acceptodds
Withdrawn from ICLR 2027

Route, Diagnose, Refine: Disagreement-Aware Multi-View Editing for 3D Asset Composition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Composing existing 3D assets from language requires realizing the requested interaction while preserving source geometry, appearance, and identity wherever changes are unnecessary. Existing 3D editors typically rely on a single or fixed set of views, leaving contact, occluded, and unseen regions unresolved. We introduce an active multi-view framework that casts 3D composition as a closed-loop route-diagnose-refine process. Our key observation is that view-conditioned source-target differential velocities share the same native 3D action space, enabling both editing and conflict diagnosis. Geometry-derived support and priority-aware ownership route each view's flow to reliable regions, while cross-view disagreement localizes unresolved edits and a three-state memory preserves resolved content. The system then acquires an informative new view and refines only where needed. Our end-to-end pipeline requires no user-specified 3D edit mask. Our method improves CLIP-T by 10.5% and DINO-I by 11.4% on our custom benchmark, while reducing LPIPS by 41.5% relative to Easy3E.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Asset Composition, 3D Editing, Multi-view Editing, Structured 3D Latent, Differential Flow, Active View Selection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RYEtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses language-guided composition of existing 3D assets. An LLM planner first generates a coarse composition, which is then refined through a closed-loop pipeline built on Easy3E. The method routes view-conditioned differential flows to supported regions, identifies unresolved voxels through cross-view disagreement, and acquires additional views for refinement.

Critical strengths
  1. Composing existing 3D assets while preserving their original properties is a practical and underexplored problem.

  2. Using cross-view flow disagreement to detect unresolved regions is intuitive and naturally motivates adaptive view selection.

  3. The pipeline requires no manual 3D masks.

Critical weaknesses
  1. The evaluation is partly circular. DINO-I, LPIPS, PSNR, and SSIM in Table 1 use Step1X-Edit targets as references, even though these targets also guide the proposed method. The method additionally acquires new views over 3.4 rounds, while baselines use fixed views. Thus, these metrics may reward agreement with the method's own guidance rather than actual composition quality. Despite acknowledging this limitation in Appendix B.3, the paper provides no source-preservation metrics against the original assets.

  2. The gains are not clearly attributable to the proposed components. In Table 7, uniform fusion already achieves CLIP-T 0.312 and DINO-I 0.813, versus 0.287 and 0.769 for Easy3E. This suggests that much of the improvement comes from the overall pipeline rather than the proposed routing, diagnosis, and view selection. The comparison is also not compute-matched: baselines run once, while the method averages 3.4 rounds at 79 s each, plus additional Step1X-Edit calls.

  3. Key baselines are missing. Interact3D directly addresses compositional generation of interacting objects, while Vinedresser3D and EditFlow3D explore iterative or agentic 3D editing. All are cited but not compared.

  4. The ablation evidence is weak. Q1 is mainly qualitative, with only one set of appendix results and no variance. Q2 uses Claude Sonnet 4.5 references derived from the same Step1X-Edit targets, while the visibility baseline achieves chance-level AUROC (0.49). Q3 relies entirely on GPT-4 ratings, and Table 4 appears to omit the "Ours" entry in the highest-rated category. Table 5 provides no baseline.

  5. The method is heuristic and difficult to reproduce. Eqs. 4, 7, and 12–17 are hand-designed without comparisons to simpler alternatives. Key hyperparameters (γ, ε, μ, η, κ, q, ρ, δ_stop, R_max) and sensitivity analyses are missing. The quality of the LLM planner's initial composition is also not evaluated, despite its importance to downstream refinement.

Questions for authors
  1. Do the evaluation cameras (0°, 120°, 270°) overlap with the conditioning views? Do baselines receive targets for the same cameras?

  2. Can the authors report source preservation against the original assets, e.g., masked metrics outside edited regions?

  3. How do baselines perform with the same number of views and refinement rounds?

  4. How does the method compare with Interact3D?

3 comments
AuthorsOpenReview ↗

Official Reply for RYEt regarding W1, W2, W3, Q1, Q2, Q3, Q4

Thank you so much for the detailed feedback, especially the concerns about evaluation bias, source preservation, and computational cost. We have added results and clarified the scope of our claims below.

W1/Q1: Evaluation and conditioning views.

Yes, the original evaluation cameras at 0, 120, and 270 degrees overlap with the initial target-view set. Fixed references avoid introducing additional ambiguity from synthesizing new target images: different image edits can realize “holding,” for example, with different grasps. However, we agree that this creates conditioning-view bias. These metrics measure fidelity to generated references and do not independently establish 3D correctness.

Single-view image-conditioned baselines receive guidance from the primary, most informative view; text-conditioned methods receive the editing instruction. Outputs are evaluated using common cameras and references.

We additionally evaluated renderings at 45, 165, and 315 degrees:

Method / evaluation camerasCLIP-T ↑DINO-I ↑LPIPS ↓
Ours / 0, 120, 270 degrees0.3170.8570.159
Ours / 45, 165, 315 degrees0.3110.8390.170
Interact3D / 45, 165, 315 degrees0.2910.7130.252

Our performance remains close to the original-camera results, and Ours outperforms the adapted Interact3D implementation under the additional-view evaluation. This broadens the camera coverage, but does not eliminate the limitations of generated image references.

Q2: Direct source preservation.

Table 5 originally evaluates preservation-mask precision, recall, and F1. We agree that these mask statistics do not directly measure preservation of the output content. We therefore added a source-reference evaluation, comparing output and source renderings within the same manually annotated preservation regions for all methods:

MethodPSNR-out ↑MAE-out ↓CLIP-T ↑
Easy3E25.30.0660.291
Uniform fusion28.50.0520.305
Ours32.80.0370.309

PSNR-out and MAE-out use the source asset rather than Step1X-Edit targets as the reference, with RGB normalized to [0,1]. CLIP-T is reported alongside them to measure instruction alignment, rather than source preservation. Relative to uniform fusion, Ours improves preservation PSNR by 4.3 dB and reduces MAE-out by 28.8%, while retaining higher instruction alignment. This directly supports preservation of source appearance outside the intended edits. These are rendered-appearance measurements, rather than direct surface-geometry metrics.

W2: Why are the routing gains modest in CLIP-T?

Routing primarily improves small interaction regions, such as contact points or partially occluded attachments. Global text-image similarity can remain similar even when a local contact is distorted, since the overall objects and requested relation are still recognizable. A small CLIP-T difference therefore need not imply a small local improvement.

Within the controlled Table 7 comparison, DINO-I increases from 0.813 to 0.857, average LPIPS decreases from 0.172 to 0.159, and worst-view LPIPS decreases from 0.234 to 0.172 (26.5%). These larger image-fidelity and worst-view gains support routing's role in resolving local cross-view conflicts. We agree that additional views and initialization also contribute to the overall pipeline gains, so Table 1 alone should not be used to attribute every improvement to routing.

Q3: Multiple refinement rounds and fairness.

The baseline configurations do not implement our specific stateful refinement controller: persistent 3D latents, changing spatial ownership, and three-state handling of completed and unresolved regions. Simply rerunning a baseline is not the same operation and would require an additional adaptation. Our technical contribution is this preservation-aware refinement process, rather than repeated execution alone.

We nevertheless agree that additional rounds and image-editing calls create a quality–cost trade-off, and the end-to-end comparison is not compute-matched. We have expanded the limitations discussion to acknowledge this explicitly. The reported 3.4 rounds at approximately 79 s per round correspond to about 269 s of refinement; the 79 s already includes target-image generation. We do not claim that the end-to-end quality advantage establishes greater efficiency at equal compute.

W3/Q4: Additional baselines.

We added Interact3D above. Nano Banana API calls had a high failure rate in our runs, so we used GPT-2-image for its image-generation/editing stage. The result is therefore an adapted implementation, not an exact reproduction of the original image-model configuration. Three Vinedresser3D comparisons are at https://rebuttal2.netlify.app/. EditFlow3D remains a missing quantitative comparison, which we acknowledge.

AuthorsOpenReview ↗

Official Reply for RYEt regarding W4, W5

Thank you also for the specific suggestions on ablation evidence and reproducibility. We have made the corresponding clarifications in blue.

W4: Ablation evidence and reference quality.

Routing: Fig. 5 and Table 7 provide qualitative and controlled quantitative evidence. Sec. 4.2 now explains why small interaction-region improvements can affect DINO-I and worst-view fidelity more than global CLIP-T. App. F.5/Fig. 10 adds a representative later-round priority comparison with targets, three-view outputs, and matched earcup details. We do not treat this visualization as a substitute for a quantitative earlier-view-first ablation.

Diagnosis: Table 3 uses Claude Sonnet 4.5 labels derived from Step1X-Edit target comparisons. These are shared proxy references, not independent geometric ground truth; Sec. 4.2 now says this explicitly. Visibility-only AUROC 0.49 means that visibility alone ranked those proxy residuals poorly, not that it independently validates disagreement.

App. F.3/Table 10/Fig. 7 now reports two completed visibility experiments. The case-level analysis uses 50 manually screened correct reference edits, with consistent 3D-rendered source/target views and the native state fixed to the correct edited latent. Mean disagreement normalized by the conflict threshold, D/τCD/\tau_C, is 0.71 for both-visible regions, 0.69 for one-view regions with lower relative prediction error, and 0.93 for those with higher relative error. One denotes the conflict threshold, not the both-visible mean. Thus, reduced visibility alone produces little change in the lower-relative-error group, while inaccurate hidden-region predictions are associated with greater disagreement.

The completed false-alarm experiment compares correct regions visible in one view with those visible in at least two views, using paired, case-level false-positive rates. The measured rates are of comparable magnitude, with no pronounced increase under single-view visibility. Together, these findings support the diagnosis under partial visibility. They do not imply zero false positives or equivalence to a support-restricted score; inaccurate unseen-view predictions can still cause false conflicts, as discussed in Sec. 5. Both measured plots are in App. F.3/Fig. 7 and at https://rebuttal1.netlify.app/ (Section 01).

View selection: the existing Table 4 reports Ours as highest-rated in 135 cases, with mean 8.1. Sec. 4.2 now explicitly distinguishes these GPT-4 scores from human judgments. Table 5 describes our progression across rounds, not a matched-round baseline study. Secs. 4.2/5 now acknowledge non-monotonic average SSIM and the unmatched computational budget.

W5: Parameters and initialization.

The planner provides a coarse arrangement and initial edit region. Controlled routing comparisons use the same coarse composition, targets, masks, seeds, denoising schedule, and total model calls, so planner differences do not explain their routing-policy differences. This does not show that initialization is unimportant: a missed contact, incorrect pose, or severe interpenetration can still prevent successful refinement. Sec. 5 now states this dependency.

We clarified the contribution in Sec. 3 and the three-state preservation rationale in Sec. 3.3: supported multi-view velocity routing and disagreement-aware state continuation, distinct from pretrained initialization modules and generic iterative editing.

We agree that these implementation details are important for reproducibility. We will document the exact hyperparameter settings and full pipeline prompts in the revised supplementary material, and release the corresponding configuration files together with our code on GitHub.

Thank you for your detailed response. However, I still have concerns regarding the evaluation and ablation studies presented in the paper. Therefore, I am inclined to maintain my original score.

Reviewer_6HYBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work introduced an active multi-view framework that casts 3D composition as a closed loop route-diagnose-refine process. The proposed method works by first taking language instruction and input assets as input, and then used a priority-aware differential-flow routing technique to assign each view's flow to its supported 3D regions and resolve regions from cross-view disagreement until stopping criterion is met.

Critical strengths
  1. This paper was supported by illustrative figures throughout the main text, which is very helpful in demonstrating both their methodology and experimental results.
  2. This paper studies a well-motivated problem and area of 3D composition, which has good application value in industry.
  3. The paper seems to have relatively good performance in comparison to other baseline methods.
Critical weaknesses
  1. My main concern lies within dataset benchmark selection. It is discussed in the experimental section that 194 cases built from two datasets are considered in the evaluation suite. However, one of these datasets were from 2021 and the other is from 2024, which makes at least the 2021 one relatively outdated in this fast-paced space of research. Furthermore, 194 cases is not enough data to prove performance. Newer work such as The 3DCoMPaT200 dataset [1] in 2025 has comprises 19k 3D shapes from 200 object categories, just one example.
  2. My second concern lies in the human evaluation trail conducted in this paper across 130 participants. It was not discussed in great detail of selection process, anonymization process, and such.
  3. The technical novelty of the differential-flow routing module is questionable, as it highly resembles the structure in Easy3E.

References:

  1. 3DCoMPaT200: Language-Grounded Compositional Understanding of Parts and Materials of 3D Shapes
Questions for authors
  1. For the human evaluation, could you give more details as in how these participants were selected? There is such a drastic difference between user preference for your method vs other methods that it would be helpful to get more details on the background of these participants.
  2. Does your method improve editing in other aspects such as robustness performance?
1 comment
AuthorsOpenReview ↗

Official Reply for 6HYB regarding W1, W2, W3, Q1, Q2

Thank you so much for your careful reading and for recognizing training-free editing and stateful refinement! Your comments helped clarify our contribution and limitations. We have added and revised our manuscript in blue font.

https://rebuttal2.netlify.app/

Our experimental cases and three-view renderings are collected at the link above; the explanations below refer to these materials.

W1: Benchmark suitability and coverage.

Our evaluation contains 194 interaction-composition tasks, rather than 194 independent shapes. Each task combines supplied assets with an instruction requiring meaningful contact, attachment, or deformation. For example, a basket becoming a balloon's gondola requires suspension-line adaptation and coherent attachment; a teddy bear holding a mug requires changes around the paws and handle. A large collection of individual shapes is valuable, but does not directly supply these interaction tasks and evaluation references.

We agree that the benchmark has limited coverage. Following your suggestion, we constructed three additional cases using 3DCoMPaT200 assets: balloon–basket attachment, a teddy bear holding a mug, and a fox wearing pixel sunglasses. The instructions, input assets, and three-view comparisons with TRELLIS and Vinedresser3D are provided. (For details, please see the link at the beginning.)

The rebuttal period did not allow us to complete a full evaluation on 3DCoMPaT200. These examples provide qualitative evidence of transfer to another asset source, rather than an exhaustive benchmark or a general robustness guarantee. We will make this scope explicit. Dataset age alone is not a sufficient measure of task suitability, although broader and more recent asset coverage would strengthen the evaluation.

W2/Q1: Recruitment and human-evaluation protocol.

The questionnaire was circulated in company and university groups, and we collected 130 completed questionnaires. Each participant evaluated four cases randomly sampled from the pool of 194 benchmark tasks, giving 520 case-level preference judgments in total.

For each case, participants were shown the input assets and editing instruction together with anonymized renderings of the methods' GLB outputs. Method identities were hidden, and the result order was randomized. Participants selected the result they considered best to satisfy the composition instruction.

Human evaluation complements the automatic metrics because small contact defects, local distortions, or identity changes may not be adequately reflected in global image scores. The reported 59.4% is the share of valid preference judgments received by our method, not the fraction of participants who exclusively preferred it. Recruitment was a convenience sample, and we do not claim that the participants were representative of a broader population or uniformly experienced in 3D modeling. We will document the recruitment and presentation protocol more explicitly.

Q2: Robustness.

I have added some renderings on the link above, please feel free to check it!

The 3DCoMPaT200 cases examine application to a different asset source and interactions involving attachment, grasping, and wearing. Spatial locking also limits where target guidance can alter the 3D state, providing a safeguard for content outside the active interaction region. The displayed examples provide qualitative evidence of cross-dataset applicability; they do not establish robustness to arbitrary prompts, seeds, or inconsistent editable-region guidance.

W3: Novelty relative to Easy3E.

We acknowledge the inherited FlowEdit/Easy3E editing machinery. Our contribution is how multiple differential velocities are fused spatially and across refinement rounds: views receive local responsibility according to geometric support, conflicts identify where further editing is needed, and the native 3D state continues without discarding completed edits. Initialization and image generation are necessary supporting components, rather than our novelty claims.

The three states address an important ambiguity in iterative composition. A resolved region may contain an intentional completed edit, or it may still correspond to the source. Both should stop changing, but the former must retain its edited value while the latter retains its source anchor. A single unresolved/resolved label does not explicitly distinguish these roles. This is the rationale for the three-state design.

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

This paper addresses language-guided composition of existing 3D assets, where the requested interaction must be realized while preserving the source assets elsewhere. A GPT-based planner first arranges SuperDec superquadric decompositions of the assets into a coarse composition and an initial edit region, and Step1X-Edit produces edited target images for rendered views. Building on Easy3E's flow-based editing in TRELLIS latents, the method then runs a closed loop: each view's source–target differential velocity is routed to regions where that view has geometric support, with later-acquired views taking local priority; cross-view disagreement among these velocities marks unresolved voxels, separating them from source-stable and edited-stable ones; and a new camera that best exposes the unresolved region is added for the next round. On 194 compositions built from Mixamo characters and PartObjaverse-Tiny objects, the method achieves the best scores on all metrics in Table 1, receives 59.4% of preferences in a 130-participant user study, and is supported by ablations on routing, diagnosis and view selection.

Critical strengths
  1. The method is training-free and operates on frozen native-3D flow models, so the routing and diagnosis layer could in principle be reused with other flow-based 3D editors.
  2. The stateful continuation across rounds (Eq. 18) lets later rounds repair geometry through the sparse-structure latent without decoding and re-encoding the asset, which is a practical design for iterative refinement.
  3. The qualitative results in Fig. 4 show clear improvements at contact and occluded regions, and the blinded 130-participant user study prefers the method in 59.4% of judgments.
Critical weaknesses
  1. Limited technical novelty. The method relies heavily on existing components, including SuperDec, a GPT planner, Step1X-Edit, and Easy3E. In particular, the per-view editing operator (Eq. 3 with λ=1.0\lambda = 1.0), the two-stage sparse-structure/SLAT editing, and the source lock-back are all inherited from Easy3E. The main contribution seems to be extending Easy3E to multiple edited views, along with a few heuristic choices for view ownership (Eq. 7), disagreement estimation (Eq. 10), and view selection (Eq. 17). These ideas are reasonable, but individually they appear fairly straightforward. The ablation in Table 7 also raises questions about how much the proposed routing contributes: compared with uniform fusion of the same views, the gains are small in CLIP-T (+0.005) and average LPIPS (−0.013), and appear mainly in DINO-I and the worst-view metrics. It would also be useful to compare against Vinedresser3D (CVPR 2026), which already explores informative-view selection for native 3D editing.

  2. Unclear reliability of the disagreement signal. Since Step1X-Edit generates the target for each view independently, I am not convinced that the disagreement score in Eq. 10 necessarily indicates where the 3D edit is incomplete. It could simply be measuring inconsistencies between the generated 2D targets. This matters because the method responds to high disagreement by adding another edited view, and Eq. 7 gives the newly acquired view local priority. If that new target contains inconsistent or hallucinated details, it could actually make the 3D result worse. I would like to see some analysis of cross-view target consistency, or at least examples where this feedback loop fails. The results in Table 5 also suggest that more rounds are not always helpful: average SSIM drops from 0.903 to 0.891 between rounds 2 and 3, even though the other fidelity metrics improve. These issues deserve more discussion, especially since the paper does not include a limitations section.

Questions for authors
  1. The disagreement score (Eq. 10) averages differences across all views without the support weights used in routing. In regions that one view sees but another cannot, the unseen view's velocity is only the model's guess, so the score could be high even when the edit is already correct. Can you show that this does not happen, for example by computing the score only over views that see each region?
  2. Each view's target is generated separately by Step1X-Edit. How often are these targets inconsistent across views, and can you show failure cases where an inconsistent target leads to a wrong edit?
  3. Newly added views always take priority over earlier ones (Eq. 7). Does this ordering actually help? For example, how do results change if earlier views take priority instead, or if the order is ignored?
3 comments
AuthorsOpenReview ↗

Official Reply for pXco regarding W1 and Q1

Thank you so much for your careful reading and for recognizing training-free editing and stateful refinement! Your comments helped clarify our contribution and limitations. We have added and revised our manuscript in blue font.

https://rebuttal1.netlify.app/

Our experimental cases and three-view renderings are collected at the link above; the explanations below refer to these materials.

W1: Novelty beyond existing components.

The main challenge of our work is coordinating multiple view-conditioned edits throughout refinement. This is not a straightforward extension of single-view editing: views have different visibility and spatial applicability, and their proposed updates can conflict, so it is quite tricky how much each view should contribute to each voxel. Uniform averaging can dilute a decisive correction, while globally prioritizing one view can damage regions it poorly observes. Our geometry-supported ownership determines where each view should act, and later-view priority gives newly acquired evidence local precedence within its supported region.

Continued refinement introduces another difficulty: preserving completed edits while retaining unchanged source content. A resolved/unresolved threshold alone does not distinguish these preservation requirements. Our three-state diagnosis separates unresolved regions requiring further refinement, edited-stable regions retaining completed edits, and source-stable regions retaining their source anchors. Together with disagreement-driven view acquisition and persistent native-3D state, this forms our preservation-aware composition pipeline.

We do inherit the source lock-back mechanism from Easy3E. However, the two-stage editing procedure—sparse-structure editing followed by SLAT refinement—is commonly adopted in TRELLIS-based native-3D editing and is not specific to Easy3E. Our contributions, including priority-aware multi-view velocity routing, disagreement-based diagnosis, and three-state preservation across refinement rounds, are quite far from Easy3E. SuperDec, planning, and image editing modules primarily prepare the coarse composition and visual guidance, we do not claim these existing components as contributions.

Routing mainly affects small interaction and occluded regions, so global CLIP-T can be less sensitive to its benefits. Table 7 shows DINO-I increasing from 0.813 to 0.857, average LPIPS decreasing from 0.172 to 0.159, and worst-view LPIPS decreasing from 0.234 to 0.172 (26.5%). We clarify these contributions in Secs. 3/3.3/4.2 and provide three Vinedresser3D comparisons in App. F.6/Fig. 11. (For details, please see the link at the beginning.)

Q1: Can missing visibility cause false conflicts?

We completed two experiments (App. F.3/Table 10/Fig. 7). The disagreement analysis uses 50 manually selected correct reference edits, consistent 3D-rendered views, and the correct edited latent, separating visibility from target inconsistency.

Mean D/τCD/\tau_C is 0.71 for both-visible regions, 0.69 for one-view regions with lower relative prediction error, and 0.93 for those with higher relative error. One is the conflict threshold, not a reference-group mean. “Lower” compares unseen-view and observing-view errors; it does not certify absolute correctness.

The completed false-alarm experiment finds comparable paired, case-level false-positive rates for correct regions visible in one view versus at least two views, without a pronounced rise under single-view visibility. Together, these results support diagnosis under partial visibility, while inaccurate hidden-region predictions can still cause false conflicts. (For details, please see the link at the beginning.)

AuthorsOpenReview ↗

Official Reply for pXco regarding W2, Q2, Q3

Our experimental cases and three-view renderings of the 3D results are collected at the link below; the explanations below refer to these materials.

https://rebuttal1.netlify.app/

W2/Q2: Target inconsistency and failure cases.

(For details, please see the link at the beginning.)

We actually have two safeguards against this scenario.

Candidate targets are automatically screened by Gemini-3-Pro, without human screening, with an average of 2.3 rejected targets per case. This is a count, not an inconsistency percentage. Our original manuscript omitted this implementation detail; we will add the model and automatic screening procedure to Apps. B.2/F.4. This differs from manually selecting the 50 diagnostic cases in Q1. Section 02A compares retained sofa guidance with a small pillow discrepancy and rejected guidance with a larger conflict; deliberately using the latter produces a worse result.

Spatial locking provides a second safeguard. Target generation emphasizes the interaction region, while regions outside the next active mask retain their existing 3D values. Section 02B shows inconsistent teddy-bear leg stripes and the rider's back hood in the targets, while the locked output regions are retained. This protection depends on the masks: it cannot resolve incompatible guidance inside the editable interaction region.

Although screening and locking mitigate these failures, our original manuscript did not adequately discuss this limitation. We have added an explicit limitations discussion explaining that disagreement is a conflict signal, not a guaranteed detector of geometric error, and that inconsistent editable-region targets can degrade refinement. We also acknowledge the non-monotonic average SSIM in Table 5 (0.903 to 0.891 from rounds 2 to 3), despite improvements in PSNR, LPIPS, and worst-view SSIM.

Q3: Does later-view priority help?

(For details, please see the link at the beginning.)

We compared priority enabled/removed across multiple cases; the helmet–headphone example in Section 03 is a representative visualization. Without priority, earlier guidance can dilute a later correction in an overlapping interaction region. The displayed earcup deformation makes this effect visible, while priority preserves the correction more coherently. Together with Table 7, these experiments support giving newly acquired evidence local precedence. This priority is restricted to geometrically supported regions, rather than allowing the new view to overwrite the whole asset.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns.

The new experiments help alleviate my concerns regarding partial visibility and later-view priority. I also acknowledge the meaningful improvements in DINO-I and worst-view LPIPS.

However, my main concern regarding limited technical novelty remains. While the proposed components are practically useful, I am still not convinced that their combination constitutes a sufficiently substantial methodological contribution. Furthermore, the disagreement analysis relies on carefully selected correct edits, leaving its reliability under inconsistent targets in the actual iterative pipeline insufficiently validated.

Overall, the rebuttal strengthens the empirical evidence but does not fully resolve my primary concerns. Therefore, I will maintain my original rating of 2 (Weak Rejection).

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

Related papers

Discussion (0)

Sign in to comment.