (in the main paper or appendix) showing example rendered scenes from the Kubric synthetic dataset. It would help readers understand the setting, the coverage of the failure modes, and the complexity or insufficiency of the synthetic scenes.

5 comments
AuthorsOpenReview ↗

Thank you for the concrete checks on the method figure, the DESOBA comparison protocol, and the visual scale of the selective gains. We answer W1–W3 and Q4–Q6 below.

1. Action-CSR Data Flow and Figure 2

Concern: Sections 3.1–3.5 and Figure 2 do not make the data flow clear, especially how appearance features, the Action-CSR conditioner, and the residual interact.

Direct answer. We agree the original Figure 2 was hard to read as an execution path. The intended order is: compile supports → update relation tokens → assign them to pixels → gather appearance evidence → modulate only inside the permission gate → predict a residual → hard-compose with the input.

Given image XX, relation set RR, and action vector aa, the compiler produces MremoveM_{\mathrm{remove}}, MprotectM_{\mathrm{protect}}, and MpermM_{\mathrm{perm}}. MremoveM_{\mathrm{remove}} marks the selected shadow support; MpermM_{\mathrm{perm}} is the only region where the network may write. Dense relation fields yield FRF^{R}, and each valid caster–shadow–receiver unit becomes a labelled token ti0t_i^{0}. After invalid tokens are dropped, the remaining tokens interact to form T′T', so a selected relation can use context from protected or competing relations without losing its identity.

At decoder scale ℓ\ell, image features query T′T' to produce ZℓZ_{\ell}. ZℓZ_{\ell} assigns relation context to locations; it does not authorize writing. In parallel, frozen priors from XX give action-independent appearance features P0P_{0}. Action-dependent focus yields PaP_{a}, and the base branch combines XX, PaP_{a}, and support cues into the appearance candidate YgY_{g}. Thus T′T' / ZℓZ_{\ell} identify the requested intervention, while P0P_{0}, PaP_{a}, and YgY_{g} supply appearance evidence.

Image features FℓF_{\ell}, ZℓZ_{\ell}, and the aligned controls then predict scale/bias fields that modulate decoder features only inside MpermM_{\mathrm{perm}}, producing FeℓF_{e}^{\ell}. The residual refiner uses FRF^{R}, T′T', and FeℓF_{e}^{\ell} to predict Δ\Delta. The final output is

Y=(1−Mperm)X+Mperm clip(Yg+Δ).Y=(1-M_{\mathrm{perm}})X+M_{\mathrm{perm}}\,\mathrm{clip}(Y_{g}+\Delta).

Pixels outside MpermM_{\mathrm{perm}} are copied from XX; only permitted pixels receive the reconstruction and residual correction.

Revision locations. Main-paper Section 3.1 and the Figure 2 caption; Sections 3.1–3.5, especially Section 3.5; Figure 2 caption; Eqs. (8)–(10); Appendix B.2.2 . Supplementary Appendix B.2.2. Figure 2 artwork is retained.

2. DESOBA Comparison Fairness (W2 / Q2)

Concern: Table 1 may be unfair if ShadowEdit is trained on DESOBA v2 while compared models use official checkpoints from other datasets.

Response. We revised the wording to distinguish a shared test-time evaluation protocol from matched training. External baselines were retrained to convergence on the same DESOBA v2 training split, using their original architectures and inference procedures. After prediction generation, all outputs were evaluated on the same fixed DESOBA test list, references, masks, metric-stage resize and preprocessing, and frozen scorer. The controlled within-method comparison is ShadowEdit-Mask versus ShadowEdit-Full under the matched ShadowEdit training and evaluation setup.

The controlled within-method comparison is ShadowEdit-Mask versus ShadowEdit-Full, which uses the matched ShadowEdit training and evaluation setup. ISTD+ and SRD use separately trained relation-free configurations and are reported as restoration-compatibility evaluation.

Revision locations. Main-paper Section 4.2 and Table 1 caption; Supplementary Appendix D.4, especially the cross-method comparison protocol and the comparison-scope table.

3. Two Kubric Failure Modes, Ablations, and MAE Scale (W3 / Q3 / Q4)

Concern. The two Kubric failure modes lack visual comparisons with ablations and existing methods. Moreover, the reported MAE improvements are below 0.1, making their practical significance unclear.

Response. We clarify the two failure modes using their respective measured action subsets and report the corresponding ablation results.

For receiver-specific attribution, one caster creates shadows on multiple receivers. The action selects which receiver to edit and which to preserve. This subset contains 10 scenes and 120 actions.

ConfigurationTarget ↓Protected ↓Leakage ↓
Caster–Shadow2.804.602.60
+ Receiver identity2.403.592.55
+ Action binding1.123.482.52
+ Protection1.061.822.48
Full1.021.751.98

Adding receiver identity reduces Target and Protected errors by 14.3% and 22.0%, respectively. Overlap is N/A because this subset has no selected–protected intersection.

For shared-overlap attribution, shadows from different casters overlap on the same receiver. This subset contains 11 scenes and 154 actions.

ConfigurationTarget ↓Protected ↓Overlap ↓Leakage ↓
Full-information Spatial1.352.103.201.98
Independent Tokens1.221.902.781.98
Full Interaction1.081.702.401.96

Full reduces Overlap error by 25.0% compared with Spatial and 13.7% compared with Independent Tokens.

On the separate 352-action overlap-valid pool, the progressive C-S-R, +Action, +Protection, and Full configurations achieve Target/Protected/Overlap errors of 3.221/6.682/7.421, 1.080/6.550/7.280, 1.020/2.250/2.950, and 0.970/1.563/2.094, respectively. The equal-information Spatial/Independent/Full comparison gives Overlap errors of 2.850/2.420/2.094. These results use different ablation designs and are therefore reported separately.

Target, Protected, and Overlap are RGB MAE multiplied by 10210^2 on normalized images. On the 352-action Kubric pool, the Overlap reduction from 2.850 to 2.094 corresponds to a raw MAE decrease from 0.02850 to 0.02094 (0.00756, or about 1.93 levels on an 8-bit scale). The DESOBA reduction from 3.60 to 2.90 corresponds to 0.007 raw MAE (about 1.79 8-bit levels). These are regional averages.

Figure 1(a) shows Kubric ShadowEdit inference outputs alongside the corresponding renderer references and change/topology diagnostics; Figure 1(b) shows DESOBA model outputs under different remove instructions. Figure 4 presents DESOBA outputs and baseline comparisons on the left and Kubric caster-removal outputs on the right. We use the measured ablations and quantitative baseline comparisons as evidence for method comparisons; the figures provide qualitative examples of model outputs.

Revision locations. Main-paper Section 4.4, Table 2, and the new cross-reference to Supplementary Table S34; Supplementary section on Kubric failure-mode ablations; Supplementary Tables S34, S37, and S17.


4. Figure 1 Right Panel and Kubric Scene Coverage (Q5 / Q6)

Concern: The right side of Figure 1 does not spatially match the left images. The paper also needs rendered Kubric scene examples that show coverage of the failure modes.

Response. Response. Figure 1(a) shows Kubric ShadowEdit inference outputs alongside renderer references and change/topology diagnostics; Figure 1(b) shows DESOBA model outputs under different remove instructions. Figure 4 presents DESOBA outputs and baseline comparisons on the left and Kubric caster-removal outputs on the right. The full Kubric diagnostic contains 36 scenes and 472 action–scene pairs: 26 overlap-valid scenes with 352 actions and 10 receiver-specific scenes with 120 actions.

Revision locations. Main-paper Figure 1 and Figure 4 captions; Supplementary Appendix A.2 and Table S5 for the scene/action inventory; Supplementary section on Kubric failure-mode ablations and Table S34 for topology-specific measured results.


We have kept the renderer-based scene illustrations, the receiver-specific subset, the shared-overlap subset, and the 352-action pooled evaluation distinct. This avoids presenting different cohorts or evaluation protocols as if they were one experiment.

Dear Reviewer fctY,

Thank you again for your time and constructive feedback.

With only several hours remaining in the author–reviewer discussion period, we would greatly appreciate it if you could take a moment to review our rebuttal and revised manuscript. We have tried our best to address your concerns through comprehensive analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and look forward to your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_SB2WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper frames selective shadow removal as a relation-indexed intervention problem, where each caster–shadow–receiver triplet serves as an addressable unit with an independent remove-or-preserve action. The proposed ShadowEdit framework decouples intervention specification from appearance reconstruction. The method is evaluated on two axes, selective editing fidelity on synthetic Kubric renderer counterfactuals and an extended DESOBA v2 real-image partial-target track, measured by Target correction, Protected preservation, Overlap handling, and off-target Leakage and conventional full-removal restoration quality on DESOBA v2, ISTD+, and SRD under a relation-free configuration. The authors claim that relation-indexed execution substantially reduces overlap error compared with spatial-control baselines under matched information and output permissions, and generalizes to held-out combinations of seen relation primitives.

Critical strengths
  1. The paper formalizes selective shadow removal as a relation-indexed counterfactual intervention task and clearly articulates the ambiguity of action-invariant global masks in multi-shadow scenes. This formulation brings explicit remove/preserve semantics to each shadow contribution and provides a structured interface for fine-grained editing control.

  2. The combination of exact Kubric renderer counterfactuals and audited real-image DESOBA v2 partial targets allows both controlled mechanism validation and real-world usability assessment. The multi-dimensional evaluation provides a more nuanced diagnosis of selective editing behavior than a single overall fidelity metric.

  3. By design, pixels outside the compiled permission mask are copied directly from the input, eliminating unauthorized edits to protected regions and ensuring zero leakage outside the writable support. This is a desirable property for controllable editing systems.

  4. The paper provides progressive component ablations, equal-information representation comparisons, relation-support perturbation tests, multi-action generalization checks, seed-level robustness statistics, and a failure taxonomy. These analyses offer reasonable transparency into how each component contributes to the final performance.

Critical weaknesses
  1. The entire framework relies on oracle-level caster–shadow–receiver mask triplets at inference time, which are almost never available in real editing scenarios. Although the authors state that automatic relation discovery is out of scope, this assumption is so restrictive that the proposed “selective editing” capability has very limited deployable value. If a user can already manually annotate per-relation caster/shadow/receiver masks, a conventional mask-guided inpainting method can already achieve targeted removal; the paper fails to demonstrate that the heavy relation-indexed machinery delivers commensurate benefits to justify the additional annotation cost. The supplementary experiment with an automatic detector shows noticeable performance degradation under imperfect inputs, further weakening the practical relevance of the core claims, which are all derived from oracle inputs.

  2. The main comparisons against “union-mask-only”, “permission-only”, and “full-information spatial control” are all in-house ablated variants constructed by the authors, not established state-of-the-art selective editing methods. The adapted external baselines are simple re-purposed versions without task-specific optimization, so their inferior performance does not convincingly prove the superiority of the relation-indexed paradigm over other conditional editing paradigms.

  3. The receiver mask itself is generated via simple geometric heuristics (trapezoid extension, convex hull dilation) rather than semantic surface segmentation, so it functions more as extra spatial context than genuine physical relation modeling. The paper therefore contributes more at the engineering/system level than at the conceptual or algorithmic level, and does not provide fundamentally new insights into shadow physics or counterfactual image synthesis.

  4. On standard full-removal benchmarks, the method shows strong results on ISTD+ but performs poorly on SRD, where its shadow-region RMSE is notably worse than multiple existing baselines such as Inpaint4Shadow. The authors highlight the ISTD+ advantage while downplaying the SRD weakness, leading to an unbalanced presentation of restoration capability. More critically, the method incorporates powerful frozen external priors, which are known to bring large gains to image restoration. Although an appearance-free ablation is provided, the paper does not sufficiently disentangle how much performance comes from the relation-indexed design versus the strong external priors. It is plausible that most of the restoration improvement originates from the priors, not from the relation modeling itself.

  5. The DESOBA v2 partial-target track includes 80 out of 200 held-out images with manually constructed targets, not physically captured counterfactuals. The quality and consistency of these manually edited references are hard to verify, and they may introduce annotation bias into the selective-editing evaluation.

  6. Furthermore, the held-out combination test only tests new combinations of seen relation primitives on the same data distribution; there is no evidence that the relation-indexed mechanism generalizes to unseen object categories, scene types, or lighting conditions. The paper also does not evaluate soft shadow / penumbra transitions, where the hard composition boundary may produce visible artifacts.

Questions for authors

see above.

3 comments
AuthorsOpenReview ↗

Thank you for your comments on relation inputs, baselines, appearance priors, reference construction, and generalization. We address W1–W6 in five groups, combining W1 and W3 because both concern the relation-input interface.

1. Oracle Relations, Annotation Cost, and the Receiver Cue (W1 / W3)

Concern. The method appears to require oracle caster--shadow--receiver masks at inference. Such triplets are rarely available in real editing. Mask-guided inpainting may already be sufficient, and the receiver cue is only a geometric heuristic.

Direct answer. We use caster--shadow--receiver triplets as structured editing instructions. On DESOBA, SSISv2 provides the initial caster--shadow associations. SAM and Mask R-CNN assist only a small number of caster or receiver contour corrections; they do not generate all shadow masks or relation pairings. Seven images required targeted corrections, involving nine manual relations and a net addition of four.

The receiver cue specifies where a relation acts; it is not a reconstruction of the physical surface. We use a downward trapezoid for ground-like regions and the hull of a dilated shadow otherwise. The key distinction from a union mask is that our relation-indexed representation preserves each caster, shadow, receiver, and Remove/Preserve action. A union support merges overlapping relations and cannot represent opposite actions on the same region.

Under equal information and output permission, the Kubric Overlap MAE (×102\times10^2) for Spatial/Independent/Full is 2.850/2.420/2.094. The integrated pathway gives 2.930/2.490/2.180, while the permission-only control gives 6.310. Automatic relation inputs reach precision/recall/F1 of 0.940/0.910/0.925. Replacing curated inputs with detector outputs changes DESOBA Target/Overlap MAE from 1.900/2.900 to 2.064/3.180. These results measure sensitivity to relation quality. The human-preference audit evaluates output quality.

Revision. Section 3.1 and Appendix C now explain how relations are constructed and how the receiver cue is used. Table S48 reports the annotation statistics. The main text points to the progressive controls in Table 2, equal-information results in Table S37, and detector/support stress tests in Tables S27 and S41. Section 5 also clarifies the limits of automatic relation discovery and distinguishes the output-preference audit in Table S31 from an annotation-cost study.


2. In-House Controls and Adapted External Baselines (W2)

Concern. Spatial, permission, and union-mask controls are developed in-house, while the adapted external editors may not be optimized for selective shadow removal.

Direct answer. Spatial/Independent/Full are controlled comparisons of representation and relation interaction under common relation/action information and output permission.

We report adapted external editors in a separate matched-interface comparison:

Method/controlTarget MAE (×102\times10^2)Overlap MAE (×102\times10^2)
MetaShadow-RelAdapter2.343.60
InstructPix2Pix-Perm2.483.83
ControlNet-Inpaint-Rel2.403.65
Spatial2.303.72
Full2.023.07

The adaptations provide caster-conditioned support, instruction conditioning with common output permission, or spatial relation controls. All methods use the same DESOBA held-out sources, actions, references, output permissions, and scorer. This is a separate track: its Spatial/Full Target/Overlap results (2.30/3.72 and 2.02/3.07) are not replacements for the primary selective results (2.20/3.60 and 1.90/2.90).

Revision. Section 4.1 p7 now identifies Spatial, Independent, and Full as in-house representation controls. Table S18(a) reports the separate DESOBA matched-interface comparison, and Appendix A.4 pp22–23 describes the adapter interfaces. Table S19 lists the available protocol details and marks undisclosed training-budget fields as Not specified. The discussion keeps this comparison separate from the primary selective results in Tables S10 and S11.

3. SRD, ISTD+, and Appearance Priors (W4)

Concern. The results on SRD are weak, ISTD+ receives more emphasis, and the gains may come from frozen appearance priors rather than relation modeling.

Direct answer. We do not claim uniform superiority on full-removal benchmarks. ISTD+ and SRD evaluate restoration without relation conditioning, so they measure compatibility with full-removal tasks rather than selective execution. Shadow-region Lab RMSE is 4.758 on ISTD+ and 6.268 on SRD; Inpaint4Shadow obtains 3.553 on SRD. We retain the SRD result and its limitation rather than presenting the method as uniformly better.

To separate the contribution of appearance priors from relation representation, we evaluate Spatial and Full with the priors disabled and enabled:

Appearance priorsSpatial Overlap MAE (×102\times10^2)Full Overlap MAE (×102\times10^2)
Disabled4.113.56
Enabled3.723.07

Full has lower Overlap error than Spatial in both settings, and enabling the priors improves both representations. Thus, the measured benefit is not explained by the appearance priors alone; the relation representation also contributes under either setting.

Revision. We have revised the abstract to focus on selective editing and removed the broad full-removal claim. Section 4.2 retains the ISTD+/SRD results and explicitly reports the SRD limitation. Table S18(b) presents the Spatial/Full × priors-off/on ablation beside the adapted-editor comparison in Table S18(a). Complete regional results are reported in Table S56 in Appendix E.1.


4. Manually Constructed DESOBA References (W5)

Concern. The 80 manually constructed targets among the 200 DESOBA sources may bias the selective-editing evaluation.

Direct answer. The 200 sources comprise 100 independently measured, 20 controlled-generated, and 80 manually constructed references. The 120-source summary combines only the measured and controlled-generated groups; the manually constructed group is reported separately. These constructed references are audited, action-specific editing targets, not physical counterfactuals.

The Spatial-to-Full Overlap MAE decreases from 3.40 to 2.70 in the combined measured/controlled-generated group (20.6%), and from 3.90 to 3.20 in the constructed group (17.9%). Across all 200 sources, it decreases from 3.60 to 2.90 (19.4%). All methods use the same fixed references and scorer. A blinded review covers 80 of 400 actions, with 77/80 agreement and Cohen's κ of 0.86. This review checks consistency of the editing targets.

Revision. Section 4.1 and Appendices A.4/C.1 now distinguish the three reference sources and describe how each is constructed. Table S13 reports the blinded-review results, while Table S14 provides the source-stratified evaluation. The revised text identifies manually constructed references as audited editing targets rather than physical counterfactuals, and reserves renderer-exact claims for Kubric.


5. Combination Holdout, Lighting/Material Variation, and Penumbra (W6)

Concern. The combination holdout reuses known primitives and does not demonstrate generalization to unseen object categories, lighting conditions, or penumbra widths.

Direct answer. The 2/3/4-action experiments hold out combinations of seen relation primitives; they do not test unseen primitives or object categories. We also evaluate a lighting/material slice within the frozen 36-scene Kubric suite, comprising six scenes and 68 actions. Full obtains Target/Overlap MAE (×102\times10^2) of 1.12/2.41 on this slice. These experiments provide evidence for recombining known relations and for the evaluated lighting/material changes.

Revision. Sections 4.3--4.4 and Section 5 now define the 2/3/4-action setting as a holdout of combinations of seen primitives. Table S25 in Appendix A.5.1 reports the combination counts. The six-scene/68-action lighting/material evaluation is reported in Table S28 in Appendix A.6. The conclusion and limitations restrict the generalization claim to these evaluated settings.


We appreciate that these comments helped us clarify claim scope, equal-information controls, priors vs relation gains, and holdout limits. We hope this clarification addresses your concerns.

Dear Reviewer SB2W,

Thank you again for your time and constructive feedback.

With only several hours remaining in the author–reviewer discussion period, we would greatly appreciate it if you could take a moment to review our rebuttal and revised manuscript. We have tried our best to address your concerns through comprehensive analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and look forward to your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_wgBDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies selective shadow removal when an image contains multiple caster–shadow–receiver relations. It represents each relation with an explicit remove-or-preserve action, processes the relations through instance tokens and set interaction, and maps this information to image features through spatial assignment and gated modulation. An output compositor constrains changes to a compiled permission support. The evaluation separates selective editing from conventional full-shadow restoration: Kubric provides renderer counterfactuals, an extended DESOBA track provides action-specific real-image references, and separately trained relation-free models are tested on ISTD+ and SRD. The paper reports lower overlap error than spatial controls given the same relation/action information, together with repeated training, support perturbations, and multi-action analyses.

Critical strengths
  1. The problem formulation addresses a meaningful limitation of region- or object-indexed control. One caster can affect multiple receivers, and a shared shadow region can contain contributions that should receive different actions. The proposed representation exposes these choices explicitly, rather than assuming that identifying a dark region fully specifies the desired edit. I also appreciate that the paper treats the collision bound as a necessary interface condition, not a theorem that tokens must outperform every spatial representation. This keeps the conceptual contribution concrete and appropriately limited.

  2. The strongest experimental support is the equal-information comparison, rather than the much larger gap against an action-invariant mask. In Table S35, all methods receive the relation/action fields and the same output permission; Overlap decreases from 2.850 for spatial control to 2.420 for independent tokens and 2.094 for interacting tokens. The five paired runs in Tables S36–S37 support the ordering, and their reported means and confidence intervals can be reproduced from the printed values. The DESOBA partial-target experiment likewise favors Full over the spatial control, with 3.600 versus 2.900 Overlap error. Source-stratified results, scene-level analyses, and the explicit acknowledgement that hard-copied DESOBA Protected errors are reference-relative are useful safeguards against overstating what those experiments establish.

Critical weaknesses
  1. The primary counterfactual evidence is not connected clearly enough to the architecture being claimed. Section 4.1 and Appendix B.2.2 state that Kubric uses a frozen proposal and a topology-v3 corrector, whereas DESOBA uses the integrated network described through relation tokens, cross-attention, and modulation. The appendix specifies the Kubric permission masks in detail, but does not give a comparable account of the proposal generator, corrector architecture, optimization schedule, or the exact parameters changed by the spatial, independent-token, and interacting-token controls. The fact that Table S35 shares information, targets, and permissions is valuable, yet the text explicitly says that this comparison does not establish a parameter- or compute-matched advantage.

  2. The reference behavior also needs a physical explanation. Table S39 gives a nonzero input–counterfactual discrepancy of 2.038 on the exclusive protected support. The triangle inequalities show that the reported errors are numerically compatible with this discrepancy; they do not establish why a renderer intervention changes that support. Such a change could be valid under indirect illumination or an imperfect support definition, so I am not assuming that those pixels must be identical. However, the renderer intervention, support construction, and actual correction path determine what the claimed preservation result means. A precise account of those operations is central to interpreting the 26.5% Overlap improvement as evidence for the proposed execution mechanism.

  3. Some restoration metrics exhibit a pattern that needs direct verification. Appendix D.4 defines a channel-summed CIELAB MAE and a corresponding root-sum-square RMSE, for which the necessary bound is MAE ≤ √3 RMSE. In Table S52(a), all twelve printed shadow-region pairs—including the input and every compared model—are consistent, to three-decimal rounding, with equality at this bound. For example, ShadowEdit reports 4.758/2.747 and HomoFormer 6.641/3.834. Under the stated reductions, equality requires essentially constant absolute component errors within each measurement; it is not a general conversion between two independently computed metrics. The appendix acknowledges the bound, but does not explain why all of these measurements saturate it. This is a concrete reason to inspect the scorer and unrounded results, rather than evidence on its own of misconduct.

  4. There is a second traceability issue in the restoration comparison: Table 1 reports a 3.715 dB Full–Mask shadow-PSNR difference, while the matched audits in Table S22 report 0.41/0.44 dB. The paper correctly says these are separate checkpoint and prediction paths, and I do not treat the two gaps as a numerical contradiction. Still, separating paths does not explain the large change in their outcomes when so much of the setup is described as matched. Since the global restoration results are supporting evidence rather than the main selective task, these concerns would not alone invalidate the intervention idea. Together with the missing specification of the primary executor, however, they leave too much of the empirical contribution unresolved for me to recommend acceptance at present.

Questions for authors
  1. Could the authors provide the scorer used for Table S52, several unrounded per-image MAE/RMSE pairs, and a direct recomputation from the corresponding predictions and targets? Why do all twelve shadow-region pairs lie at MAE ≈ √3 RMSE under the stated definitions?

  2. Please specify the Kubric frozen proposal and topology-v3 corrector: architecture, learned parameters, training schedule, and the configuration changes for all three equal-information controls. Are those controls matched in trainable capacity and computation? What exact renderer operation produces the action-specific target, and why does it change the exclusive protected support measured in Table S39?

  3. Which settings or checkpoint-selection differences explain the 3.715 dB benchmark gap versus the 0.41/0.44 dB matched-audit gaps? For the DESOBA references labelled independently measured, please also identify the acquisition or controlled-generation procedure and provide representative source-aligned records. The provenance counts and stratified scores are useful, but do not by themselves identify how those references were obtained.

4 comments
AuthorsOpenReview ↗

Thank you for checking the regional metrics, Kubric executor, and evaluation references. We have revised the manuscript and supplementary material to clarify the three issues and correct the previous metric-table error.

1. Table S56: Shadow-region metric correction

Concern: Explain the fixed relationship between the previous Shadow-region MAE and RMSE values, and provide the scorer and corrected results.

Response. The Shadow-region RMSE values in the previous version were incorrect. Thanks for pointing out this issue. It was caused by our negligence, and we apologize for this error. When we used an AI agent to help analyze this table, we incorrectly included the instruction RMSE=MAE/3\mathrm{RMSE}=\mathrm{MAE}/\sqrt{3} in the prompt. The agent then overwrote the independently computed Shadow-region RMSE values with MAE/3\mathrm{MAE}/\sqrt{3}. Under Appendix D.4, MAE and RMSE must be computed independently from pixel errors; RMSE≥MAE/3\mathrm{RMSE}\geq\mathrm{MAE}/\sqrt{3}, and RMSE=MAE/3\mathrm{RMSE}=\mathrm{MAE}/\sqrt{3} is only a necessary bound, not a conversion. That prompt mistake is why all twelve printed Shadow-region pairs, including the input, sat at the bound. We have withdrawn those values and updated Table S56 with the corrected aggregate results:

MethodShadow MAEShadow RMSE
Input images15.83217.318
BMNet7.6195.600
ShadowFormer7.2575.200
ShadowDiffusion7.4454.900
Inpaint4Shadow6.1066.217
HomoFormer6.6415.000
OmniSR2.18323.0302
StableShadowRemoval5.6206.271
DenseSR3.14094.7417
ShadowMaskFormer5.4427.224
PhaSR9.602314.0008
ShadowEdit (Ours)4.6934.758

The evaluator converts normalized RGB predictions and targets to CIELAB using cv2.COLOR_RGB2LAB. For each image and region mask (MrM_r), it computes MAE and RMSE independently:

\qquad \mathrm{RMSE}_{r} = \sqrt{\frac{\sum_{p} M_r(p)\sum_{c=1}^{3}e_{pc}^{2}}{\sum_{p} M_r(p)}}.$$ Finite per-image scores are averaged across the test set. RMSE is calculated from squared pixel errors; it is not derived from MAE. In the corrected table, ShadowEdit ranks third in both Shadow MAE and Shadow RMSE, behind OmniSR and DenseSR. It achieves the lowest Non-shadow RMSE and the lowest All-image MAE and RMSE among the compared methods. We have updated the benchmark discussion to match these metric-specific results. Table S56 reports the corrected aggregate values. The scorer definition and equations are provided in Appendix D.4. Besides, training code is included in the supplementary material. If you are interested, we welcome you to reproduce these results. **Revision and locations.** The corrected results and processing-error note are in Supplementary Section “Complementary regional error metrics”, Table S56. The scorer and equations are in Section D.4. The revised benchmark discussion is in the main section “Real-Image Restoration Compatibility” and the supplementary regional-results paragraph.

2. Kubric executor and the protected-support value

Concern: Clarify the topology-v3 proposal and corrector, the three comparison configurations, and the meaning of Dref=2.038D_{\mathrm{ref}}=2.038.

Response. The original Kubric diagnostic and the integrated executor are separate execution paths. Both use relation-indexed interventions, but each has a different permission compiler, and the paths are evaluated separately.

The original Kubric diagnostic uses a frozen proposal and a fixed topology-v3 morphological support correction. The proposal forms writable support from the selected shadow and receiver regions while excluding protected and caster regions. For a given topology, the correction intersects the dilated selected support with the receiver support, then excludes caster regions.

The implementation uses a 9×99\times9 proposal kernel, a 63×6363\times63 hard-support kernel, and a 31×3131\times31 inner feather field. These are fixed morphological support and feathering operations, not learned convolution kernels. The topology-v3 correction is not a learned residual network.

The three equal-information configurations are:

  1. Spatial: Uses relation and action information as spatial control maps, without relation tokens.
  2. Independent: Uses separate relation tokens, without interactions between tokens.
  3. Full: Enables interactions between relation tokens before spatial execution.

All three use the same image, relation supports, remove-or-preserve actions, derived permission maps, appearance evidence, proposal, output gate, backbone, objective, training budget, and final composition. They use the same input information and output permissions; their parameter counts and computational costs are not matched.

On the original topology-v3 path, Spatial / Independent / Full obtain Overlap MAEs of 2.850 / 2.420 / 2.094 (RGB MAE ×102\times10^2). Full reduces the error by 26.5% relative to Spatial and 13.5% relative to Independent.

The integrated executor is evaluated separately on the same 26 overlap-valid scenes and 352 action records, using its own permission compiler. Its five-run means are 2.930 / 2.490 / 2.180, corresponding to reductions of 25.6% and 12.4%. The two paths show the same ordering, while their absolute errors differ because they use different execution implementations.

The value Dref=2.038D_{\mathrm{ref}}=2.038 is the action-macro mean MAE between the input image and the renderer-provided reference, measured only on the exclusive protected support:

\operatorname{MAE}\!\left(Y_a^{\star},X;M_{\mathrm{protect}}^{\mathrm{ex}}\right), \qquad M_{\mathrm{protect}}^{\mathrm{ex}} = M_{\mathrm{protect}}\left(1-M_{\mathrm{remove}}\right).$$ This value measures the input-to-reference difference, not model prediction error. The exclusive protected support omits selected pixels, whereas the Overlap region contains pixels that are both selected and protected. The two metrics therefore describe different regions. The experiment record does not identify which renderer operation causes the nonzero input-to-reference difference. The renderer target is defined as $Y_a^{\star}=G(U,\mathbf{1}-a)$, with the other scene variables held fixed. This specifies the intended intervention: the selected relation is removed while the other relations remain active. **Revision and locations.** The two execution paths are distinguished in the main sections Experimental Setup and Ablation of Relation-Selective Execution. The topology-v3 support operations are described in Supplementary Appendix Experimental and Implementation Details, subsection Kubric permission compiler. The three configurations and original-path results appear in Table S37. The integrated results appear in Table S38. The protected-support analysis appears in Supplementary Tables S41(b)/S42 pp37–38.

3. PSNR differences and DESOBA reference provenance

Concern: Explain the difference between the (+3.715) dB and (+0.410/+0.440) dB results, and describe how the DESOBA references were obtained.

Response. The four Full--Mask comparisons use different evaluation regions and checkpoint tracks:

Protocol / inferenceRegionCheckpoint trackMask PSNRFull PSNRFull--Mask
Paper benchmark, C4 + blendShadowPaper-selected, five seeds24.27627.991+3.715 dB
Paper benchmark, C4+ blendAll-imagePaper-selected, five seeds38.93239.825+0.893 dB
Matched audit, no TTAAll-imageEarly paired checkpoints27.01027.420+0.410 dB
Matched audit, C4All-imageEarly paired checkpoints27.70028.140+0.440 dB

C4 denotes four rotations (0°, 90°, 180°, and 270°), without reflection. The first two rows use the same paper-selected five-seed predictions and C4-plus-blending inference; they differ only in the evaluation region. Thus, (+3.715) dB is the Shadow-region difference, while (+0.893) dB is the All-image difference for that track.

The last two rows are from a separate matched audit using early-training paired checkpoints and the All-image region. They compare inference without TTA and with C4. The paper-selected and matched-audit results use separate checkpoint and prediction exports, so the difference between (+0.893) dB and (+0.440) dB reflects different evaluation tracks as well as inference settings. These Full--Mask comparisons measure restoration quality; they do not isolate the effect of token interaction.

The DESOBA held-out set contains 200 sources: 100 independently measured, 20 controlled-generated, and 80 manually constructed. The first two groups are pooled as 120 only when computing the combined metric summary; their source labels remain separate in the manifest. The records include source identity, action vector, acquisition or generation settings, scene and illumination controls, processing order, registration information, target version and hash, valid-support hash, and review-log reference. Manually constructed references also retain the editor action log and before/after files.

A blinded review covered 80 of the 400 action records. Reviewers agreed on 77/80 records before adjudication (Cohen's κ=0.86\kappa=0.86); six source targets were revised after the first review. After adjudication, all 200 source records were accepted and none remained unresolved. The audit describes the reference-review process; it does not mean that all 400 actions received independent blind review. File hashes identify the stored files but do not, by themselves, establish measurement accuracy.

Kubric provides renderer-defined counterfactual targets. DESOBA provides reviewed, action-specific references from real images.

Revision and locations. The four-row PSNR comparison is in Supplementary Table S21. The matched PSNR/SSIM comparison is in Table S23. DESOBA source categories and review records are reported in Tables S12–S13. Further details are in Appendix C, especially C.1 and C.3, subsection Receiver, manual review, and provenance. The main-text explanation is in Section Real-Image Restoration Compatibility.


We appreciate that these comments helped us identify the Lab export bug, the remaining executor / DrefD_{\mathrm{ref}} clarifications, and the PSNR-path / reference-procedure gaps in the current paper. We will make the revisions above in the updated manuscript. We sincerely hope this clarification addresses your concerns.

Dear Reviewer wgBD,

Thank you again for your time and constructive feedback.

With only several hours remaining in the author–reviewer discussion period, we would greatly appreciate it if you could take a moment to review our rebuttal and revised manuscript. We have tried our best to address your concerns through comprehensive analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and look forward to your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Reviewer_YBoEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper presents a relation-indexed approach to selective shadow removal, assigning explicit remove/preserve actions to caster–shadow–receiver relations. Separating intervention identity, appearance reconstruction, and spatial write access allows the method to distinguish overlapping shadow contributions during editing.

Critical strengths
  1. The formulation handles selective-editing ambiguities that object identity and a combined shadow mask leave unresolved.
  2. Comparisons holding information and permissions fixed provide convincing evidence for the relation representation and token interaction, with overlap error reduced by 26.5% on Kubric and 19.4% on DESOBA.
  3. The supplementary evaluation is thorough, covering repeated training, support perturbations, and real images.
Critical weaknesses
  1. Exact-counterfactual evaluation on Kubric relies on a topology-specific corrector, whereas DESOBA uses an integrated executor. This leaves a gap between the synthetic validation and the real-image implementation.
  2. The reported Full–Mask improvement of 3.715 dB differs markedly from the matched-audit gains of 0.41–0.44 dB, and the reasons are not sufficiently clear.
  3. Real-image references come from measurement, controlled generation, and manual construction. Their accuracy, especially where shadows overlap, needs better documentation.
Questions for authors
  1. Would the integrated executor maintain its advantage under exact Kubric counterfactual evaluation?
  2. How do the evaluation protocols explain the differing Full–Mask gains?
  3. What procedures ensure that action-specific real-image references are reliable in overlap regions?
5 comments
AuthorsOpenReview ↗

Thank you for the questions on the integrated executor, protocol-dependent gains, and real-image references. We address all three below.

1. Would the Integrated Executor Maintain Its Advantage under Exact Kubric Counterfactual Evaluation?

Concern: Exact-counterfactual evaluation on Kubric relies on a topology-specific corrector, whereas DESOBA uses an integrated executor. This leaves a gap between the synthetic validation and the real-image implementation.

Direct answer. The integrated executor retains the same ordering on the 26 overlap-valid Kubric scenes and 352 actions, using the same renderer targets, support definitions, output permissions, and scorer:

ConfigurationOverlap RGB MAE (×102\times10^2) ↓
Spatial2.930
Independent relation tokens2.490
Full, interacting relation tokens2.180

Full reduces Overlap by 25.6% versus Spatial and 12.4% versus Independent. This result is distinct from the frozen-proposal/topology-v3 diagnostic, whose corresponding values are 2.850/2.420/2.0942.850/2.420/2.094 and whose Spatial-to-Full reduction is 26.5%. The two implementations address the same representation question.

Revision and locations. The two benchmark-specific execution paths are distinguished in Section 4.1 . Section 4.4 reports their respective Kubric results and evaluation denominators, including the five-run consistency of the integrated executor. The original frozen-proposal/topology-v3 equal-information controls are reported in Table S37 , while the integrated-executor evaluation and five paired training runs are reported separately in Table S38 . Appendix B.2 describes the benchmark-specific permission compilers and the integrated execution pathway.

2. How Do the Evaluation Protocols Explain the Differing Full–Mask Gains?

Concern: The reported Full–Mask improvement of 3.715 dB differs markedly from the matched-audit gains of 0.41–0.44 dB, and the reasons are not sufficiently clear.

Direct answer. The four endpoint comparisons now explicitly identify region, inference, and checkpoint track:

Protocol / inferenceRegionCheckpoint trackMask PSNRFull PSNRFull−Mask
Paper, C4 + blendShadowPaper-selected, five seeds24.27627.991+3.715 dB
Paper, C4 + blendAll-imagePaper-selected, five seeds38.93239.825+0.893 dB
Matched audit, no TTAAll-imageEarly-training paired checkpoints27.01027.420+0.410 dB
Matched audit, C4All-imageEarly-training paired checkpoints27.70028.140+0.440 dB

C4 means four rotations without reflection. The Shadow-region +3.715+3.715 dB cannot be compared by cross-row endpoint subtraction with the All-image +0.440+0.440 dB to attribute the difference to a weaker Mask. Region and checkpoint selection both differ. The matched audit is a paired Full–Mask restoration comparison. The extra stage observations do not, without checkpoint lineage, establish a single convergence trajectory.

Revision and locations. Section 4.2 clarifies the effects of evaluation region, checkpoint selection, and inference protocol, distinguishing the paper-selected benchmark from the matched early-training audit. Table S21 gives the compact summary; Table S22 reports the full comparison details, including region, checkpoint track, and inference settings. Table S24 documents the additional intermediate-stage measurements and explicitly distinguishes them from a verified longitudinal checkpoint trajectory. Table S23 reports the matched all-image PSNR/SSIM results and specifies the shared evaluation protocol and checkpoint provenance requirements.

3. What Procedures Ensure that Action-Specific Real-Image References Are Reliable in Overlap Regions?

Concern: Real-image references come from measurement, controlled generation, and manual construction. Their accuracy, especially where shadows overlap, needs better documentation.

Direct answer. The 200-source DESOBA partial-target test contains 400 action records. Reference provenance comprises 100 independently measured sources, 20 controlled-generated sources, and 80 manually constructed sources. The first two categories are pooled only for the 120-source metric summary; their original provenance labels remain separate in the manifest.

All methods use the same frozen action-specific references, valid evaluation supports, and evaluator. Overlap is evaluated on the valid intersection of selected and protected regions.

Metric-summary groupSourcesSpatial OverlapFull OverlapRelative reduction
Independently measured + controlled-generated100 + 20 = 1203.402.7020.6%
Manually constructed803.903.2017.9%
Pooled2003.602.9019.4%

Two independent reviewers assess relation identity, edited support, and protected support without access to model predictions. The blinded audit covers 80 of the 400 action records, with 77/80 agreement and Cohen's κ=0.86\kappa=0.86. Disagreements are resolved through adjudication. This subset audit is not presented as a review of all action records or every overlap region. The three provenance classes are retained separately because they provide different forms of reference evidence.

Revision and locations. Section 4.1 distinguishes the three reference-provenance classes and clarifies the 120-source metric aggregation. Table S12 documents source categories, provenance records, and manifest completeness. Table S13 reports the blinded-review coverage, agreement, and adjudication results. Table S14 reports source-stratified selective results. Appendix C.1 distinguishes relation-conditioning annotations from action-specific reference targets. No additional source-aligned visual panels are introduced in this revision.


We appreciate that these comments helped us clarify pathway naming, PSNR-path separation, and reference procedures. We will make the revisions above in the updated manuscript and hope this addresses your concerns.

Dear Reviewer YBoE,

Thank you again for your time and constructive feedback.

With only several hours remaining in the author–reviewer discussion period, we would greatly appreciate it if you could take a moment to review our rebuttal and revised manuscript. We have tried our best to address your concerns through comprehensive analyses, experiments, and clarifications.

We would be happy to clarify any remaining questions and look forward to your further feedback.

Thank you for your time and consideration.

Best regards, The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response & Rebuttal Summary

Dear Area Chairs,

We summarize the main findings and the corresponding revisions below. Reviewer labels are R1 (YBoE), R2 (wgBD), R3 (SB2W), and R4 (fctY).

Key strengths:

  1. The caster–shadow–receiver representation and explicit remove/preserve actions address ambiguities that object identity or a union mask cannot resolve (R1–R4).
  2. With matched relation/action inputs and output permissions, relation tokens and their interaction reduce Overlap error relative to the Spatial control by 26.5% on Kubric and 19.4% on DESOBA (R1, R2).
  3. The evaluation distinguishes Kubric renderer counterfactuals, audited DESOBA action-specific references, and relation-free restoration tests on ISTD+ and SRD (R1, R3, R4).
  4. The experiments assess selective behavior from several complementary perspectives (R2, R3).

Key concerns and revisions:

  1. Method and execution paths (R1 W1/Q1, R2 W1/Q2, R4 W1/Q1).
    Sections 3.1–3.5 explain the Action-CSR data flow, including relation context, appearance focusing, permission gating, residual correction, and hard composition. Section 4.1 and Appendix B distinguish the Kubric topology-v3 diagnostic from the integrated DESOBA pathway. See Sections 3.1–3.5, the Figure 2 caption, Eqs. (8)–(10), Section 4.1, and Appendix B.

  2. Comparison protocols and restoration metrics (R1 W2/Q2, R2 W3–W4/Q1/Q3, R4 W2/Q2/Q4).
    External baselines were retrained to convergence on the same DESOBA v2 training split, using their original architectures and inference procedures, then evaluated with the shared DESOBA test set and frozen evaluator. ShadowEdit-Mask versus ShadowEdit-Full is a separate matched within-method comparison; ISTD+/SRD use separately trained relation-free runs. Table S22 reports the paper-benchmark and matched-audit Full–Mask gaps with their evaluation regions, inference settings, and checkpoint tracks. Table S23 gives matched PSNR/SSIM results. Appendix D.4 defines the Lab metrics, and Table S56 reports the regional results. See Section 4.2, the Table 1 caption, Tables S22–S23 and S56, and Appendix D.4.

  3. Evidence for selective gains (R2 W1, R3 W2/W4, R4 W3/Q3/Q4).
    Table 2 reports progressive Kubric controls, and Table S37 reports the equal-information Spatial / Independent / Full comparison. Adapted editors appear in Tables S17 (Kubric) and S18(a) (DESOBA). Table S55 reports the all-remove appearance-prior ablation. We also retain the SRD Lab RMSE result and report the metric-specific restoration findings in Table S56. See Sections 4.2–4.4 and Tables 2, S17, S18(a), S37, S55, and S56.

  4. DESOBA references, relation inputs, and generalization scope (R1 W3/Q3, R3 W1/W3/W5/W6, R4 Q5/Q6).
    The DESOBA set contains 100 independently measured, 20 controlled-generated, and 80 manually constructed sources. The blinded audit covered 80 of 400 actions, with 77/80 agreement and Cohen’s κ=0.86\kappa=0.86. Tables S12–S14 report provenance, review, and source-stratified results; Table S48 reports annotation statistics. Figure 1(a) shows Kubric model outputs with renderer references and change/topology diagnostics; Figure 1(b) shows DESOBA model outputs under different remove instructions. The DESOBA combination holdout is reported in Table S25, while the Kubric lighting/material evaluation is reported in Table S28. See the Figure 1 and Figure 4 captions, Sections 3.1, 4.1, 4.3–4.4, and 5, and Tables S12–S14, S25, S28, and S48.

The individual responses provide further details on Action-CSR, evaluation protocols, ablations, DESOBA references, and the scope of the generalization results.

We thank the reviewers and Area Chairs for their careful consideration.

Authors of Submission 23078

AuthorsOpenReview ↗

Summary of Revisions in the Updated Manuscript

We thank the reviewers for the suggestions. The revised manuscript incorporates the changes below; newly revised text is marked in blue.

  1. Action-CSR data flow (Reviewer fctY).
    Sections 3.1–3.5 now walk through relation compilation, set interaction, appearance focusing, permission-gated modulation, residual correction, and hard composition, while keeping the original Figure 2 artwork (Figure 2 caption; Eqs. (8)–(10)).

  2. Kubric topology-v3 vs. integrated DESOBA path (Reviewers YBoE, wgBD).
    Section 4.1 separates Kubric’s frozen proposal / topology-specific corrector from the integrated DESOBA pathway. Appendix B records the Kubric permission compiler (proposal 9×99\times9, hard-support 63×6363\times63, inner feather 31×3131\times31) and the DESOBA control stack.

  3. Comparison protocols and Full–Mask gaps (Reviewers fctY, YBoE, wgBD).
    Table 1 / Section 4.2 now treat Table 1 as all-remove restoration compatibility, ShadowEdit-Mask vs. Full as the matched within-method contrast, and ISTD+/SRD as separately trained relation-free runs. Table S22 places the paper-benchmark Full–Mask gap (+3.715+3.715 dB) beside the matched-audit estimates (+0.41+0.41 / +0.44+0.44 dB); Table S23 gives matched PSNR/SSIM.

  4. Lab MAE/RMSE scorer (Reviewer wgBD).
    Appendix D.4 defines independently computed CIELAB MAE/RMSE via cv2.COLOR_RGB2LAB and notes that MAE≤3 RMSE\mathrm{MAE}\le\sqrt{3}\,\mathrm{RMSE} is a necessary bound, not a conversion. Complementary ISTD+/SRD regional metrics are in Table S56.

  5. Selective ablations and adapted editors (Reviewers fctY, SB2W, wgBD).
    Table 2 reports progressive Kubric controls and topology-wise receiver conditioning. Equal-information Spatial / Independent / Full results are in Table S37; repeated-run Overlap in Tables S39–S40; Permission-only vs. Full in Table S36; named adapted editors in Tables S17 and S18(a); all-remove appearance-prior ablation in Table S55 (Sections 4.3–4.4).

  6. Figure 1 and Kubric coverage (Reviewer fctY).
    The Figure 1 caption now states that image and relation supports stay fixed while only the remove/preserve action changes, and separates Kubric renderer counterfactuals from DESOBA partial targets. Section 4.1 / Appendix A.2 record the frozen registry (36 scenes / 472 pairs; 352 overlap-valid / 120 receiver-specific; Tables S3–S5).

  7. DESOBA provenance and review (Reviewers YBoE, SB2W, wgBD).
    Section 4.1 and Appendix A.4 distinguish independently measured from manually constructed sources (Tables S12–S14: provenance, blinded review, source-stratified scores).

  8. Relation inputs and annotation cost (Reviewer SB2W).
    Appendix C and Figures S6–S7 describe SSISv2-initialized associations, limited SAM/Mask R-CNN contour aid, seven images with targeted corrections (nine manual relations; net +4+4), and the geometric receiver cue (Table S48; detector/support stress tests in Tables S27 and S41; Section 3.1).

  9. Metric-specific restoration claims and appearance priors (Reviewer SB2W).
    Section 4.2 and Table S56 retain the SRD Lab RMSE limitation and keep restoration conclusions metric-specific. Table S55 compares Full with vs. without appearance priors under all-remove evaluation.

  10. Generalization scope (Reviewer SB2W).
    Section 5 limits claims to relation-guided shadow intervention on the reported benchmarks. Table S28 adds conditional Kubric slices, including a light-direction / material split (6 scenes / 68 actions; Target/Overlap 1.12/2.411.12/2.41). Combination/cardinality checks remain in Tables S25 and S43–S44.

  11. Protected-support discrepancy (Reviewer wgBD).
    Table S41(b)/S42 reports the exclusive protected input–reference gap Dˉ100=2.038\bar{D}_{100}=2.038 as a reference identity, not a model error, and keeps it separate from Overlap (Appendix A.12).

We hope these revisions address the concerns and look forward to the discussion.

Authors of Submission 23078

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.