acceptodds
Withdrawn from ICLR 2027

ConcordFlow: View-Aware Complex 3D Editing via Flow Aggregation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing image-guided 3D editing methods provide effective control when a requested change can be clearly specified from a single view. Complex instructions, however, may involve multiple changes best observed from different viewpoints. Applying these changes sequentially through multi-round editing can lead to error accumulation, while jointly processing multiple view-specific conditions requires resolving spatially varying editing effects and potential conflicts among updates. To address these challenges, we introduce ConcordFlow, a mask-free and training-free framework built on TRELLIS for view-aware complex 3D editing that jointly integrates multiple conditions in a single 3D editing process. In the sparse-structure stage, trajectory-guided spatial weighting (TGSW) derives spatial weights from independent editing and restoration trajectories to identify each condition's editing influence. Disagreement-guided reweighting (DGR) then adjusts these weights during joint editing using source–target velocity disagreement. In the subsequent structured-latent stage, disagreement–visibility adaptive fusion (DVAF) likewise compares source- and target-conditioned velocities and combines the resulting disagreement with geometric visibility to guide feature generation. To evaluate our framework, we introduce ComplexEdit3D-Bench for view-aware complex 3D editing. On this benchmark, ConcordFlow achieves stronger condition alignment and better overall source preservation than the compared multi-round methods.

Primary area:
generative models
Keywords:
3D Editing, 3D Generation, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WyySOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ConcordFlow, a training-free and mask-free single-round 3D asset editing framework designed to address the task of complex 3D asset editing conditioned on multi-view images. To tackle the issues of error accumulation and corruption of irrelevant regions that plague existing multi-round editing methods, the framework introduces Trajectory-Guided Spatial Weighting (TGSW), Disagreement-Guided Reweighting (DGR), and Disagreement-Visibility Adaptive Fusion (DVAF) into the sparse-structure and latent-feature generation stages of TRELLIS, respectively. This enables efficient joint coordination of multi-view editing conditions and precise local control, substantially improving editing quality and preserving source-asset integrity while greatly reducing computational overhead. Qualitative and quantitative experiments demonstrate that ConcordFlow outperforms existing mainstream multi-round editing baselines.

Critical strengths

S1. The paper explicitly points out that conventional multi-round serial editing is prone to error accumulation and mutual interference between successive modifications, making the topic highly practical and forward-looking.

S2. The method is elegantly designed, complete, and theoretically sound.

S3. The framework requires no fine-tuning of the underlying diffusion model, and through single-round joint generation it avoids the substantial computational overhead of repeatedly performing 150-view rendering in TRELLIS multi-round editing, yielding a clear inference-speed advantage.

S4. This paper brings comprehensive experimental evaluation and a benchmarking contribution.

Critical weaknesses

W1. Strong dependence on 2D image editing quality: ConcordFlow depends on the upstream 2D editing model. If the upstream model fails to produce reasonable results, the editing results will likewise degrade.

W2. Struggle to handle order-dependent edits. if the complex instruction contains causal/sequential dependencies, independently generated multi-view 2D target images may exhibit severe spatial-structural conflicts, and the framework currently struggles to handle such order-dependent edits

W3. Hyperparameter sensitivity and the setting of temperature coefficients: The framework introduces several temperature-coefficient hyperparameters (τtgsw,τdgr,τdvaf\tau^{tgsw}, \tau^{dgr}, \tau^{dvaf}) as well as step-truncation parameters Kdgr,KdvafK^{dgr}, K^{dvaf}. Although the appendix provides step-sensitivity tests for KdgrK^{dgr} and KdvafK^{dvaf} (Table 8), there is no discussion of the sensitivity of the Softmax temperature coefficients τ\tau. If τ\tau is too small, the weights may concentrate excessively on a single view; if too large, the scheme degenerates to uniform averaging.

Questions for authors

Although the paper claims to perform single-round editing, I would like to ask whether the proposed framework can also support multi-round editing. I think the proposed method is also able to do multi-round editing. If so, how does your method perform on such cases? Furthermore, do you have any ideas for addressing such order-dependent edits? How does the proposed method can be expanded to such cases?

4 comments

Your suggestions prompted us to examine temperature sensitivity and clarify how ConcordFlow can handle order-dependent edits. Changes in the revised manuscript are highlighted in orange.

W1. Strong dependence on upstream 2D editing quality

We agree that ConcordFlow depends on the quality of the edited 2D target images. This limitation is shared by 3D editing methods that use edited images to condition an image-to-3D generator: missing or incorrect edits in the target images can carry over to the final 3D asset. ConcordFlow coordinates edits across views, but is not designed to correct errors introduced by the upstream 2D editor.

W2 & Q. Can ConcordFlow handle order-dependent edits over multiple rounds, and if so, how?

Order-dependent instructions do not necessarily require multiple editing rounds if their image conditions can be constructed in advance. The limitation in Sec. 5 concerns our single-round framework: conditions requiring intermediate edited results cannot be handled within that round. For such cases, we can apply ConcordFlow repeatedly in a multi-round workflow, organizing edits into successive rounds according to their dependencies. Each invocation remains a single-round joint edit, and its result becomes the source for constructing the next round's conditions. In our two-round example, the first round replaces the hat and cup, while the second adds details to the resulting hat and camera. We have added this workflow and the visualization to Appendix D.12.

W3. Missing analysis of softmax temperature sensitivity

To assess how sensitive editing performance is to the choice of softmax temperature, we vary one module’s temperature at a time while keeping the other two at their defaults. Across the tested temperature settings, Local DINO-I ranges from 0.760 to 0.767 and CLIP-T from 0.252 to 0.255, indicating that condition alignment is robust to temperature variations within this range. We have added these results to Appendix D.13.

TGSW temperature (DGR = 1.0, DVAF = 1.0).

TemperatureCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
0.20.01026.370.9530.0460.7620.255
0.50.01026.510.9540.0450.7670.255
1.00.01026.540.9530.0450.7640.254
1.50.01026.570.9530.0450.7640.254
2.00.01026.630.9540.0450.7630.254

DGR temperature (TGSW = 0.5, DVAF = 1.0).

TemperatureCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
0.20.01026.230.9520.0460.7600.253
0.50.01026.440.9530.0460.7630.254
1.00.01026.510.9540.0450.7670.255
1.50.01026.560.9540.0450.7620.254
2.00.01026.590.9540.0440.7600.254

DVAF temperature (TGSW = 0.5, DGR = 1.0).

TemperatureCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
0.20.01026.420.9530.0460.7640.255
0.50.01026.520.9530.0450.7650.255
1.00.01026.510.9540.0450.7670.255
1.50.01026.540.9540.0450.7660.253
2.00.01026.510.9540.0450.7650.252
AuthorsOpenReview ↗

Follow-up on Additional Experiments and Clarifications

Dear Reviewer,

Thank you again for your thoughtful feedback. May we ask whether our additional experiments and clarifications have addressed your concerns? Briefly:

  • Order-dependent editing: We clarify how ConcordFlow can be applied in successive rounds and provide a two-round example in which later edits build on the hat and camera produced in the first round.
  • Temperature sensitivity: Across the tested temperature settings, Local DINO-I ranges from 0.760 to 0.767 and CLIP-T from 0.252 to 0.255, supporting robustness within the tested range.
  • Upstream editing quality: We acknowledge that dependence on 2D target-image quality is a shared limitation of 3D editing methods conditioned on edited images. ConcordFlow coordinates view-specific conditions but does not correct errors introduced by the upstream editor.

The two-round visualization and temperature analysis are included in the revised manuscript. Please let us know if any concerns remain. We would be happy to clarify them.

Best regards,

The Authors

Thank you for your detailed explain. I decide to keep my score.

Thank you for your time and constructive feedback. Your suggestions have helped us improve the paper, and we appreciate your consideration of our responses.

Reviewer_w5pMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses complex 3D editing tasks in which different edits are best observed from different viewpoints. Rather than performing editing iteratively across multiple rounds, ConcordFlow generates a target image for each view and integrates them into a single editing pass using TRELLIS. Specifically, TGSW and DGR adaptively weight the contributions of different views during the sparse-structure stage, while DVAF fuses multi-view information in the SLat stage. The paper also introduces ComplexEdit3D-Bench, a benchmark consisting of 100 samples for evaluating complex 3D editing.

Critical strengths
  1. The problem is practical and well motivated. Fig. 1 clearly illustrates the error accumulation caused by multi-round editing.

  2. The proposed framework is training-free and mask-free. The VR/VO/NV analysis also provides useful insights into the roles of individual components.

  3. The introduction of ComplexEdit3D-Bench, particularly its annotated 3D masks, is a useful contribution to evaluating complex 3D editing.

Critical weaknesses
  1. The reported improvements are not clearly attributable to the proposed components. All baselines in Table 1 use multi-round editing, making it difficult to distinguish the benefits of single-round editing from those of the proposed techniques. More importantly, Table 7 shows that the uniform single-round variant, with all proposed components disabled, achieves better source preservation than the full method (PSNR: 27.01 vs. 26.51; LPIPS: 0.033 vs. 0.045). This suggests that the preservation improvements primarily come from the single-round formulation, while the proposed components introduce a trade-off between preservation and editing alignment.

  2. Important baselines are missing. The paper does not compare against directly conditioning TRELLIS on all target images jointly, which would be a natural baseline for evaluating the proposed multi-view fusion strategy. Comparisons with existing multi-view editing methods discussed in the related work, such as Edit360, EditP23, and Pro3D-Editor, are also absent.

  3. The experimental evidence for improved editing alignment is limited. In Table 1, alignment is primarily evaluated using CLIP-T, where the improvements over baselines are approximately 0.01, without reporting variance or statistical significance. Meanwhile, the CD metric shows little difference across methods. The improvements in Local DINO-I reported in the ablation study are also relatively small, making it difficult to assess their significance without variance estimates.

  4. The technical design is largely heuristic and incremental. The weighting and fusion strategies in Eqs. 5, 8, and 13 appear to be hand-designed, without sufficient comparisons against simpler alternatives. Furthermore, the velocity-disagreement signal is inherited from VS3D, which further limits the methodological novelty.

Questions for authors
  1. How does ConcordFlow compare with directly conditioning TRELLIS on multiple target images jointly? Could the authors also include the uniform single-round variant from Table 7 as a baseline in the main comparison?

  2. Could the authors report the variance or confidence intervals for CLIP-T and Local DINO-I to establish whether the observed improvements are statistically meaningful?

  3. Does the reported runtime account for the additional editing and restoration runs required by TGSW?

6 comments

Thank you for the constructive feedback. Below, we clarify what our fusion modules contribute beyond single-round editing and present additional comparisons and evaluations. Changes in the revised manuscript are highlighted in orange.

W1 & Q1. Source-preservation gains and comparisons with Uniform Fusion and native TRELLIS

We agree that single-round editing helps preserve the source by reducing cross-round error accumulation. Our fusion modules address a separate problem: coordinating potentially conflicting edits across views within that round. Uniform Fusion, now included in Tab. 1, preserves the source better but achieves lower condition alignment than ConcordFlow. Its better preservation may partly reflect more conservative updates: equal-weight fusion does not account for differences in spatial relevance or conflicts among conditions, which can weaken requested edits.

Comparison with a CFG-tuned baseline. Tuning CFG itself changes the trade-off between condition alignment and source preservation. To test whether CFG tuning alone can match ConcordFlow’s editing–preservation trade-off, we vary Uniform Fusion’s sparse-structure target CFG and SLat CFG one at a time, keeping the target images and evaluation protocol fixed. In each sweep, we keep the other stage’s CFG at the value used in the original submission. Increasing the sparse-structure target CFG improves Uniform Fusion’s alignment at the cost of source preservation, whereas varying SLat CFG has a smaller effect on alignment. Uniform Fusion does not match ConcordFlow on Local DINO-I or CLIP-T in either sweep. These results have been added to Appendix D.7.

CFG values are listed as (sparse-structure target CFG / SLat CFG).

SettingCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
Sparse-structure CFG sweep
Uniform Fusion (1.5/5.0)0.00928.230.9600.0280.6150.230
Uniform Fusion (3.5/5.0)0.00927.800.9590.0300.6520.233
Uniform Fusion (5.5/5.0)0.01027.010.9560.0330.6670.235
Uniform Fusion (7.5/5.0)0.01026.390.9530.0370.6860.240
Uniform Fusion (9.5/5.0)0.01025.960.9510.0400.6930.243
SLat CFG sweep
Uniform Fusion (5.5/1.0)0.01027.010.9570.0340.6500.233
Uniform Fusion (5.5/3.0)0.01027.160.9570.0330.6630.236
Uniform Fusion (5.5/5.0)0.01027.010.9560.0330.6670.235
Uniform Fusion (5.5/7.0)0.01026.900.9550.0340.6700.237
Uniform Fusion (5.5/9.0)0.01026.690.9520.0350.6680.237
ConcordFlow (5.5/5.0)0.01026.510.9540.0450.7670.255

Comparison with native TRELLIS multi-image generation. We compare ConcordFlow with native TRELLIS [1] conditioned jointly on the same target images. This pipeline generates a new asset from noise without using the source 3D state, rather than editing the existing asset. ConcordFlow achieves higher Local DINO-I and PSNR than native TRELLIS multi-image generation. We have added these results to Appendix D.9.

MethodCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
TRELLIS [1]0.04717.680.9060.1300.5040.239
ConcordFlow0.01026.510.9540.0450.7670.255

W2. Missing comparisons with existing multi-view editing methods

Our original comparison focused on TRELLIS-based editing methods to keep the underlying 3D backbone consistent, as described in the Baselines paragraph of Sec. 4.1. Following your suggestion, we have extended the comparison to Edit360 [2], EditP23 [3], and Pro3D-Editor [4] on ComplexEdit3D-Bench. We apply the view-specific edits sequentially and evaluate the final asset after all editing rounds using the benchmark evaluation protocol. ConcordFlow achieves higher condition alignment and better source preservation on the reported metrics. These results have been added to Appendix D.10.

MethodCD ↓PSNR ↑SSIM ↑LPIPS ↓CLIP-T ↑
Edit360 [2]—17.060.8940.1730.192
EditP23 [3]0.04617.280.8880.1510.223
Pro3D-Editor [4]—21.510.9190.0900.227
ConcordFlow0.01026.510.9540.0450.255

CD is not reported for Edit360 and Pro3D-Editor because their evaluated outputs contain Gaussians but no triangle meshes.

W3 & Q2. Are the reported gains in editing alignment statistically meaningful?

Thank you for raising this concern. Absolute gains in CLIP-T can be modest in 3D editing. For example, Tab. 1 of VoxHammer [5] and Tab. 1 of VineDresser3D [6] show CLIP-T gains of 0.004 and 0.005 over TRELLIS, respectively, on their own benchmarks. In our main comparison, ConcordFlow improves CLIP-T by 0.0085–0.0109 over the five multi-round baselines.

Paired confidence intervals. To assess whether the reported gains are statistically distinguishable from zero, we analyze per-case scores from the multi-round comparisons and component ablations. For each comparison, we subtract the baseline or ablation score from ConcordFlow’s score on each of the same cases. We report the mean of these differences and its 95% confidence interval, estimated by resampling the paired cases. Most paired comparisons show positive gains with 95% confidence intervals entirely above zero, supporting the reported improvements. We have added this analysis and the combined table to Appendix D.11.

MethodΔ CLIP-T (95% CI)Δ Local DINO-I (All) (95% CI)
FlowEdit [7]+0.0099 [0.0066, 0.0133]—
VoxHammer [5]+0.0087 [0.0061, 0.0113]—
VineDresser3D [6]+0.0085 [0.0056, 0.0114]—
Nano3D [8]+0.0098 [0.0065, 0.0132]—
PartFlow [9]+0.0109 [0.0077, 0.0142]—
ConcordFlow w/o DGR+0.0015 [−0.0002, 0.0032]+0.0294 [0.0201, 0.0395]
ConcordFlow w/o TGSW+0.0021 [0.0007, 0.0036]+0.0173 [0.0100, 0.0253]
ConcordFlow w/o DVAF+0.0060 [0.0043, 0.0079]+0.0181 [0.0136, 0.0235]

Edit completion. To provide more direct evidence of improved editing alignment, we report edit completion rates for the same multi-round comparisons and component ablations. We use Qwen3.8-27B [10] to evaluate each sub-instruction across all cases. The judge compares renders of the final 3D asset against the actual 2D target image used by the method for that edit, with the sub-instruction clarifying what to check. An edit passes if it realizes the requested change shown in the target image without visibly damaging directly adjacent structures that the target image leaves intact. We report the percentage of individual edits that pass and the percentage of cases in which every edit passes. ConcordFlow completes all edits in 88% of cases, compared with 78% for the best multi-round method, and also outperforms each component ablation. We have added the evaluation protocol and results to Appendix D.8.

MethodIndividual Edit Completion (%) ↑Complete Case Completion (%) ↑
FlowEdit [7]85.871.0
VoxHammer [5]89.778.0
VineDresser3D [6]87.174.0
Nano3D [8]76.856.0
PartFlow [9]70.049.0
ConcordFlow w/o DGR89.375.0
ConcordFlow w/o TGSW91.482.0
ConcordFlow w/o DVAF92.785.0
ConcordFlow94.888.0

W4. Hand-designed weighting rules and novelty beyond VS3D

Thank you for raising this concern. Here, we elaborate on the design rationale presented in Sec. 3.3 of the original submission and report additional comparisons with simpler alternatives.

Eq. 5 — Trajectory-guided spatial weighting. As discussed in Sec. 3.3 (lines 208–211 and 216–220), a large forward displacement can reflect either a requested edit or unintended drift. Swapping the source and target conditions provides a reverse signal for the requested edit, but does not explicitly instruct the model to undo unrelated changes. Eq. 5 therefore favors locations where both displacements are substantial and approximately cancel. To test the value of the restoration signal, we compare with Edit-only weighting, which uses only the forward displacement magnitude while retaining DGR and DVAF. With restoration, ConcordFlow achieves higher Local DINO-I alongside better source preservation. We have added this benchmark-wide quantitative comparison to Appendix D.4.

MethodCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
Edit-only0.01026.020.9510.0470.7490.254
ConcordFlow0.01026.510.9540.0450.7670.255

Eq. 8 — Disagreement-guided reweighting. Although Eq. 8 expresses the disagreement term in exponential form, it is equivalent to multiplying the normalized trajectory-based and disagreement-based weights and then renormalizing across conditions. Among the settings with DVAF enabled in Tab. 4, combining the two weighting signals yields higher Local DINO-I and CLIP-T than using either signal alone.

Eq. 13 — Disagreement–visibility adaptive fusion. As discussed in Sec. 3.3 (lines 287–296), when the disagreement weights are similar across views, they offer little guidance on which view to favor, so Eq. 13 gives more weight to geometric visibility. When disagreement distinguishes among views, the rule gives it more weight. To evaluate the benefit of entropy-based adaptive fusion, we compare full DVAF with fixed disagreement–visibility ratios, including both single-signal endpoints. We have added these results to Appendix D.5. Full DVAF achieves higher Local DINO-I than either single-signal endpoint or any tested fixed mixture, with similar source-preservation scores.

SLat fusion (Disagreement / Visibility)CD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
0% / 100% (Visibility-only)0.01026.500.9540.0450.7520.250
10% / 90%0.01026.540.9540.0450.7540.251
30% / 70%0.01026.540.9540.0450.7550.252
50% / 50%0.01026.540.9540.0450.7570.253
70% / 30%0.01026.550.9540.0450.7580.254
90% / 10%0.01026.480.9530.0450.7570.255
100% / 0% (Disagreement-only)0.01026.450.9530.0460.7560.255
ConcordFlow0.01026.510.9540.0450.7670.255

Relation to VS3D. We credit VS3D [11] for the idea of comparing source- and target-conditioned model responses. Its TAR module compares the resulting SLat features for one source–target image pair and uses agreement to decide where to inject source features. ConcordFlow instead compares velocities at a shared latent state for each view-specific pair and uses those comparisons to decide how much each view contributes to a joint update. Our contribution is the coordination of multiple, potentially conflicting conditions, not the idea of disagreement itself.

Reviewer_sxS6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript studies view-aware complex 3D editing, where multiple source–target image pairs from different viewpoints specify changes to the same 3D asset. Each target image contains only the edits assigned to its view and may retain the original content in regions modified by other views, creating conflicts between conditions. ConcordFlow performs joint editing using a frozen TRELLIS backbone: TGSW estimates spatial weights from independent editing and restoration trajectories; DGR adjusts these weights using source–target velocity disagreement at the same latent state; and DVAF combines disagreement propagation with geometric visibility during structured-latent generation. The work introduces ComplexEdit3D-Bench, compares against multi-round versions of five editing methods, and provides extensive ablation studies.

Critical strengths

1.The problem is well defined, and the method addresses it coherently. TGSW, DGR, and DVAF handle spatial contribution assignment, biases caused by imperfect restoration, and subsequent feature fusion, respectively. 2.The experiments provide solid evidence of the method’s effectiveness. Tables 4 and 7 cover all component combinations, the category-wise results support the complementarity of TGSW and DGR, and the user study shows a clear preference for the proposed method. 3.The method does not require manual editing masks. It nevertheless outperforms several baselines that receive manually annotated masks, which makes it appealing in practice.

Critical weaknesses

1.Comparisons with joint-editing are missing. Uniform fusion in Table 7 already achieves better source-preservation scores, so comparisons with a tuned uniform-fusion baseline or simple visibility-based weighting are needed to establish the value of the more elaborate weighting scheme. 2.The evaluation does not directly establish whether all requested edits are completed. CLIP-T and average Local DINO-I may overlook failures on individual sub-instructions, while best-result voting in the user study does not measure complete instruction fulfillment. 3.The comparison between multi-round and single-round editing involves several confounding factors. Since later target images are generated from each method’s intermediate outputs, the differences in the main table may also reflect errors from 2D image editing and intermediate reconstruction.

Questions for authors

1.With the same target images and computational budget, how does the method compare with visibility-based weighting or uniform fusion with tuned guidance strength? I would like to understand whether it improves the editing–preservation trade-off, rather than simply producing stronger edits. 2.Could authors provide benchmark-wide comparisons of TGSW against Edit-only weighting, and of full DVAF against visibility-only weighting and fixed-ratio fusion? These would help establish the value of restoration trajectories and entropy-based fusion beyond the examples shown. 3.Could authors report the completion rate for individual sub-instructions and the fraction of samples in which all requested edits are completed? These measures would more directly assess the proposed complex-editing task than average similarity scores. 4.How are the 3D evaluation masks defined for addition, removal, and movement edits? For example, when a part is moved, does the mask cover both its original and new locations? If the author's response resolves my issue, I will consider adjusting my rating.

5 comments

We appreciate your thoughtful review. Below, we address your concerns with additional experiments and clarifications. Changes in the revised manuscript are highlighted in orange.

W1 & Q1. Missing comparisons with joint-editing baselines

To test whether ConcordFlow offers a better editing–preservation trade-off than CFG tuning alone, we vary the sparse-structure target CFG and SLat CFG one at a time for Uniform Fusion and Visibility Fusion. In each sweep, we keep the other stage’s CFG at the value used in the original submission: 5.5 for the sparse-structure target CFG and 5.0 for the SLat CFG. Both baselines use their respective fusion rules in both stages, and all methods share the same single-round setup, target images, and evaluation protocol. We have added these comparisons to Appendix D.7.

Comparison with CFG-tuned baselines. Increasing the sparse-structure target CFG improves alignment for both baselines at the cost of source preservation, whereas varying SLat CFG has a smaller effect on alignment. Neither baseline matches ConcordFlow on Local DINO-I or CLIP-T in either sweep. At their highest Local DINO-I settings, both baselines also have lower PSNR and SSIM than ConcordFlow.

CFG values are listed as (sparse-structure target CFG / SLat CFG).

SettingCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
Uniform Fusion: sparse-structure CFG sweep
Uniform Fusion (1.5/5.0)0.00928.230.9600.0280.6150.230
Uniform Fusion (3.5/5.0)0.00927.800.9590.0300.6520.233
Uniform Fusion (5.5/5.0)0.01027.010.9560.0330.6670.235
Uniform Fusion (7.5/5.0)0.01026.390.9530.0370.6860.240
Uniform Fusion (9.5/5.0)0.01025.960.9510.0400.6930.243
Visibility Fusion: sparse-structure CFG sweep
Visibility Fusion (1.5/5.0)0.00928.100.9600.0300.6410.234
Visibility Fusion (3.5/5.0)0.00927.410.9570.0330.6710.237
Visibility Fusion (5.5/5.0)0.01026.670.9540.0370.6900.240
Visibility Fusion (7.5/5.0)0.01026.150.9510.0400.6990.243
Visibility Fusion (9.5/5.0)0.01125.710.9490.0420.7040.244
Uniform Fusion: SLat CFG sweep
Uniform Fusion (5.5/1.0)0.01027.010.9570.0340.6500.233
Uniform Fusion (5.5/3.0)0.01027.160.9570.0330.6630.236
Uniform Fusion (5.5/5.0)0.01027.010.9560.0330.6670.235
Uniform Fusion (5.5/7.0)0.01026.900.9550.0340.6700.237
Uniform Fusion (5.5/9.0)0.01026.690.9520.0350.6680.237
Visibility Fusion: SLat CFG sweep
Visibility Fusion (5.5/1.0)0.01026.870.9560.0360.6870.239
Visibility Fusion (5.5/3.0)0.01026.850.9550.0360.6900.240
Visibility Fusion (5.5/5.0)0.01026.670.9540.0370.6900.240
Visibility Fusion (5.5/7.0)0.01026.470.9520.0380.6890.240
Visibility Fusion (5.5/9.0)0.01026.260.9500.0390.6870.240
ConcordFlow (5.5/5.0)0.01026.510.9540.0450.7670.255

Computational cost. We report runtimes in seconds following the protocol in Tab. 3. Although ConcordFlow incurs some runtime overhead, it achieves substantially better condition alignment than both baselines.

Method150-view Rendering w/o ↓150-view Rendering w/ ↓
Uniform Fusion85247
Visibility Fusion88250
ConcordFlow128289

The “w/” column adds the estimated 150-view rendering time.

W2 & Q3. Report completion rates for individual sub-instructions and the fraction of cases with all edits completed

Thank you for suggesting a more direct evaluation of edit completion. CLIP-T and Local DINO-I measure condition alignment through similarity, but they do not directly indicate whether each requested edit was completed. We therefore use Qwen3.8-27B [1] as a vision-language judge to assess the completion of each sub-instruction. The judge uses the actual 2D target image provided to each method as the primary reference and the corresponding sub-instruction to clarify what to check. We count an edit as complete if the 3D result realizes the local change shown in the target image without visibly damaging directly adjacent structures that the target image clearly leaves intact. Across all cases, we report completion rates at both the sub-instruction and case levels. A case counts as complete only if every sub-instruction in that case meets the above criterion. ConcordFlow achieves the highest completion rates at both levels, outperforming the compared multi-round methods and all component ablations. We have added the evaluation protocol and results to Appendix D.8.

MethodIndividual Edit Completion (%) ↑Complete Case Completion (%) ↑
FlowEdit [2]85.871.0
VoxHammer [3]89.778.0
VineDresser3D [4]87.174.0
Nano3D [5]76.856.0
PartFlow [6]70.049.0
ConcordFlow w/o DGR89.375.0
ConcordFlow w/o TGSW91.482.0
ConcordFlow w/o DVAF92.785.0
ConcordFlow94.888.0

W3. Different target images across multi-round baselines

We acknowledge that multi-round baselines generate target images after the first round from their own intermediate 3D results, so those images may differ across baselines. In contrast, ConcordFlow uses target images prepared in advance from renders of the original source asset. These differences can affect the results in Tab. 1, which compares complete editing workflows on the same source assets and instructions rather than isolating the effect of our fusion components. To assess those components separately, Tabs. 4 and 7 compare Uniform Fusion, component ablations, and ConcordFlow under the same single-round setup and identical target images. ConcordFlow’s higher condition alignment in these controlled comparisons supports the contribution of our components beyond the single-round formulation itself.

Q2. Benchmark-wide comparisons with Edit-only weighting and simpler DVAF variants

Thank you for suggesting these quantitative comparisons. We evaluated both design choices on ComplexEdit3D-Bench and have added the results to Appendices D.4 and D.5, respectively.

Restoration trajectories in TGSW. Edit-only weighting estimates spatial weights from the editing trajectory alone, while keeping DGR and DVAF enabled. As discussed in Sec. 3.3 (lines 208–211 and 216–220), restoration trajectories help distinguish intended editing influence from unintended displacement, reducing spurious contributions in unrelated regions. Compared with Edit-only, ConcordFlow better preserves unedited content while achieving higher condition alignment.

MethodCD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
Edit-only0.01026.020.9510.0470.7490.254
ConcordFlow0.01026.510.9540.0450.7670.255

Entropy-adaptive fusion in DVAF. As explained in Sec. 3.3 (lines 287–296), DVAF favors visibility when disagreement weights are similar across views, and favors disagreement when these weights distinguish among views. We compare this entropy-based mixing rule with fixed disagreement–visibility ratios, including both single-signal endpoints. Full DVAF achieves higher Local DINO-I than all tested ratios, with similar source-preservation scores.

SLat fusion (Disagreement / Visibility)CD ↓PSNR ↑SSIM ↑LPIPS ↓Local DINO-I ↑CLIP-T ↑
0% / 100% (Visibility-only)0.01026.500.9540.0450.7520.250
10% / 90%0.01026.540.9540.0450.7540.251
30% / 70%0.01026.540.9540.0450.7550.252
50% / 50%0.01026.540.9540.0450.7570.253
70% / 30%0.01026.550.9540.0450.7580.254
90% / 10%0.01026.480.9530.0450.7570.255
100% / 0% (Disagreement-only)0.01026.450.9530.0460.7560.255
ConcordFlow0.01026.510.9540.0450.7670.255

Q4. How are 3D evaluation masks defined for different edits?

Our benchmark groups edits into addition, removal, and replacement. For additions, we mask the region where new geometry is intended to appear. For removals, we mask the original geometry to be removed. For replacements, we mask both the original and intended replacement regions. Movement follows the same mask annotation rule as replacement: the mask covers both the part’s original location and its intended destination.


[1] Qwen Team. Qwen3.8-Max: A New Bar for Coding and Cowork. 2026.

[2] Vladimir Kulikov et al. FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models. ICCV, 2025.

[3] Lin Li et al. VoxHammer: Training-Free Precise and Coherent 3D Editing in Native 3D Space. 3DV, 2026.

[4] Yankuan Chi et al. Vinedresser3D: Towards Agentic Text-guided 3D Editing. CVPR, 2026.

[5] Junliang Ye et al. Nano3D: A Training-Free Approach for Efficient 3D Editing Without Masks. ICLR, 2026.

[6] Jiawei Weng et al. Feedforward 3D Editing Learns from Semantic-Part Transformation. arXiv:2605.27351, 2026.

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

Related papers

Discussion (0)

Sign in to comment.