acceptodds
Under review as a conference paper at ICLR 2027

Pace3D: Controlling Structural Progress in 3D Editing

17 views

Abstract

Training-free 3D editors typically stop at a fixed sampling endpoint, implicitly assuming elapsed time implies structural completion. This can fail in hierarchical 3D generators: sparse-structure latents evolve continuously, while decoded support changes discretely as occupancy logits cross thresholds. Sampling can therefore end with incomplete target support, unrecoverable by downstream feature generation. We call this failure *structural progress–clock mismatch*. **Pace3D** uses the structure decoder as an online progress sensor. A target-conditioned proxy provides a structural reference; when decoded structure lags behind its reference phase, Pace3D rewinds the clock while retaining the current latent. Source Alignment stabilizes the extended trajectory, while Cross-Condition Consensus preserves unchanged source features. Diagnostics show that larger terminal phase lag corresponds to more missing target structure. On Eval3DEdit, Pace3D further achieves the best target alignment while maintaining strong source fidelity.

Primary area:
generative models
Keywords:
3D editing, generative models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HZT1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Pace3D for 3D editing. Pace3D first embeds the current 3D object into 3D occupancy, which is labelled as "source". It also use image editing model (didn't tell the specific model used?) and image-to-3D model (TRELLIS-image-large) to get a proxy target occupancy. Then during the sparse structure editing, it continually judges how well is the editing path aligned with the pesudo GT path obtained from the source occupancy latent to the proxy target occupancy latent. If the structure editing falls behind, it will reset the sampling timestep. Such resets can happen multiple times (but less than or equal to 6 times). Overall, Pace3D use longer latency (than Nano3D) as trade-off for more accurate 3D editing results.

Critical strengths
  1. Figure 3 clearly shows the motivation of Pace3D: the edited structure can fall behind a GT editing path. I also think checking the structure rather than only checking the timestep is reasonable.

  2. Table 2 shows that the editing alignment is improved over related baselines.

  3. Table 6's equal-wall-clock comparison with Nano3D shows that the performance gains not 100% comes from longer computation time.

  4. Table 4's comparison with "Fixed" shows the benefit of Pace3D's timestep resetting operation.

Critical weaknesses
  1. The paper measures unfinished structure for baselines, but not for Pace3D. Section 5.1 counts the unfinished addisions and deletions for baselines, but same comparison after Pace3D is applied is missing. Table 3 and table 4 report the similarity of embeddings, which can not be interpreted as the quality of structure editing. The paper's claim is that baselines finish earlier then ready, while Pace3D makes sure finishing when ready via rewind. But these erperiments is not strong enough to support such claim. Also, the changes outside the editing parts should also be measured, other wise a method may finished with "over-edited" structures, including some changes that we do not want.

  2. Eval3DEdit has a total of 100 cases, but in table 8 (the proxy vs GT comparison), only 12 cases were used, making the relibility of the proxy unjustified. Also in paper's main experiments on Eval3DEdit. The authors did not make it clear that those experiments are conducted on all 100 cases or a selected subset from it.

  3. In table 3, the controller makes the PC-PC worse, from 0.7604 to 0.7383, though PC-Txt improves, making me more concerned about whether Pace3D has "over-editing" issues.

  4. In table 4, which is an important ablation, the paper does not make it clear how random and fixed variants are implemented.

Questions for authors

please see the weaknesses above

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/4]

Thank you for the constructive feedback. We have added geometric measurements and VLM assessments of edit completion and unintended changes to regions that should be preserved.


Image Editing Model Used in Eval3DEdit

The target images in Eval3DEdit were generated using Nano Banana, as described in Appendix B.1, Step 4 of AnchorFlow (Zhou et al., 2026). We use the target images provided by the benchmark without modification for the image-conditioned comparisons. These images provide the target condition for editing. We will clarify this distinction in the revised paper.


[W1] Evaluation of Structural Completion and Over-Editing

Embedding similarity does not directly measure structural completion. Our additional geometric evaluation shows that Pace3D reduces both missing structure and unintended changes. On the 12 paired cases, Missing Structure decreases from 0.418 to 0.350 at 64364^3 (Table R1), while extra additions decrease from 53.9% to 28.1% and incorrect deletions from 30.7% to 15.9% (Table R3).

1. Results on the 12 Paired Cases

Table R1 reports the cumulative component ablation on the 12 paired cases used in Sec. 5.1. Alongside the metrics from Table 2, we evaluate Missing Structure (Appendix C, Eq. 48) and voxel IoU against the paired target. Both geometric metrics are computed from voxelized final meshes at 64364^3 and 2563256^3. Before voxelization, the source, target, and edited meshes share the same position, orientation, and scale. All methods use the same spatial extent and voxel grid at each resolution.

Table R1. Results on the 12 paired cases. Components are added cumulatively; the final row is the full Pace3D model.

VariantPC-PC ↑DINO-src ↑FID ↓PC-Img ↑PC-Txt ↑MV-Img ↑MV-Txt ↑DINO-tgt ↑Missing Structure ↓ (64364^3)IoU ↑ (64364^3)Missing Structure ↓ (2563256^3)IoU ↑ (2563256^3)
Baseline0.65310.7503131.420.39300.28360.73970.49960.74510.4180.6490.3080.595
+ Controller0.63590.7498138.200.39460.28880.74610.50180.74990.3750.6030.2810.527
+ Src. Align.0.68100.7567115.320.39580.28450.74480.50310.75740.3470.8130.1850.776
+ CCC (full Pace3D)0.68800.7579115.010.39960.28330.74740.50360.75790.3500.8130.1860.780

Compared with the baseline, full Pace3D reduces Missing Structure by 16.3% at 64364^3 and 39.6% at 2563256^3, with IoU gains of 0.164 and 0.185, respectively. These comparisons are made within each voxel resolution.

The controller alone reduces Missing Structure but also lowers IoU, from 0.649 to 0.603 at 64364^3 and from 0.595 to 0.527 at 2563256^3. Adding Source Alignment raises IoU to 0.813 and 0.776 while further reducing Missing Structure. This is consistent with Sec. 4.2: additional sampling gives unfinished edits more opportunity to complete, but also allows unintended changes to accumulate. Source Alignment complements the controller by limiting this geometric leakage.

2. Interpretation of the Metrics in Tables 3–4

Eval3DEdit does not provide paired target 3D ground truth, so we use embedding-based metrics to assess source fidelity and target alignment across the benchmark. PC-PC compares the source and edited point clouds in Uni3D space, while PC-Img and PC-Txt measure alignment with the target image and text. DINO-src and DINO-tgt compare rendered outputs with source and target images using DINOv2; MV-Img and MV-Txt measure multi-view alignment using CLIP (Appendix B.2). By measuring visual and semantic agreement with the target image and text, the target-alignment metrics provide indirect evidence of whether the requested edit has been realized. They do not directly quantify missing geometry or changes to regions that should be preserved, so we complement them with the paired geometric measurements in Tables R1 and R3.

AuthorsOpenReview ↗

Rebuttal by Authors [2/4]

3. Assessment of Overall Edit Completion

Across all 100 Eval3DEdit cases, Pace3D achieves a VLM-assessed overall edit completion rate of 71%, compared with 59% for Nano3D† and AnchorFlow. This evaluation includes both structural and appearance edits.

We use gpt-6-sol as the judge. Each case is presented with the editing instruction, source and target images, and six pairs of source and edited renderings from matching cameras spaced 60° apart. The judge identifies the requested change before returning a yes/no decision. A case is counted as complete only when all requested changes are fulfilled. For structural edits, added parts must appear as recognizable geometry at the intended locations, and parts requested for removal must be absent from the provided views. Recoloring alone does not count as completing a structural edit.

Each output is judged once, with uncertain cases reviewed manually. Method names are hidden, and all methods use the same prompt and rendering settings. Completion is assessed separately from visual quality. The full prompt is provided in Appendix H of the revised manuscript.

Table R2. VLM-assessed overall edit completion on all 100 Eval3DEdit cases.

MethodOverall edit completion (%) ↑
Edit3608
EditP2330
AnchorFlow59
3DEditFormer55
PartFlow38
Nano3D39
Nano3D†59
Pace3D (Ours)71

† denotes the SS-NFE-matched Nano3D variant in Table 2.

These judgments use evidence from the rendered views and complement the direct geometric evaluation on paired cases.

4. Evaluation of Over-Editing

Pace3D reduces both extra additions and incorrect deletions on the 12 paired cases (Table R3). On all 100 Eval3DEdit cases, its VLM-assessed over-editing rate is 13%, compared with 24% for Nano3D† (Table R4).

For the paired evaluation, we voxelize the final meshes on a common 64364^3 grid after aligning the source, target, and edited meshes in position, orientation, and scale. Extra additions count occupied output voxels where both the source and target are empty. Incorrect deletions count voxels occupied in both the source and target but missing from the output. Each count is divided by the source’s occupied voxel count. Table R3 reports the mean percentages over the 12 cases.

Table R3. Over-editing on the 12 paired cases, measured from voxelized final meshes.

MethodExtra additions (%) ↓Incorrect deletions (%) ↓
Baseline (FlowEditing only)53.930.7
Pace3D (Ours)28.115.9

For the benchmark-wide assessment, we use gpt-6-sol with the same cases, generated outputs, and rendered views as in Table R2. The judge identifies the requested edit and checks for major unrequested structural changes, including unintended removal or duplication of parts, collapsed geometry, and extra structures or fragments. Each output is judged once, and uncertain cases are reviewed manually. Method names are hidden, and all methods use the same prompt and rendering settings.

Table R4. VLM-assessed over-editing on all 100 Eval3DEdit cases.

MethodOver-editing rate (%) ↓
Edit36085
EditP2358
AnchorFlow55
3DEditFormer23
PartFlow18
Nano3D21
Nano3D†24
Pace3D (Ours)13

† denotes the SS-NFE-matched Nano3D variant in Table 2.

Under this evaluation protocol, Pace3D achieves the highest overall edit completion rate and the lowest judged over-editing rate among the compared methods.

AuthorsOpenReview ↗

Rebuttal by Authors [3/4]

[W2] Clarification of Data Sources

The main comparison in Table 2 uses all 100 Eval3DEdit cases. Each case provides a source 3D asset and a target image, but no paired target 3D ground truth. Table 7 also uses these 100 cases and compares using the proxy directly as the final sparse structure with using it as a progress reference in Pace3D.

The 12 paired cases in Sec. 5.1 and Appendix D.4 were prepared separately and are not a subset of Eval3DEdit. We used existing 3D generation and editing models to prepare candidate source–target pairs, followed by manual adjustment and screening. As described in Appendix C (“Data and case selection”), we retained cases with clear, spatially coherent structural changes and well-aligned source–target geometry, excluding minor, fragmented, or ambiguous changes.

These paired references support the structural-progress diagnostic in Sec. 5.1 and the proxy analyses in Tables 8–9. Table 8 compares editing results under proxy and oracle control. Table 9 evaluates phase error, progress ordering, and agreement with the oracle’s reset decisions, reporting a mean phase error of 0.133, an AUC of 0.857, and an F1 score of 0.592. These diagnostics assess the proxy on the 12 curated cases; they do not establish phase-estimation accuracy across the full benchmark.


[W3] Interpreting the PC-PC Drop and Mitigating Over-Editing

Using the controller alone can indeed cause over-editing. Rewinding retains the current latent and extends the sampling horizon, so unintended changes can accumulate along with the requested structural edits (Secs. 4.1–4.2). In Table 3, DINO-src decreases from 0.7957 to 0.7877 and PC-PC from 0.7604 to 0.7383, even though target alignment improves. Table R1 additionally shows lower Missing Structure but lower IoU with the controller alone.

PC-PC measures similarity to the source point cloud, so a correct structural edit can also lower this score. A PC-PC decrease alone therefore does not establish over-editing. The paired geometric measurements provide additional evidence about the changes relative to the intended target.

Source Alignment applies a weak pull toward the source latent at each SS step, attenuating accumulated deviations while allowing changes reinforced by the target condition to persist (Sec. 4.2 and Appendix A.4). Adding it raises DINO-src to 0.7914 and PC-PC to 0.7815, while also improving PC-Txt and MV-Img. The IoU recovery in Table R1 is consistent with this complementary role.

Cross-Condition Consensus then preserves local features during SLat generation by comparing source- and target-conditioned predictions at the same noisy state and aggregating their disagreement across steps (Sec. 4.3). It favors source features at shared locations with low disagreement, while new voxels and locations with stronger disagreement remain target-driven.

For reference, we reproduce the component ablation from Table 3 of the submitted paper below.

VariantDINO-src ↑PC-PC ↑PC-Txt ↑MV-Img ↑
Baseline0.79570.76040.24090.7071
+ Controller0.78770.73830.25320.7218
+ Src. Align.0.79140.78150.25370.7229
+ CCC (full Pace3D)0.80090.79970.25660.7291

The full model improves both source fidelity and target alignment over the baseline in Table 3. Table R3 further shows fewer extra additions and incorrect deletions in regions that should remain unchanged.

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

This paper propose Pace3D, which addresses the incomplete geometric edits in hierarchical 3D generators, when sparse support is finalized, downstream feature generations cannot recover the missing support. Pace3D solve this issue by matching current decoded structure with a source-to-proxy reference path and rewinds sampling time when progess is insufficient. Experiments are conducted on Eval3DEdit, which show superior results across most metrics.

Critical strengths

Clear motivations and focusing on a practical bottleneck. The paper propose to impose interventions early and before when feature refinements cannot repair, which is efficient and an economical way.

Meaningful empirical study. The proposed method is evaluated both qualitatively and quantitatively, both showing promisin results. This can be of practical use.

Critical weaknesses

Inconsistent results between tables. In tab. 10, the mean differences are different to those reported in tab. 2. Concretely, if we calclulate the difference, in tab. 2, the PC(PC) is -0.0157, while in tab. 10, it is -0.0228. MVImg in tab. 2 is 0.0163, whereas in tab. 10, it is 0.0115. This means there are inconsistencies in the experiments and therefore reduced the soundness of this paper.

Questions for authors
  1. Please clarify on the inconsitencies shown above in the weaknesses part.
3 comments

Thank you for your positive assessment. For the case-averaged metrics, Table 2 included all 100 cases, including the measured scores of the two invalid PartFlow outputs. In Table 10, the invalid-output filter excluded the corresponding two pairs, and the paired differences were computed over the remaining 98 cases. We have corrected the conversion issue and recomputed the PartFlow means and paired differences using the same 100 cases. The Pace3D scores in Table 2 remain unchanged.

We traced the issue to the conversion of Eval3DEdit’s GLB models into PartFlow’s input format. The two affected cases, action_change/0 and object_removal/11, had geometry offset from the origin. For these cases, our Blender and trimesh steps used different scaling centers. Applying the same scale and translation without accounting for this difference moved part of the geometry outside the voxel grid, resulting in PartFlow outputs with 256 or fewer occupied voxels.

The measured scores of the two invalid PartFlow outputs were included in Table 2, while the corresponding pairs were excluded from Table 10. In the submitted version, the PC-PC difference from the 100-case means in Table 2 was (0.7997-0.8154=-0.0157), compared with (-0.0228) in the 98-case paired analysis in Table 10. For MV-Img, the corresponding differences were (0.7291-0.7128=0.0163) and (0.0115), respectively.

We corrected the scaling-center inconsistency and reran PartFlow. All 100 cases now produce valid outputs. The score changes for the two affected cases are shown below.

Metricaction_change/0: before → afterobject_removal/11: before → after
PC-PC0.0130 → 0.85000.4500 → 0.9000
DINO-src0.4800 → 0.79000.7720 → 0.8500
PC-Img0.0700 → 0.27600.2510 → 0.3150
PC-Txt0.0650 → 0.23000.2680 → 0.2750
MV-Img0.3400 → 0.72000.5380 → 0.7500
MV-Txt0.3700 → 0.49000.4950 → 0.5200
DINO-tgt0.2000 → 0.57000.4550 → 0.6100

The corrected PartFlow results for Table 2 are reported below alongside the unchanged Pace3D scores.

MethodPC-PC ↑DINO-src ↑FID ↓PC-Img ↑PC-Txt ↑MV-Img ↑MV-Txt ↑DINO-tgt ↑
PartFlow0.82830.809750.730.36460.24900.71870.47900.5672
Pace3D (Ours)0.79970.800950.060.36530.25660.72910.48990.5733

The updated results for the Pace3D–PartFlow comparison in Table 10 are shown below. The two corrected cases are included in the paired analysis, and differences are computed as Pace3D − PartFlow using the same 100 cases as Table 2. The mean paired differences therefore agree with the differences between the means in Table 2. For example, the PC-PC difference is (0.7997-0.8283=-0.0286), and the MV-Img difference is (0.7291-0.7187=0.0104).

MetricMean paired difference95% confidence interval
PC-PC−0.0286[−0.0560, 0.0018]
DINO-src−0.0088[−0.0189, 0.0031]
PC-Img+0.0007[−0.0086, 0.0086]
PC-Txt+0.0076[−0.0002, 0.0165]
MV-Img+0.0104[0.0016, 0.0207]
MV-Txt+0.0109[0.0047, 0.0175]
DINO-tgt+0.0061[−0.0010, 0.0148]

These updates do not change the relative ranking of Pace3D and PartFlow on any of the reported metrics. FID is computed over rendered image sets and is reported separately from the case-averaged paired differences. The corrected FID scores are 50.73 for PartFlow, giving a Pace3D − PartFlow difference of −0.67.

We will update the corresponding results in Tables 2 and 10 and clarify the preprocessing and handling of invalid outputs in the revised manuscript.

Thank you for the clarifications and results, which suggests 1) the relative conclusion does not change before and after fixing the experimental errors, 2) the soundness assessment of this paper needs to be revised and the relative performance to existing methods is near a boundary of noise.

Therefore, I would like to maintain my score, but would be open to a rejection. I'll see other reviewer's comments on this point about experimental merits in later discussion.

Thank you for taking a look at the corrected results. We understand your concern and would like to clarify the preprocessing issue and explain why we believe the experimental results remain meaningful.

We ran PartFlow using its official implementation. The released code does not include the preprocessing needed to convert Eval3DEdit’s GLB assets into its input format, so we implemented this conversion separately. By manually checking all outputs against the corresponding asset renderings, we identified two problematic cases with off-center geometry. Although the files could be loaded, inconsistent scaling centers in our Blender/trimesh conversion placed part of the geometry outside the voxel grid. We corrected this preprocessing error, reran both cases, and recomputed the means and paired differences using the same 100 valid cases. We reported the before-and-after metrics for both cases in our earlier response.

Regarding performance, our main benchmark results are averaged over three random seeds, as stated in Section 5 and Appendix B.2. Table 10 also reports paired bootstrap confidence intervals over evaluation cases. After the correction, the gains over PartFlow are +0.0104 on MV-Img, with a 95% CI of [0.0016, 0.0207], and +0.0109 on MV-Txt, with a 95% CI of [0.0047, 0.0175]. Both intervals exclude zero, providing statistical evidence against attributing these multi-view alignment gains solely to noise. PC-Txt and DINO-tgt also show positive mean gains of +0.0076 and +0.0061, with 95% CIs of [-0.0002, 0.0165] and [-0.0010, 0.0148], respectively. These intervals extend only slightly below zero, although we do not claim statistical significance for these two metrics.

We also compared against Nano3D under matched SS-NFE and comparable end-to-end runtime in Table 6. Pace3D achieves 0.7291 MV-Img, compared with 0.7190 for SS-NFE-matched Nano3D and 0.7202 for its comparable-runtime variant. In the latter comparison, Pace3D takes 11.14s versus 11.31s, including our proxy construction. These controls support the benefit of our method beyond simply spending more computation.

Figures 4 and 6 provide qualitative comparisons across object addition, removal, replacement, pose changes, and appearance edits. Examples such as adding apples to the basket in Figure 4, and removing the basket lid or changing the robot’s pose in Figure 6, show how the requested changes are realized while retaining unrelated source content. These examples make the differences in edit completion visible alongside the quantitative results.

Our central claim is that monitoring structural progress and adapting the sampling horizon helps complete requested structural changes while limiting unintended changes. The table below reproduces our geometric evaluation on 12 separately prepared 3D pairs, reported in the response to Reviewer inQM, Rebuttal by Authors [3/3], Section 2. Pace3D achieves the highest structural completion and voxel IoU at both resolutions. Together with the compute controls and qualitative comparisons, these results support our claim.

MethodStructural completion (%) ↑ (64³)IoU ↑ (64³)Structural completion (%) ↑ (256³)IoU ↑ (256³)
Edit36021.50.23943.00.177
EditP2323.90.47949.90.405
AnchorFlow45.50.54262.60.468
3DEditFormer52.00.77472.60.730
PartFlow40.20.77168.20.760
Nano3D42.80.77866.30.759
Nano3D†59.30.77377.90.717
Pace3D (Ours)65.00.81381.40.780

† denotes the SS-NFE-matched Nano3D baseline.

We will release the code, preprocessing scripts, and processed evaluation data. In the final version of the paper, we will also document the original geometry centers and alignment corrections for the affected cases so that the results can be independently checked.

With the discussion period about to close, we cannot complete further experiments or prepare a detailed per-sample comparison of the corrected input and output coordinates before the deadline. For reference, the original source assets are publicly available in the Eval3DEdit benchmark introduced in AnchorFlow (Zhou et al., CVPR 2026). The appendix and our responses provide detailed settings, corrected results, confidence intervals, compute controls, and direct geometric measurements. We hope this evidence can be considered as a whole when assessing our experimental contribution.

Thank you again for your time and consideration.

Reviewer_inQMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Pace3D, a training-free 3D editing framework that addresses the structural progress–clock mismatch in hierarchical 3D generation. The authors observe that a fixed sampling schedule may terminate before the desired structural changes are fully realized. To address this issue, Pace3D introduces a Structural-Progress Controller that adaptively extends the sampling horizon based on decoded structural feedback. In addition, Source Alignment stabilizes the editing trajectory, while Cross-Condition Consensus selectively preserves unchanged source features. Experiments on Eval3DEdit demonstrate improved target alignment while maintaining competitive source fidelity compared with existing 3D editing methods.

Critical strengths

S1. Interesting and Well-Motivated Problem.

The paper identifies the structural progress–clock mismatch in hierarchical 3D editing, where a fixed sampling schedule may lead to incomplete target structures. This is an interesting observation that provides a meaningful motivation for adaptive sampling in 3D editing.

S2. Effective Training-Free Editing Framework.

The proposed framework combines structural progress control, source alignment, and cross-condition consensus to improve target realization while preserving source content. The training-free design is also practically appealing, as it does not require additional model training or per-instance optimization.

S3. Comprehensive Experimental Evaluation.

The paper provides extensive comparisons with existing 3D editing methods, including quantitative evaluations, qualitative results, and a user study. The component ablations and compute-matched comparisons further support the effectiveness of the proposed framework.

S4. Promising Editing Quality.

Pace3D achieves the best reported target-alignment results on Eval3DEdit while maintaining competitive source fidelity. The qualitative examples also demonstrate improvements in structural completeness and preservation of unrelated source content.

Critical weaknesses

W1. Limited Generalization Beyond Hierarchical 3D Generative Models.

The proposed method is primarily designed for hierarchical 3D generators with separate sparse-structure and feature-generation stages. Its applicability to other 3D representations or generative architectures remains unclear. Further discussion or experiments would help demonstrate the generality of the proposed framework.

W2. Limited Direct Evaluation of Structural Completion.

Although improving structural completeness is a central goal of the paper, the main evaluation primarily relies on semantic alignment and perceptual metrics. While the authors analyze structural lag and missing structure on a small set of paired examples, a more comprehensive quantitative evaluation of structural completion would better support the effectiveness of the proposed approach.

Questions for authors

Q1. Could the authors discuss whether the proposed structural-progress controller can be extended to other 3D generative architectures that do not explicitly separate sparse-structure generation from feature generation? What modifications would be required?

Q2. Could the authors provide a more comprehensive quantitative evaluation of structural completion, such as the proportion of successfully generated target components or the amount of missing target geometry across the full benchmark?

9 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/3]

Thank you for the careful review and for recognizing the motivation, framework design, and empirical results of our work. We clarify the architectural scope of Pace3D and provide additional evaluations of structural completion below.


[W1 & Q1] Scope and Applicability

1. Extensions and component-wise applicability

We investigate an extension of Pace3D by adapting the Structural-Progress Controller (SPC) and Source Alignment (SA) to a shared geometry–appearance latent. This adaptation computes structural feedback from decoded occupancy and applies source alignment along the unified latent trajectory. The full three-component framework is designed for hierarchical pipelines with a sparse-structure stage followed by a separate feature-generation stage. In this framework, SPC and Source Alignment operate on the sparse-structure trajectory, while Cross-Condition Consensus uses the subsequent feature-generation stage.

2. Adaptation to UniLat3D

UniLat3D (Wu et al., CVPR 2026 Findings) jointly models geometry and appearance in a shared latent representation generated by a single flow model, and its decoder provides voxel occupancy predictions. We apply SPC and SA to this unified latent trajectory. SPC uses decoded occupancy and a source-to-proxy reference to guide horizon adjustment, while SA regularizes the latent updates. We omit Cross-Condition Consensus because its current formulation requires a separate feature-generation stage.

Both configurations are evaluated on the same 100 samples from Eval3DEdit using the same editing seed. The baseline uses FlowEdit-style editing on a 30-step time grid, starting at step 10 with an initial noise time of approximately 0.857 and performing 20 updates. The source and target CFG scales are set to 1 and 10, respectively, following UniLat3D’s convention.

With SPC enabled, we retain these base editing settings and generate a target proxy with an additional 20-step SDEdit run using a separate seed fixed at 0. SPC measures structural progress solely through decoded occupancy matching, with a lag threshold of 0.1 and at most six rollback operations. The table compares training-free editing with and without SPC and SA using the metrics in Table 2.

3. Preliminary results and discussion

MethodPCPC ↑DINOsrc\mathrm{DINO}_{\mathrm{src}} ↑FID ↓PCImg ↑PCTxt ↑MVImg ↑MVTxt ↑DINOtgt\mathrm{DINO}_{\mathrm{tgt}} ↑
UniLat3D + FlowEdit-style baseline0.83310.805850.200.35960.24640.71040.47490.5614
+ SPC & Source Alignment0.84530.809948.840.36170.24620.71300.47430.5631

On UniLat3D, adding SPC and Source Alignment improves all three source-fidelity metrics. The three image-based target-alignment metrics also increase slightly, while the two text-based scores decrease slightly: PCTxt changes from 0.2464 to 0.2462, and MVTxt from 0.4749 to 0.4743. These results provide preliminary evidence for the joint adaptation of the two components to UniLat3D.

Our sparse-structure analysis in Sec. 3 does not establish that structure-driven updates preserve the desired appearance in this shared geometry–appearance latent. The full three-component framework, including Cross-Condition Consensus, has not been evaluated in this single-stage setting. We will clarify this scope and include the adaptation details and results in the revised manuscript.

AuthorsOpenReview ↗

Rebuttal by Authors [2/3]

[W2 & Q2] Quantitative Evaluation of Structural Completion

We have added direct geometric evaluation on 12 separately prepared paired cases and VLM-based evaluation on all 100 Eval3DEdit cases. Compared with Nano3D†, Pace3D achieves 65.0% vs. 59.3% structural completion at 64364^3 on the paired set, and 71% vs. 59% VLM-assessed overall edit completion on Eval3DEdit. Since Eval3DEdit does not provide paired target 3D assets, we use render-based assessment for the full benchmark and the separate paired references for direct geometric measurements.

1. VLM-assessed overall edit completion on Eval3DEdit

We evaluate all 100 cases in Eval3DEdit, including both structural and appearance edits, using gpt-6-sol. Each case is presented with the editing instruction, source and target images, and six paired renderings of the source and edited assets. Source and edited renders share the same camera positions, with adjacent views 60° apart. The inputs are labeled, and the judge is asked to identify the intended change before returning a yes/no completion decision. Each edited result is assessed once by the VLM, and uncertain cases are referred for human review.

We use an all-or-nothing criterion: a case is counted as complete only when all requested changes are judged to be fulfilled based on the provided views, with no partial credit. For structural edits, added parts must appear as recognizable geometry at the intended locations, and parts requested for removal must be absent from the provided views. Recoloring alone does not count as completing a structural edit. The judge is instructed to assess completion independently of visual quality. Method names are not provided, and all methods are evaluated with the same prompt and rendering settings. The full evaluation prompt is provided in Appendix H.

MethodVLM-assessed overall edit completion (%) ↑
Edit3608
EditP2330
AnchorFlow59
3DEditFormer55
PartFlow38
Nano3D39
Nano3D†59
Pace3D (Ours)71

† denotes the SS-NFE-matched Nano3D variant reported in Table 2.

Under this protocol, Pace3D achieves the highest VLM-assessed overall edit completion rate among the compared methods. This result indicates that more editing requests are judged to be fully completed from the provided rendered views.

AuthorsOpenReview ↗

Rebuttal by Authors [3/3]

2. Supplementary geometric evaluation on paired cases

To complement the benchmark-wide assessment with direct geometric measurements, we evaluate the 12 paired references used in Sec. 5.1. These references were prepared separately and are not a subset of the 100 Eval3DEdit cases. Their preparation and selection are described in Appendix C (“Data and case selection”).

Following Appendix C, Eq. (48), we measure Missing Structure, MstructM_{\mathrm{struct}}, as the fraction of required voxel additions and deletions that remain unresolved relative to the paired target. We report its complement as structural completion:

Cstruct=1−Mstruct.C_{\mathrm{struct}} = 1-M_{\mathrm{struct}}.

The metrics are computed from voxelizations of the final edited meshes and the paired source and target meshes. Before computing the voxel metrics, the source, target, and edited meshes from all methods have consistent position, orientation, and scale. For each pair, all meshes are voxelized over the same spatial domain using the same grid at each resolution. The table reports the mean structural completion over the 12 pairs as a percentage, together with 3D voxel IoU against the paired target. We evaluate both metrics at 64364^3 and 2563256^3 and compare methods within each resolution.

MethodStructural completion (%) ↑ (64364^3)IoU ↑ (64364^3)Structural completion (%) ↑ (2563256^3)IoU ↑ (2563256^3)
Edit36021.50.23943.00.177
EditP2323.90.47949.90.405
AnchorFlow45.50.54262.60.468
3DEditFormer52.00.77472.60.730
PartFlow40.20.77168.20.760
Nano3D42.80.77866.30.759
Nano3D†59.30.77377.90.717
Pace3D (Ours)65.00.81381.40.780

† denotes the SS-NFE-matched Nano3D variant reported in Table 2.

Pace3D achieves the highest structural completion and voxel IoU at both resolutions. Its structural completion exceeds Nano3D† by 5.7 percentage points at 64364^3 and 3.5 percentage points at 2563256^3. The corresponding IoUs of 0.813 and 0.780 also exceed those of all compared methods, indicating closer agreement with the paired target geometry on these 12 cases.


We thank the reviewer for the constructive feedback. We have clarified the scope of Pace3D and provided additional results on UniLat3D adaptation, benchmark-wide edit completion, and paired geometric evaluation. We will incorporate these clarifications and results into the revised manuscript.

Comments on the paper

3 comments
Area_Chair_MZqBOpenReview ↗

discussion period reminder

Dear reviewer:

At first, thank you for being a reviewer for ICLR 2027, and thanks for your contributions so far.

We are half way through the discussion period. Please check if the authors' response have addressed your concerns, any remaining/extended concerns, and update your final rating with justification as early as possible.

Thank you for your contributions again

AC

AuthorsOpenReview ↗

Global Response

We thank the AC and all reviewers for their constructive feedback.

To address the main concerns raised by the reviewers, we made the following key changes.

  • We strengthened the evaluation of structural completion and over-editing (raised by inQM [W2, Q2] and HZT1 [W1, W3]) by adding direct geometric measurements on 12 separately prepared paired cases and VLM assessments on all 100 Eval3DEdit cases. Pace3D achieves 65.0% and 81.4% structural completion at 64³ and 256³, compared with 59.3% and 77.9% for Nano3D with a matched sparse-structure evaluation budget. Across the full benchmark, Pace3D achieves 71% VLM-assessed overall edit completion and 13% over-editing, compared with 59% and 24% for the same baseline. Additional geometric ablations show that full Pace3D reduces extra additions from 53.9% to 28.1% and incorrect deletions from 30.7% to 15.9% relative to baseline. We also clarified that PC-PC measures similarity to the source, and geometric ablations show that Source Alignment reduces unintended changes during extended sampling.
  • We expanded proxy validation (raised by HZT1 [W2] and inQM’s follow-up) by comparing against ground-truth references and testing randomly removed voxels and missing target components. On the 12 paired cases, the SDEdit proxy yields 65.00% structural completion at 64³, close to 65.64% with the ground-truth target reference. Across all 100 benchmark cases, removing 50% of occupied proxy voxels changes DINO-src from 0.8009 to 0.7988 and MV-Img from 0.7291 to 0.7229. We also clarified that an incomplete proxy can still provide useful structural progress cues for extending sampling, while the target condition continues to guide geometry generation.
  • We extended the evaluation beyond hierarchical generators (raised by inQM [W1, Q1]) by applying the Structural-Progress Controller and Source Alignment to the single-stage UniLat3D model on all 100 Eval3DEdit cases. Adding these two components improves all three source-fidelity metrics: PC-PC from 0.8331 to 0.8453, DINO-src from 0.8058 to 0.8099, and FID from 50.20 to 48.84. The three target-alignment metrics also improve.
  • We specified the data sources and controller ablations (raised by HZT1 [W2, W4]) by clarifying that the main comparison uses all 100 Eval3DEdit cases and their provided target images, generated with Nano Banana, while the 12 paired references were prepared separately. We detailed how the Random and Fixed variants are constructed and provided the explicit Fixed sampling schedule. Pace3D exceeds both schedules on all four reported metrics, reaching 0.2566 PC-Txt and 0.7291 MV-Img, compared with 0.2489 and 0.7191 for Fixed.
  • We corrected the inconsistency between Tables 2 and 10 (raised by WTEP) by fixing a scaling-center mismatch in PartFlow preprocessing and rerunning the two affected cases. We recomputed the PartFlow means and paired differences using the same 100 valid cases. The corrected PC-PC and MV-Img differences are −0.0286 and +0.0104, respectively, matching the differences between the means in Table 2.

We are grateful to reviewers inQM, WTEP, and HZT1 for their thoughtful comments, which helped us improve the clarity and rigor of the paper, and we thank the Area Chair for their consideration.

AuthorsOpenReview ↗

Index to Author Responses

Dear Area Chair, Senior Area Chair, Program Chairs, and Reviewers,

For ease of reference, we provide the following index, linking each concern to our response and identifying the relevant section.

Response guide

Reviewer inQM

  • [W1/Q1] Applicability beyond hierarchical models: we tested the Structural-Progress Controller and Source Alignment on single-stage UniLat3D and clarified how these components can be adapted to this setting.
    See Rebuttal by Authors [1/3], “[W1 & Q1] Scope and Applicability.”

  • [W2/Q2] Evaluation of structural completion: we added geometric completion measurements on 12 paired cases and VLM-based edit-completion assessment on all 100 benchmark cases.
    See Rebuttal by Authors [2/3], “VLM-assessed overall edit completion on Eval3DEdit,” and Rebuttal by Authors [3/3], “Supplementary geometric evaluation on paired cases.”

  • [Follow-up] Reliability of the proxy: we explained how partial proxy structure can still guide sampling and tested robustness to randomly removed voxels and missing target components.
    See “Response on Effectiveness of SDEdit-based proxy and Robustness of Controller”: Part [1/2] for the explanation; and Part [2/2] for the experiments.

Reviewer WTEP

  • Inconsistent results in Tables 2 and 10: we corrected the preprocessing error, reran the affected cases, and recomputed the means and paired differences using the same 100 valid cases. We updated Tables 2, 6, and 10 and the data-processing description in Appendix B.1 of the revised paper, with these changes highlighted in blue.
    See Official Comment by Authors, which explains the preprocessing issue and reports the corrected results for Tables 2 and 10.

Reviewer HZT1

  • [W1] Missing direct measurements of completion and over-editing: we added measurements of missing structure, extra additions, and incorrect deletions on 12 paired cases, plus VLM assessments of completion and over-editing on all 100 benchmark cases.
    See Rebuttal by Authors [1/4], “Results on the 12 Paired Cases,” and Rebuttal by Authors [2/4], “Assessment of Overall Edit Completion” and “Evaluation of Over-Editing.”

  • [W2] Evaluation coverage and the 12-case proxy validation: we clarified that the main evaluation uses all 100 benchmark cases. Proxy–ground-truth comparisons use 12 separately prepared pairs because Eval3DEdit does not provide paired target 3D assets.
    See Rebuttal by Authors [3/4], “[W2] Clarification of Data Sources.”

  • [W3] PC-PC drop and possible over-editing: we clarified that the drop occurs with the controller alone. Component ablations show that adding Source Alignment raises PC-PC above the baseline while reducing unintended changes.
    See Rebuttal by Authors [3/4], “[W3] Interpreting the PC-PC Drop and Mitigating Over-Editing.”

  • [W4] Unclear Random and Fixed variants: we specified how both schedules are constructed, how their compute budgets are set, and the exact schedule used for Fixed.
    See Rebuttal by Authors [4/4], “[W4] Implementation Details of the Random and Fixed Variants”

Thanks to the reviewers for their thoughtful feedback and to the chairs for their time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.