- Rating
- 4(+1)
- Confidence
- 3
- Informativeness
- 3(+1)
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes CUESplat, a text-driven 3D Gaussian editing framework with three main components, Multi-view Edit Coupling (MEC), Structural Edit Inference (SEI), and Selective Gaussian Adaptation (SGA). MEC improves multi-view consistency by sharing editing information across views during image generation. SEI predicts edited geometry to update cross-view correspondences during generation and identifies reliable addition and removal regions afterward. Finally, SGA uses the edited multi-view images and structural evidence from SEI to determine which Gaussians to remove or add and which existing Gaussians to refine during optimization. The paper is well-motivated, with a reasonable framework design and comprehensive experiments that demonstrate promising results.
Critical strengths
- The paper has a clear and meaningful motivation, addressing the unwanted geometry problem in existing Gaussian editing methods.
- The three main components are reasonably designed and work well together. The ablation studies are comprehensive and demonstrate the effectiveness of each component.
- The authors demonstrate that SEI and SGA can be integrated into DGE and EditSplat with consistent performance improvements, showing their effectiveness across different image editing pipelines.
- The authors provide an honest discussion of failure cases and limitations, which helps clarify the limitations of the proposed method.
Critical weaknesses
- Although the ablation studies are detailed, they are only conducted on appearance editing tasks. Since SEI and SGA are particularly important for object addition and removal, the authors should explain this choice or provide additional ablation experiments on these tasks.
- The method heavily relies on the accuracy of depth predictions in SEI, and inaccurate predictions may lead to errors in structural updates. The current experiments lack an evaluation of the method's robustness to depth prediction errors.
- The framework involves repeated depth prediction and two stages of Gaussian optimization, which may introduce considerable computational overhead. A runtime comparison with existing methods would be helpful.
Questions for authors
- In Eq. (6), why is source confidence used for removal and edited confidence for addition? Why not use the same confidence source for both or combine them?
- The RGB change threshold in Eq. (15) is based on the median, MAD, and 65th percentile. If all pixel differences are very small, such as 0.001, the threshold would also be 0.001, and could still reach 1. Could this give too much weight to small changes or noise? Would an absolute threshold help?
- Can the method handle other editing tasks, such as pose changes or object replacements involving large geometry changes? It would be helpful to see some examples.
- How does the Gaussian selection guided by SEI compare with simply projecting 2D edit masks onto the original 3DGS? If ground-truth masks are available, can the authors also compare the selected regions with them?
- The method uses several thresholds for structural evidence and Gaussian updates. How were thresholds such as and the Gaussian removal thresholds chosen? Appendix C.4 analyzes some hyperparameters, but more sensitivity experiments on the remaining thresholds would be helpful.
10 comments
Q4: Separate confidence gating for removal and addition in Eq. 6.
A4: We thank the reviewer for this question. Decoupling for removal and for addition gates each structural operation by the branch that actually observes the target surface. Removal relies on reliable source geometry via because the edited branch only sees newly exposed background, whereas addition relies on reliable edited geometry via because the new object is absent from the source scene.
Combining both confidences conflates two distinct surfaces and suppresses valid structural updates, as shown in the table below:
| Confidence Formulation | CLIP-S ↑ | CLIP-D ↑ | DINO ↑ | Gemini ↑ | MEt3R ↓ |
|---|---|---|---|---|---|
| Ours Decoupled | 0.2420 | 0.1113 | 0.7050 | 1.83 | 0.2530 |
| Minimum of and | 0.2395 | 0.1079 | 0.6997 | 1.67 | 0.2668 |
| Product | 0.2389 | 0.1071 | 0.7011 | 1.75 | 0.2655 |
Both joint formulations degrade performance across every metric. We have added this ablation and clarification to Section 3.2 of the revised manuscript.
Q5: Adaptive RGB threshold sensitivity to minor noise.
A5: We thank the reviewer for pointing out this edge case. In practical editing tasks, foreground modifications usually produce non-uniform RGB differences (190 out of 192 key views across our 16 appearance editing tasks naturally yield adaptive thresholds above 0.020). To safeguard against extreme near-zero difference edge cases, incorporating an absolute noise floor in normalized RGB space into Eq. 15 prevents uniformly small differences from triggering false activations:
As shown in the table below, setting suppresses spurious background activations in the 2 low-difference views:
| Affected Views | Activated Pixel Ratio | |
|---|---|---|
| 0.010 (Ours) | 0 / 192 | 13.875% |
| 0.000 | 2 / 192 | 14.512% |
| 0.005 | 2 / 192 | 14.512% |
| 0.020 | 0 / 192 | 13.875% |
We have updated Eq. 15 and included this analysis in the revised manuscript.
Q6: Pose changes and large geometric edits.
A6: We thank the reviewer for this question. We evaluate this capability on our 12 newly added tasks, which include challenging large-geometry object replacements (turning a bear statue into a stone elephant and replacing a bonsai with a potted cactus) alongside local, global, and street-scale edits:
| Method | CLIP-S ↑ | CLIP-D ↑ | DINO ↑ | Gemini ↑ | MEt3R ↓ |
|---|---|---|---|---|---|
| CUESplat | 0.2697 | 0.1739 | 0.7622 | 3.08 | 0.2991 |
| DGE | 0.2354 | 0.0890 | 0.7401 | 2.67 | 0.3231 |
| GaussianEditor | 0.2425 | 0.1070 | 0.7545 | 1.83 | 0.3348 |
| DreamCatalyst | 0.2414 | 0.1099 | 0.7523 | 1.92 | 0.3212 |
| TrAME | 0.2345 | 0.0885 | 0.7442 | 2.25 | 0.3163 |
| GaussCtrl | 0.2450 | 0.1103 | 0.7305 | 2.83 | 0.3286 |
| CoreEditor | 0.2228 | 0.0664 | 0.7250 | 2.08 | 0.3254 |
| Instruct-GS2GS | 0.2342 | 0.1286 | 0.7121 | 1.33 | 0.4808 |
| EditSplat | 0.2424 | 0.1252 | 0.7416 | 2.42 | 0.3238 |
| ViP3DEdit | 0.2192 | 0.0504 | 0.7072 | 1.58 | 0.3441 |
| Omni-3DEdit | 0.2215 | 0.0833 | 0.7148 | 2.17 | 0.3569 |
CUESplat ranks first across all metrics. CUESplat handles large geometric and pose shifts because SEI separates removed source surfaces from newly added regions via signed depth residuals, allowing SGA to remove selected source Gaussians rather than stretching existing geometry. We have included these results and visual comparisons in the revised manuscript.
Q7: SEI-guided Gaussian selection versus 2D mask projection.
A7: We thank the reviewer for this insightful question. Directly unprojecting 2D edit masks fails because they lack depth selectivity and directional edit evidence. Specifically, 2D projection sweeps all Gaussians along camera rays, and during geometric expansion, such as turning a person into a wider Minecraft character, it mistakenly selects the newly occluded background wall as an editable target instead of keeping it fixed and initializing new primitives in front.
We evaluate the methods at two levels on the Person scene with 32,815 source Gaussians across 12 key views. First, we measure 3D Gaussian over-selection across the whole scene:
| Selection Strategy | Selected Source GS | Scene Ratio | Depth-Inconsistent GS |
|---|---|---|---|
| SEI-Guided Selection (Ours) | 1,471 | 4.48% | 0 |
| Depth-Truncated Projection | 13,181 | 40.17% | 0 |
| Naive 2D Projection | 17,952 | 54.71% | 4,771 |
SEI restricts selection to the true foreground target (4.48% of scene Gaussians) with zero depth-inconsistent primitives. In contrast, naive projection selects 54.71% of the scene, and depth truncation still over-selects 40.17% because it cannot separate the edited target from the newly occluded wall.
Second, we evaluate pixel-level selection routing inside the expansion region, defined as the newly covered background wall where the target silhouette extends beyond the source person across the same 12 views:
| Method | Assigned to Source Edit | Routed to New Addition | Kept Unchanged |
|---|---|---|---|
| SEI (Ours) | 1.16% | 45.10% | 53.74% |
| Unsigned 2D Mask | 97.85% | 0.00% | 2.15% |
Within the newly occluded wall expansion zone, SEI protects 98.84% of the newly occluded wall pixels from unintended source edits, whereas an unsigned 2D mask blindly assigns 97.85% of the background wall pixels to source Gaussian editing and corrupts the wall. We have added this analysis to the revised manuscript.
Q8: Selection and sensitivity of structural and update thresholds.
A8: We thank the reviewer for this suggestion. Rather than per-scene tuning, the local evidence threshold and the per-Gaussian removal threshold in Eq. 20 are both set to and remain fixed. To verify sensitivity beyond Appendix C.4, we evaluate and across a fourfold range around our default value:
| Threshold | Value | CLIP-S ↑ | CLIP-D ↑ | DINO ↑ | Gemini ↑ | MEt3R ↓ |
|---|---|---|---|---|---|---|
| 0.010 (Ours) | 0.2420 | 0.1113 | 0.7050 | 1.83 | 0.2530 | |
| 0.005 | 0.2412 | 0.1105 | 0.7026 | 1.75 | 0.2619 | |
| 0.020 | 0.2416 | 0.1110 | 0.7034 | 1.75 | 0.2568 | |
| 0.010 (Ours) | 0.2420 | 0.1113 | 0.7050 | 1.83 | 0.2530 | |
| 0.005 | 0.2415 | 0.1111 | 0.7027 | 1.75 | 0.2612 | |
| 0.020 | 0.2413 | 0.1109 | 0.7032 | 1.83 | 0.2637 |
Performance remains highly stable across the entire fourfold range. Setting cleanly balances noise suppression and structural completeness. We have added this sensitivity analysis to Appendix C.4 of the revised manuscript.
Response to Reviewer Yb2c
Q1: Ablations on object addition and removal.
A1: We thank the reviewer for this suggestion. To verify how our core modules support structural changes, we conduct an ablation study directly on the object addition and removal tasks:
Unlike appearance editing, object addition and removal require revising visible surfaces rather than only updating colors. In our pipeline, MEC aligns cross-view edits along evolving correspondences to maintain consistent 3D boundaries (improving CLIP-D by 18.0% and MEt3R by 9.9%), SEI extracts paired depth evidence to separate true structural changes from 2D appearance shifts (boosting Gemini quality by 18.6% and CLIP-D by 10.1%), and SGA executes local structural updates by removing selected Gaussians, reusing revealed surfaces, and initializing missing geometry (improving Gemini quality by 10.7% and CLIP-D by 10.6%).
We have added this structural ablation study to the revised manuscript.
Q2: Robustness of SEI to depth prediction errors.
A2: We thank the reviewer for raising this point. To evaluate robustness against depth estimation noise, we independently perturb valid pixels in both source and edited depth maps using multiplicative noise d~=deϵ, where ϵ follows a zero-mean Gaussian distribution with standard deviation of 0.01, 0.02, and 0.04::
Under depth noise, all metrics remain virtually unchanged. This robustness stems from three mechanisms in SEI. Anchoring the source-GS residual to source Gaussians avoids single-view scale drift, cross-view support suppresses uncorrelated depth jitter across views, and geometric confidence filters out minor residual fluctuations before structural updates. We have added this noise sensitivity evaluation to the revised manuscript.
Q3: Runtime comparison.
A3: We thank the reviewer for this suggestion. We evaluate the end-to-end runtime on the Bear scene alongside all appearance editing benchmark metrics across all 11 methods under identical hardware and evaluation settings:
CUESplat completes a full scene edit in 6.85 min (411 s), ranking second fastest among all 11 methods while achieving the best performance across all metrics, because depth inference runs only on sparse key views and SGA updates only selected editable Gaussians. We have added this runtime comparison to the revised manuscript.