- 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
-
Composing existing 3D assets while preserving their original properties is a practical and underexplored problem.
-
Using cross-view flow disagreement to detect unresolved regions is intuitive and naturally motivates adaptive view selection.
-
The pipeline requires no manual 3D masks.
Critical weaknesses
-
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.
-
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.
-
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.
-
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.
-
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
-
Do the evaluation cameras (0°, 120°, 270°) overlap with the conditioning views? Do baselines receive targets for the same cameras?
-
Can the authors report source preservation against the original assets, e.g., masked metrics outside edited regions?
-
How do baselines perform with the same number of views and refinement rounds?
-
How does the method compare with Interact3D?
3 comments
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, , 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.
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:
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:
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.