- 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
- 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.
- 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.
- 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
-
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.
-
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
- 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?
- Keep regions are copied from the source at the end of both stages. How much of the preservation gain remains without this final copy?
- 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
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.
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.
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:
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.