acceptodds
Under review as a conference paper at ICLR 2027

UMAS3D: Unified Mask-Free 3D Editing with Region-Adaptive Source Preservation

Abstract

3D editing aims to modify an existing 3D asset according to user intent while preserving geometry and appearance outside the intended edit region. Existing methods often rely on task-specific training, source-3D inversion, or external 3D masks, and lack a unified framework for diverse editing conditions and tasks. We propose UMAS3D, a unified 3D editing framework that is training-free, inversion-free, and free from external 3D masks. UMAS3D converts text, bounding boxes, reference images, and their combinations into common target-view guidance, after which addition, deletion, modification, appearance, and pose editing share the same downstream 3D generation process. Our key insight is to couple online edit-region estimation with region-aware source preservation. Mask-Free Adaptive Region Modeling (MARM) initializes an Edit/Keep prior from source-target visual differences and progressively refines it as the target structure evolves, while Source-Guided Preservation Steering (SGPS) constrains the current Keep-region states to retain source geometry and appearance without suppressing target-driven changes. Experiments show that UMAS3D supports addition, deletion, modification, appearance, and pose editing while effectively preserving unedited source content. These advances make UMAS3D a unified framework for flexible and source-preserving 3D editing across diverse editing conditions. Project Page: https://paperprojectpage.github.io/UMAS3D

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Editing, Training-Free Editing, Mask-Free Editing, Multimodal Guidance, Adaptive Region Refinement, Source-Guided Preservation

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 4 reviews, above 27% of the others.
Reviewer_pvbAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UMAS3D, a training-free and inversion-free 3D editing framework built on TRELLIS that does not require an external 3D mask. Text, bounding boxes, reference images and their combinations are converted by Gemini models into a single edited target image, after which addition, deletion, modification, appearance and pose edits share the same 3D pipeline. MARM initializes an Edit/Keep partition by clustering source voxels into regions and comparing DINOv2 features between the source view and the target image, then reassigns regions during Stage 1 based on how much their occupancy deviates from the source. SGPS keeps the Keep regions close to the source by correcting the predicted clean state and applying small gradient steps on occupancy (Stage 1) and SLAT features (Stage 2). On 150 objects from Pxform and Omni3DEdit, UMAS3D achieves the best results on all preservation and target-quality metrics against six baselines, and is preferred in a 40-participant user study.

Critical strengths
  1. The framework handles five editing tasks and five input combinations with one training-free pipeline, without source inversion or an external 3D mask, which is practical for real use.
  2. Updating the Edit/Keep partition during generation addresses a real limitation of fixed masks computed before the target shape exists, and Fig. 8 shows the partition changing during Stage 1.
  3. Preservation is measured directly against the source within ground-truth Keep regions, and UMAS3D leads on all metrics in Table 3 and in the user study (57–61% preference).
Critical weaknesses
  1. The evaluation does not fully support the main claims. The paper's headline contribution is unifying text, boxes and reference images through a generated target image. Using a ground-truth rendering as the shared target image is a reasonable controlled protocol (Sec. 4.1), but as a result the multimodal pipeline is evaluated only qualitatively. In addition, Sec. 4.1 states that baseline details are provided in the Appendix, but I could not find them there. Results are reported only as an average over all tasks, although tasks such as pose editing, where parts move, seem to conflict with region-based source preservation. Moreover, the paper cites AnchorFlow (Zhou et al., CVPR 2026) but does not include it in the comparison. Similarly, TanGO (Lim et al., ECCV 2026), although built on VecSet-based 3D generators, is neither cited nor compared; a discussion of these concurrent mask-free methods would strengthen the paper.

  2. The source of the preservation gains is not isolated. Both stages end by copying the source into the Keep regions, so preservation largely depends on how accurately MARM separates Edit from Keep. However, the paper reports neither the accuracy of the MARM partition nor the effect of this final copy separately from the SGPS steering. It is also not stated which 3D masks VoxHammer and Easy3E receive; if they are coarser than the partition MARM derives from the target image, the preservation gap in Table 3 may mainly reflect the masks rather than the editing method. Likewise, the external 3D masks in Table 5, which mask-free initialization outperforms (Pres-IoU 0.709 vs. 0.644), are not described. Finally, since the quantitative targets are clean ground-truth renderings, it remains untested how robust MARM is when the target image is generated and may contain inconsistent content.

Questions for authors
  1. In the quantitative setting, the target image is a ground-truth rendering, so unedited parts match the source exactly. In practice, image editors such as Gemini 3 Pro Image often slightly change unedited parts as well. How accurate is the MARM partition in that case?
  2. Keep regions are copied from the source at the end of both stages. How much of the preservation gain remains without this final copy?
  3. Which 3D masks do VoxHammer and Easy3E receive in Table 3, and where do the external 3D masks in Table 5 come from?
2 comments
AuthorsOpenReview ↗

Author Response to Reviewer pvbA

W1. Evaluation Completeness and Comparison Fairness

W1.1. End-to-end multimodal evaluation. The main-paper quantitative comparison uses GT-target renderings as shared guidance to isolate downstream 3D editing performance. We additionally evaluate the complete pipeline on 30 objects with five input combinations: user conditions → generated target image → MARM/SGPS → edited 3D (Appendix A.3). Four rendered views are compared with dataset-provided target reference images using PSNR, SSIM, DINO-I, and LPIPS; no GT target 3D is required.

InputPSNR ↑SSIM ↑DINO-I ↑LPIPS ↓
Text18.79270.91250.90960.1044
Text+BBox17.89890.90850.88280.1159
Text+Ref19.60860.91790.80900.1092
BBox+Ref19.30990.91790.90020.0976
Text+BBox+Ref18.78730.91300.89920.1056

These measure image-space target-view consistency, not GT-3D quality.

W1.2. Baseline protocol and task-wise results. We compare six approaches: trained 3DEditFormer; inversion-based VoxHammer/Vinedresser3D; FlowEdit-based Nano3D; and feed-forward EditP23/Easy3E. All use the same 150 source objects, target-view guidance, GT targets, and seven metrics; each retains its pretrained weights and required input adaptations (Section 4.1; Appendices A.5–A.6). Appendix A.7 gives the complete five-task baseline comparisons. UMAS3D's Pres-IoU / Pres-PSNR / DINO-I by task are: Addition 0.7877/27.7919/0.9153; Deletion 0.7301/25.8058/0.8733; Modification 0.8508/24.9799/0.9230; Appearance 0.7183/24.5461/0.8915; Pose 0.5173/21.4386/0.8627. UMAS3D leads the overall five-task aggregate on all seven metrics (Table 3).

W1.3. Pose editing. MARM updates Edit/Keep assignments as the target structure changes, allowing moving parts to be edited, while SGPS constrains only current Keep regions. On Pose, UMAS3D achieves the highest Pres-IoU 0.5173, lowest CD-L1 0.0248, and highest F@1% 0.6537 and NC 0.8124. 3DEditFormer leads the remaining three metrics (Appendix A.7).

W1.4. AnchorFlow and TanGO. AnchorFlow uses Hunyuan3D 2.1 with latent-anchor alignment; TanGO uses tangent-space token guidance and a VecSet representation. UMAS3D instead employs TRELLIS with MARM's online 3D Edit/Keep refinement and SGPS source constraints in occupancy/SLAT spaces.

W2. Source-Preservation Contributions and Mask Fairness

W2.1 & Q2. MARM, SGPS, and final copy. MARM maintains an adaptive Edit/Keep partition; SGPS applies Keep-region source constraints during structure and SLAT generation. Appendix A.10 isolates intermediate guidance and end-of-stage source composition:

ConfigurationPres-IoU ↑Pres-PSNR ↑Pres-SSIM ↑DINO-I ↑CD-L1 ↓F@1% ↑NC ↑
SGPS Final-only0.527523.32400.82780.82560.02520.68820.8263
UMAS3D w/o Final Copy0.669321.62660.77000.87280.03700.78840.8343
UMAS3D0.721724.94730.85860.89360.01340.81940.8995

Final-only keeps only the final Keep-region Source-SLAT composition, with no intermediate SGPS; w/o Final Copy retains intermediate Stage-1/2 guidance but removes both end-of-stage Grid64/SLAT restorations. Thus source preservation is not solely due to final copying. Appendix A.4 visualizes MARM refinement.

W2.2 & Q3. External masks and GT-mask evaluation. VoxHammer, Easy3E, and the external-mask ablation use estimated, not oracle masks: PartField proposes parts, Qwen3-VL selects relevant parts from instructions, and task-specific geometric rules yield required 3D masks (Appendix A.5). UMAS3D requires no external masks. On 30 EditHero objects, GT masks are used only to define evaluation regions, never as input or as replacements for MARM predictions. UMAS3D achieves Pres-IoU 0.9184, Pres-PSNR 14.0144, Pres-SSIM 0.5262, DINO-I 0.9128, CD-L1 0.0059, F@1% 0.9136, NC 0.9057 (full baseline table: Appendix A.8).

W2.3 & Q1. Robustness to generated target images. MARM does not label every source–target difference as Edit. DINOv2 correspondences, visual changes, and region-level evidence determine the prior; minor unintended image changes with strong correspondence may remain Keep. Structural feedback updates the partition, and SGPS preserves the current Keep regions. Beyond controlled GT-target renderings, the complete Gemini-guided pipeline is evaluated in A.3 as above. Appendix A.4/A.11 and the paper's project page visualize MARM refinement and generated-target editing; these results support practical operation but do not separately quantify every type of target-image error.

I thank the authors for their detailed rebuttal. The additional experiments and clarifications have addressed most of my concerns, particularly regarding the end-to-end evaluation of the full pipeline, task-wise comparisons, the use of estimated 3D masks for the baselines, and the ablation demonstrating that SGPS contributes to preservation beyond the final copy operation.

While I still consider the methodological novelty to be somewhat limited, as the framework primarily integrates target image generation, adaptive refinement of a correspondence-based Edit/Keep partition, and source-guided correction of the Keep regions within TRELLIS, the proposed approach offers practical benefits without requiring additional training. The additional results also provide better empirical support for its effectiveness. Nevertheless, my concerns regarding methodological novelty remain, and I therefore maintain my rating of Weak Accept.

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

UMAS3D edits source 3D assets with frozen TRELLIS models. A multimodal frontend turns text, boxes and reference images into an edit description plus one target view; MARM initializes and then online-updates a binary Edit/Keep label over eight source regions; SGPS preserves source occupancy in Stage 1 and source SLAT features in Stage 2 before the Keep tokens are copied back and decoded. On 150 objects across five tasks the method reports the best value on all seven aggregate metrics over six baselines, under a protocol that gives every method the same manually selected view rendered from the ground-truth target asset.

The main questions concern whether the metrics catch unwanted edits, whether the baseline comparison is fair, and whether source details are preserved in the edited asset.

Critical strengths
  • S1 — Relevant problem and mechanism. Preserving source details as an edit region changes is a useful objective. Coupling MARM's region updates with SGPS's preservation in both TRELLIS stages is the clearest potential technical increment.

  • S2 — Large reported preservation gains under a controlled protocol. In Table 3, Pres-IoU rises from 0.6030 to 0.7217 and Pres-PSNR from 20.0870 to 24.9473 dB versus the strongest compared baselines. These are reported averages using the same selected GT target view, without per-task results or seed variation.

  • S3 — A relevant adaptation ablation. Table 5's E-to-full comparison turns on the adaptive region update within the reported configuration and improves all seven metrics, including Pres-PSNR by 6.3370 dB, and Fig. 8 shows how the partition changes.

Critical weaknesses
  • W1 — A larger wrong edit may not lower the Pres-* scores unless its effects are measured in the evaluation GT Keep region. Changes confined to GT Edit are not directly penalized by these scores (Eqs. 19–28). The GT Keep mask is independent of MARM's predicted partition (p26, L1388–1393), while Table 3's whole-object scores do not isolate edit completion. Please report GT-Edit completion and GT-Keep leakage by task, with a source-copy control.

  • W2 — The paper does not say what base models the baselines use or whether they are the same as UMAS3D's, so it is unclear whether the visual-quality comparison is fair. If UMAS3D uses a stronger base model, its meshes may look better for that reason alone. Sharing a target image does not equalize base models, and the appendix does not list them (p7, L348–353; Appendix A.1–A.7). Please list each model and checkpoint and, where feasible, compare using the same base model.

  • W3 — Three project-page examples raise the question of whether unrelated source details are repeatedly changed during editing. The views and lighting differ, and the page cannot be paused or zoomed. Please provide matched-camera, pausable or downloadable before/after renders of these regions.

    • Addition (lower-left example, backpack added to the mecha). The chest texture appears to change, and the chest geometry may also change.
    • Deletion (upper-right example, monster). The eyeball appears to change in both geometry and texture.
    • Pose (upper-right example). The shoulder armor appears to change noticeably after the pose edit.
  • W4 — The "inversion-free" description does not establish how well source details survive encoding, and the SLAT-space loss does not directly check visible texture. The method avoids reverse-trajectory inversion but still encodes the source into TRELLIS (Algorithm 1). Copying Keep tokens before decoding, without a reported rendered-image loss (§3.5), does not establish fidelity of the decoded mesh or texture. Please report source encode–decode error for geometry and texture, including for supplied textured meshes if supported.

  • W5 — The paper does not yet clarify whether a user can directly supply a textured mesh for editing. Exporting a mesh does not establish that one can be used as input. Please describe the input and encoding path if supported; W4 addresses the fidelity of that path.

Other Issues

  • At p7, L351–353, the quantitative protocol uses one target view; the other canonical views are source views. It would help to clarify whether multiple target views are supported and, if so, how view count and consistency affect results.
  • At §3.5, SGPS depends on TRELLIS occupancy and SLAT tokens. A brief discussion of portability to other representations, or of the TRELLIS-only scope, would be helpful.
  • At §3.5 (Eqs. 11–15), a plain-language clarification that the inference-time updates act on the sample state using occupancy and SLAT losses, rather than an RGB loss, would help.
  • At §3.5 (L188–190) and Fig. 2, it would help to add a short, intuitive walkthrough alongside the MARM and SGPS module names: how the Edit/Keep regions are iteratively refined during TRELLIS Stage 1, how source structure is preserved there, and how SLAT features are preserved in Stage 2. Could the authors also explain why the pipeline is organized this way and what each step contributes to editing and source fidelity?
  • At p2, L64, a brief definition of "user-facing" would help; does it refer to conditions provided by the user?
Questions for authors

Please refer to previous sections.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer ErsG

W1. Editing Quality and Source Preservation

Localized 3D editing must complete the target edit while preserving unedited source regions. We evaluate these complementary goals with Pres-IoU, Pres-PSNR, and Pres-SSIM for source consistency in Keep regions, and DINO-I, CD-L1, F@1%, and NC for overall consistency with the target 3D asset (Appendix A.6). Appendix A.7 reports full five-task baseline comparisons. UMAS3D results are:

TaskPres-IoU ↑Pres-PSNR ↑Pres-SSIM ↑DINO-I ↑CD-L1 ↓F@1% ↑NC ↑
Addition0.787727.79190.86000.91530.00900.84570.9152
Deletion0.730125.80580.85100.87330.01420.81810.9265
Modification0.850824.97990.91650.92300.00840.84890.9225
Appearance0.718324.54610.88190.89150.01090.93270.9213
Pose0.517321.43860.78390.86270.02480.65370.8124

Appendix A.8 uses GT masks only for final-result evaluation; A.10 separates intermediate SGPS from final copying. These are not direct GT-Edit completion, GT-Keep leakage, or pure source-copy measurements.

W2. Backbone Comparability and Evaluation Fairness

UMAS3D uses a frozen TRELLIS backbone; 3DEditFormer, VoxHammer, Vinedresser3D, Nano3D, and Easy3E also use TRELLIS-family models. Each retains its own original pretrained configuration; the methods are not assumed to use identical checkpoints. EditP23 uses a different architecture. All methods use the same 150 samples, target-view guidance, GT targets, and seven evaluation metrics; input formats and operations are adapted according to the original implementations (Section 4.1; Appendix A.5). In Table 6, the same frozen TRELLIS is used in both ablation variants: adaptive MARM refinement raises Pres-PSNR by 6.3370 dB and improves all six other metrics.

W3. Preservation of Local Source Details

We acknowledge residual geometry/texture changes in complex areas such as the mecha chest, monster eyeball, and shoulder armor. Appendix Figure A6 shows source/edited views. UMAS3D improves overall preservation: Pres-IoU 0.7217 vs. 0.6030 and Pres-PSNR 24.9473 vs. 20.0870 dB (best baselines; Table 3). This does not imply perfect preservation of every local detail.

W4. Source 3D Encoding–Decoding Fidelity

Inversion-free excludes reverse-trajectory source inversion, not source encoding. TRELLIS still extracts source structure and SLAT features for preservation guidance. We added a matched-view source 3D encoding–decoding reconstruction test (Appendix A.9): PSNR 25.7378, SSIM 0.9723, DINO-I 0.9908, silhouette IoU 0.9866. These values indicate high reconstruction consistency.

W5. Textured Mesh Input and Encoding Pipeline

UMAS3D takes an existing source 3D asset as input, including textured meshes supported by the current pipeline. Source views are rendered, an edit-relevant view is selected, and user conditions generate the target image. TRELLIS encoding obtains source Grid64 occupancy and SLAT features for source-consistency guidance in Stages 1 and 2. Before decoding, source SLAT features are restored in Keep regions. The target image guides 3D editing, not single-view source reconstruction (Sections 3.2–3.5; Appendix A.5).

Other Issues

O1. Multiple target views. One target view follows the TRELLIS single-image-conditioned pipeline; other canonical views provide source information. All baselines share the one-target-view protocol. Joint multi-view guidance remains future work (Section 3.3).

O2. SGPS portability. Source-consistency constraints could be adapted to voxel-based generators with accessible intermediate states and spatial correspondence, but are validated here only on TRELLIS (Section 3.5).

O3. SGPS inference-time optimization. SGPS updates 3D sampling states via occupancy and SLAT losses, not rendered-image RGB loss. Compared with RGB-loss guidance:

MethodPres-IoU ↑Pres-PSNR ↑Pres-SSIM ↑DINO-I ↑CD-L1 ↓F@1% ↑NC ↑
RGB Loss0.710723.83260.83160.86810.02520.76560.8581
UMAS3D0.721724.94730.85860.89360.01340.81940.8995

UMAS3D improves all seven metrics; only the sampling state is optimized, while TRELLIS weights remain frozen (Section 3.5).

O4. MARM/SGPS across stages. MARM updates Edit/Keep regions in Stage 1. SGPS constrains source occupancy in Stage 1, source SLAT features in Stage 2, and restores Keep-region source features before decoding (Sections 3.4–3.5).

O5. User-facing conditions. These are user-supplied text, bounding boxes, reference images, and combinations; they exclude generated target images and intermediate 3D representations (Section 3.3).

The rebuttal has addressed some of my concerns, and I really appreciate it. However, considering that the novelty is somewhat limited and there are some obvious artifacts after editing, I will maintain my initial rating of weak rejection.

Reviewer_P6pPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose UMAS3D, a training-free 3D editing pipeline that supports multiple forms of user input, including text prompts, reference images, and bounding boxes. UMAS3D eliminates the need for explicit 3D editing masks, which can be challenging to obtain, thereby simplifying the editing process and improving usability. Experimental results demonstrate that UMAS3D outperforms both training-based and training-free 3D editing baselines.

Critical strengths
  • This paper presents a comprehensive training-free framework for 3D editing that supports different editing formats, demonstrating improvements over existing baselines on the benchmark constructed by the authors.

  • Overall, the proposed method is straightforward and easy to understand.

  • Ablation studies are provided to validate the effectiveness of individual components in the proposed editing pipeline.

Critical weaknesses
  • The fairness and reliability of the benchmark constructed by the authors in Table 3 remain questionable. First, the benchmark consists of 150 3D objects manually selected from two sources (Pxform and Omni3DEdit). However, no concrete details are provided regarding the selection criteria, raising concerns about potential selection bias in favor of the proposed method. Second, the relative performance of the baselines is inconsistent with the results reported in the Pxform paper. Specifically, Pxform reports Nano3D > VoxHammer > 3DEditFormer, whereas UMAS3D reports the opposite ranking: 3DEditFormer > VoxHammer > Nano3D. This discrepancy further undermines the credibility of the constructed benchmark. Third, the editing tasks and input formats in the benchmark appear to be tailored to the proposed UMAS3D framework, yet the authors don't provide details on how baseline methods with different input requirements are adapted and evaluated under the same settings. These issues raise significant concerns about the fairness of the comparisons and the validity of the quantitative results reported in this submission.

  • The overall technical contribution of this paper is limited. Most components of the proposed framework rely on well-established techniques from prior work, while the main novel component, adaptive editing region refinement, does not demonstrate substantial improvements in the ablation studies. Additionally, directly preserving the source voxels may introduce inconsistencies or visible artifacts at the boundaries between edited and unedited regions.

  • Runtime is an important metric for training-free methods, as they often incur higher computational costs than training-based approaches. However, the authors only report the runtime of UMAS3D without comparisons to existing baselines. Moreover, the reported runtime of approximately four minutes per 3D object seems relatively high, raising concerns about the practical efficiency of the proposed framework.

  • Some notations, like F_pre and F_edit in Eq (2), are not clearly defined, making certain parts of the methodology difficult to follow.

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer P6pP

W1. Benchmark Reliability and Fairness

W1.1. Dataset Selection

We randomly selected 150 3D editing samples from Pxform and Omni3DEdit, including 32 Addition, 29 Deletion, 30 Modification, 29 Appearance, and 30 Pose samples. Selection did not depend on UMAS3D results (Section 4.1 and Appendix A.1).

W1.2. Baseline Ranking Discrepancy

Table 3 in the original paper reports aggregate results across five editing tasks, whereas Pxform uses a different task scope and evaluation protocol. To examine this discrepancy, we separately evaluated the 91 Addition, Deletion, and Modification samples:

MethodPres-IoU ↑Pres-PSNR ↑Pres-SSIM ↑DINO-I ↑
3DEditFormer0.606617.63750.66980.8620
VoxHammer0.607117.65750.67030.8625
Nano3D0.607617.67750.67080.8631
UMAS3D0.790126.23190.87580.9045

On these three tasks, the baseline ranking is Nano3D > VoxHammer > 3DEditFormer, consistent with the trend reported in Pxform. The three-task results are presented in Section 4.2, Table 4, and the complete five-task results in Appendix A.7.

W1.3. Fair Comparison Across Different Input Formats

All methods are evaluated under the same experimental setting: the same 150 3D editing samples, shared target-view guidance, a common evaluation protocol, and the same seven metrics. Each baseline is adapted to its original input interface and available editing operations, without modifying pretrained model weights. The input adaptations, auxiliary conditions, and evaluation settings are detailed in Section 4.1 and Appendices A.5–A.6.

W2. Technical Novelty and Boundary Consistency

W2.1. Technical Contributions

The main technical contributions of UMAS3D are two cooperating mechanisms:

  1. MARM: Adaptively refines the 3D Edit/Keep regions as the target structure evolves during generation, rather than relying on a fixed region partition.
  2. SGPS: Applies source-consistency constraints in both structure and SLAT feature spaces to reduce unintended changes in unedited regions.

Together, they enable localized 3D editing without additional training, source inversion, or external 3D editing masks (Sections 3.4–3.5).

W2.2. Effectiveness of Adaptive Region Refinement

The original paper already evaluates MARM through ablation experiments. Variant E and full UMAS3D use the same initialization and both SGPS stages; the only difference is whether adaptive region refinement is enabled. Enabling MARM improves Pres-PSNR by 6.3370 dB, with improvements across the other six metrics. Quantitative results are reported in Section 4.4, Table 6, while Appendix Figure 8 visualizes the Edit/Keep regions from initialization through adaptive refinement.

W2.3. Boundary Consistency

UMAS3D does not simply concatenate source and edited voxels along a fixed boundary. MARM adaptively updates the Edit/Keep regions as the target structure evolves, while SGPS applies source-consistency constraints in structure and SLAT feature spaces to mitigate boundary inconsistencies and artifacts associated with fixed region partitions. Appendix Figure 8 directly visualizes the adaptive refinement of the Edit/Keep regions; the mechanisms are detailed in Sections 3.4–3.5.

W3. Runtime Efficiency

We measured runtime on 50 randomly selected objects under identical hardware and settings.

MethodRuntime (s/object) ↓
3DEditFormer48.9
TRELLIS58.2
Vinedresser3D110.1
EditP23129.2
UMAS3D (every 2 steps)190.0
Nano3D214.8
UMAS3D216.7
VoxHammer258.1
Easy3E345.2

Standard UMAS3D takes 216.7 s/object on average, comparable to Nano3D and faster than VoxHammer and Easy3E. Meanwhile, UMAS3D achieves 0.7217 Pres-IoU (best baseline: 0.6030) and 24.9473 dB Pres-PSNR (best baseline: 20.0870 dB), demonstrating better preservation of unedited regions.

Furthermore, performing the source-consistency optimization every two steps reduces runtime to 190.0 s/object (a 12.3% reduction). This shows a favorable quality–runtime tradeoff (Appendix A.5).

W4. Notation Clarification

We have formally defined both functions immediately after Eq. (2) in the revised Section 3.3: Fpre\mathcal{F}_{\mathrm{pre}} (Gemini 3.6 Flash) integrates the available multimodal conditions into a unified editing description EE; Fedit\mathcal{F}_{\mathrm{edit}} (Gemini 3 Pro Image) generates the target image ItI_t from the source view, EE, and the available bounding-box and reference-image conditions.

I appreciate the authors' responses and the updated results, particularly regarding runtime efficiency. However, I remain concerned about the effectiveness of the quantitative benchmark and the technical novelty of the proposed method. Therefore, I have decided to maintain my original score.

Reviewer_LiG8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UMAS3D, a unified 3D editing framework that is training-free, inversion-free, and does not require externally provided 3D masks. Its core consists of two modules, MARM and SGPS. MARM initializes Edit/Keep regions based on visual differences between source and target views and adaptively refines these regions during generation. SGPS preserves unedited regions by imposing source-consistency constraints in both structure space and SLAT feature space, reducing unintended modifications to unrelated content.

Critical strengths
  1. Practical research problem. Eliminating the need for user-provided external 3D masks lowers the barrier to localized 3D editing.

  2. Adaptive region refinement. The method integrates editing-region localization into the generation process, offering greater flexibility than approaches that rely entirely on predefined, fixed masks.

Critical weaknesses
  1. The reliability of mask-free 3D region localization is insufficiently established. MARM infers 3D Edit/Keep regions by combining source-target visual differences from a single target view with multi-view DINOv2 feature clustering on the source 3D object. However, feature similarity and spatial proximity do not necessarily correspond to semantic part boundaries or actual editing regions. This is particularly concerning for complex structures, occlusions, and fine-grained local edits. Moreover, the use of only eight adaptive regions may be too coarse to capture complex editing boundaries. The paper lacks direct quantitative evaluation of the predicted regions against ground-truth 3D editing regions.

  2. The occupancy-deviation-based adaptive refinement lacks a reliable semantic basis. MARM determines region transitions based on relative structural deviations among regions sharing the same current state. However, structural deviation does not necessarily indicate whether a change is intended by the user. Unintended structural changes introduced by the generative model may trigger incorrect Edit/Keep transitions, potentially amplifying generation errors rather than correcting them. Conversely, regions that should be edited but have not yet undergone sufficient structural changes may remain incorrectly protected. The reliability of this feedback mechanism is not sufficiently investigated.

  3. The contribution of mask-free localization is not sufficiently isolated. The reported preservation metrics evaluate final editing results rather than the accuracy of MARM itself. Since SGPS additionally employs source-state alignment, sample-level optimization, and final source-feature composition, improvements in preservation do not necessarily demonstrate accurate Edit/Keep localization. More controlled experiments, particularly comparisons with oracle region assignments, are needed to establish the actual contribution of MARM.

  4. Minor presentation issue. Figure 2 contains a typographical error in the label "Source 3D With Spatial oordinates", where the initial letter "C" in "Coordinates" is missing. The figure should be corrected for clarity and presentation quality.

Questions for authors
  1. Can the authors provide quantitative evaluations of MARM's predicted Edit/Keep regions against ground-truth 3D editing regions?

  2. How does occupancy-deviation-based refinement distinguish intended structural changes from unintended changes introduced by the generative model?

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer LiG8

We thank the reviewer for the constructive feedback.

W1. Accuracy of 3D Editing-Region Localization

W1.1. Region initialization. Initial region assignments may be imperfect, motivating MARM's adaptive refinement. MARM uses multi-view source DINOv2 features, 3D spatial continuity, and source–target view differences to initialize Edit/Keep regions. These assignments are priors, not fixed final partitions: MARM re-evaluates region states as the target structure evolves, reducing reliance on imperfect initialization. Appendix Figure 8 visualizes the process (Section 3.4).

W1.2. Region granularity and localization. The eight regions model adaptive states, not a one-to-one correspondence to semantic parts. We evaluated different region counts (Appendix A.5):

KKPres-IoU ↑Pres-PSNR ↑DINO-I ↑CD-L1 ↓
40.714524.69770.88110.0149
80.721724.94730.89360.0134
120.728624.93460.88680.0155
160.723924.54750.88670.0150

K=8K=8 leads on four of the full seven metrics; larger KK does not consistently improve performance. We also evaluate final edits using GT-defined regions (W3 & Q1).

W2. Reliability of Adaptive Feedback Refinement

W2.1 & Q2. Incorrect updates from unintended structural changes. Occupancy deviation alone does not determine editing intent. Source–target visual differences, multi-view DINOv2 features, and spatial continuity establish an initial Edit/Keep prior. Occupancy deviation is then used only as structural feedback to revise this prior, not as an independent semantic signal. To reduce erroneous transitions, MARM holds the partition for the first 35 steps; later changes require consistent evidence in at least two of three checks, with a stricter Keep→Edit threshold (Section 3.4; Appendix A.5). This reduces transient errors but cannot guarantee semantic correctness in every case.

W2.2. Missed updates for weak intended changes. MARM periodically reassesses region states instead of retaining a fixed partition (Appendix Figure 8). In the controlled ablation, Variant E and UMAS3D share the same initialization and both SGPS stages; only adaptive refinement differs. Refinement increases Pres-PSNR by 6.3370 dB and improves the six other metrics (Table 6, Section 4.4).

W3 & Q1. Independent MARM Contribution and GT-Mask Evaluation

MARM predicts/refines Edit/Keep regions; SGPS applies source constraints to the resulting Keep regions (Sections 3.4–3.5). The Variant-E comparison above isolates the effect of adaptive refinement, and Appendix Figure 8 shows its evolving partitions. We further evaluated 30 EditHero objects using GT 3D masks to define evaluation regions for the final edited assets (Appendix A.8):

MethodPres-IoU ↑Pres-PSNR ↑Pres-SSIM ↑DINO-I ↑CD-L1 ↓F@1% ↑NC ↑
TRELLIS0.14707.48350.37350.86440.05180.30670.6481
EditP230.17165.19080.40630.44660.05860.35920.6712
Easy3E0.26199.79140.44300.86020.02170.47260.7570
VoxHammer0.903913.95060.52510.90360.00830.90610.9040
Nano3D0.877213.30830.49550.90370.00740.91530.9038
Vinedresser3D0.910513.89970.52640.91210.00660.91270.9024
3DEditFormer0.874212.42470.47350.84910.02210.83130.8518
UMAS3D0.918414.01440.52620.91280.00590.91360.9057

UMAS3D obtains the strongest overall results. GT masks are used only for evaluation, never as UMAS3D inputs or replacements for MARM predictions. Accordingly, this experiment evaluates final edit quality within GT-defined regions, not direct MARM-predicted-versus-GT mask overlap accuracy. The controlled comparison of Variant E with full UMAS3D holds initialization and both SGPS stages fixed, isolating refinement’s contribution to final editing quality; Appendix Figure 8 provides the corresponding region evolution. These are complementary evidence for refinement, not direct localization-accuracy measurements.

W4. Figure 2 Typo

We corrected “Spatial oordinates” to “Spatial Coordinates.”

Thank you for the additional experiments. The region-count ablation and the Variant-E comparison provide evidence that adaptive refinement improves the final editing results.

However, the paper does not directly evaluate the accuracy of MARM’s predicted Edit/Keep regions. The GT-mask experiment measures final preservation quality instead of region-localization accuracy.

I therefore keep my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.