acceptodds
Under review as a conference paper at ICLR 2027

CUESplat: Guiding Gaussian Scene Updates with Paired 3D Edit Inference

Abstract

Existing text-driven Gaussian editing methods commonly edit rendered views and fit the source Gaussians to the resulting images. Due to geometric changes and multi-view inconsistencies generated during the editing process, image fitting alone is not enough to specify which source surfaces to retain or reconstruct. We present CUESplat, which uses paired 3D edit inference to guide these scene updates. During denoising, a geometry predictor updates edited depths in the source scene frame and returns them to subsequent steps for cross-view coupling. After candidate generation, paired source and edited geometry provides displacement evidence, checked against source geometry and across views, to guide local Gaussian updates. To preserve usable source content, CUESplat reuses existing surfaces before introducing geometry for missing content. This couples structural revision with image fitting for appearance modification, object addition, and object removal. Experiments demonstrate that our method outperforms existing state-of-the-art Gaussian editing methods on three tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Text-driven 3D Scene Editing, 3D Gaussian Splatting, Multi-view Consistency, Structural Edit Inference, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Yb2cOpenReview ↗
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
  1. The paper has a clear and meaningful motivation, addressing the unwanted geometry problem in existing Gaussian editing methods.
  2. The three main components are reasonably designed and work well together. The ablation studies are comprehensive and demonstrate the effectiveness of each component.
  3. 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.
  4. The authors provide an honest discussion of failure cases and limitations, which helps clarify the limitations of the proposed method.
Critical weaknesses
  1. 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.
  2. 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.
  3. 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
  1. In Eq. (6), why is source confidence cisc_i^s used for removal and edited confidence ciec_i^e for addition? Why not use the same confidence source for both or combine them?
  2. 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 wirgbw_i^{rgb} could still reach 1. Could this give too much weight to small changes or noise? Would an absolute threshold help?
  3. 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.
  4. 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?
  5. The method uses several thresholds for structural evidence and Gaussian updates. How were thresholds such as τe\tau_e 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
AuthorsOpenReview ↗

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:

MethodCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
CUESplat (Full)0.24200.11130.70501.830.2530
w/o Multi-view Edit Coupling (MEC)0.22500.10090.69331.500.2733
w/o Structural Edit Inference (SEI)0.22330.09670.69521.330.2849
w/o Selective Gaussian Adaptation (SGA)0.22150.08920.69311.500.2658

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\tilde{d} = d e^{\epsilon}, where ϵ\epsilon follows a zero-mean Gaussian distribution with standard deviation of 0.01, 0.02, and 0.04::

MethodCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
CUESplat (Full)0.24200.11130.70501.830.2530
MapAnything Depth + 1% Noise0.24110.11080.70411.830.2542
MapAnything Depth + 2% Noise0.24090.11020.70371.830.2549
MapAnything Depth + 4% Noise0.23870.10930.70101.670.2580

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:

MethodRuntime ↓CLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
CUESplat6.85 min / 411 s0.26080.17650.77453.130.2448
DGE4.05 min / 243 s0.23750.10490.76582.690.4043
DreamCatalyst8.82 min / 529 s0.24010.08280.77071.880.2967
GaussianEditor9.32 min / 559 s0.23460.08870.75711.630.2889
EditSplat9.72 min / 583 s0.24450.11330.76612.250.4779
GaussCtrl10.33 min / 620 s0.24530.09590.76882.690.3185
ViP3DEdit10.55 min / 633 s0.23490.06310.75301.310.2723
CoreEditor11.42 min / 685 s0.23170.06220.76112.130.4005
Instruct-GS2GS11.45 min / 687 s0.22700.13270.72451.500.3463
TrAME14.15 min / 849 s0.24300.08640.76932.310.3121
Omni-3DEdit27.98 min / 1679 s0.21900.04840.76701.880.2476

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.

Q4: Separate confidence gating for removal and addition in Eq. 6.

A4: We thank the reviewer for this question. Decoupling cisc_i^s for removal and ciec_i^e for addition gates each structural operation by the branch that actually observes the target surface. Removal relies on reliable source geometry via cisc_i^s because the edited branch only sees newly exposed background, whereas addition relies on reliable edited geometry via ciec_i^e 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 FormulationCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
Ours Decoupled0.24200.11130.70501.830.2530
Minimum of cisc_i^s and ciec_i^e0.23950.10790.69971.670.2668
Product cis⋅ciec_i^s \cdot c_i^e0.23890.10710.70111.750.2655

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 τabs=0.010\tau_{abs} = 0.010 in normalized RGB space into Eq. 15 prevents uniformly small differences from triggering false activations:

τirgb=max⁡(τabs,median⁡(di)+2.5⋅1.4826MAD⁡(di),Q0.65(di))\tau_i^{rgb} = \max(\tau_{abs}, \operatorname{median}(d_i) + 2.5 \cdot 1.4826 \operatorname{MAD}(d_i), Q_{0.65}(d_i))

As shown in the table below, setting τabs≥0.010\tau_{abs} \ge 0.010 suppresses spurious background activations in the 2 low-difference views:

τabs\tau_{abs}Affected ViewsActivated Pixel Ratio
0.010 (Ours)0 / 19213.875%
0.0002 / 19214.512%
0.0052 / 19214.512%
0.0200 / 19213.875%

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:

MethodCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
CUESplat0.26970.17390.76223.080.2991
DGE0.23540.08900.74012.670.3231
GaussianEditor0.24250.10700.75451.830.3348
DreamCatalyst0.24140.10990.75231.920.3212
TrAME0.23450.08850.74422.250.3163
GaussCtrl0.24500.11030.73052.830.3286
CoreEditor0.22280.06640.72502.080.3254
Instruct-GS2GS0.23420.12860.71211.330.4808
EditSplat0.24240.12520.74162.420.3238
ViP3DEdit0.21920.05040.70721.580.3441
Omni-3DEdit0.22150.08330.71482.170.3569

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 StrategySelected Source GSScene RatioDepth-Inconsistent GS
SEI-Guided Selection (Ours)1,4714.48%0
Depth-Truncated Projection13,18140.17%0
Naive 2D Projection17,95254.71%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:

MethodAssigned to Source EditRouted to New AdditionKept Unchanged
SEI (Ours)1.16%45.10%53.74%
Unsigned 2D Mask97.85%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 τe\tau_e and the per-Gaussian removal threshold τg\tau_g in Eq. 20 are both set to 0.0100.010 and remain fixed. To verify sensitivity beyond Appendix C.4, we evaluate τe\tau_e and τg\tau_g across a fourfold range around our default value:

ThresholdValueCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
τe\tau_e0.010 (Ours)0.24200.11130.70501.830.2530
τe\tau_e0.0050.24120.11050.70261.750.2619
τe\tau_e0.0200.24160.11100.70341.750.2568
τg\tau_g0.010 (Ours)0.24200.11130.70501.830.2530
τg\tau_g0.0050.24150.11110.70271.750.2612
τg\tau_g0.0200.24130.11090.70321.830.2637

Performance remains highly stable across the entire fourfold range. Setting 0.0100.010 cleanly balances noise suppression and structural completeness. We have added this sensitivity analysis to Appendix C.4 of the revised manuscript.

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

This paper proposes a text-driven 3DGS editing method. It uses paired source and edited geometry to decide which changes are supported (SEI), updates the Gaussians accordingly (SGA), and feeds geometry back during multi-view denoising (MEC).

Critical strengths
  1. The motivation is reasonable. Existing render–edit–fit pipelines fit the edited images without checking whether a change is geometrically supported, and the method addresses this with corresponding modules.

  2. The method gets the best results on all five metrics for appearance editing, and the ablations (Table 3) show that each module contributes. SEI and SGA also improve DGE and EditSplat when plugged into them (Table 2).

Critical weaknesses
  1. The evaluation is small and unclear. Table 1 reports “16 appearance-editing tasks”, but neither the table nor its caption says which scenes or datasets they come from. Only the text reveals that the benchmark consists of eight self-selected, mostly object-centric scenes from four sources, not a standard benchmark. I am concerned about performance on large scenes.

  2. Addition and removal are pooled in Table 1(b) and compared with only two baselines, and the ablations cover only appearance editing. It is therefore unclear how much each module helps for object removal and addition, which are the cases that need structural changes.

Questions for authors
  1. The abstract and Figure 1 are hard to understand. The abstract is full of new terms (paired 3D edit inference, displacement evidence, cross-view coupling) and never describes the problem concretely. Figure 1 is only a pipeline diagram. It says image fitting leaves “unwanted geometry”, but gives no example of this failure.

  2. Many gains are small (e.g., DINO 0.7745 vs. 0.7707, and several ablation differences below 0.01), and no variance over runs is reported. The Gemini score uses only four views per task, so it is hard to tell whether the improvements are significant.

7 comments
AuthorsOpenReview ↗

Response to Reviewer sQoH

Q1: Benchmark scale and large-scene evaluation.

A1: We thank the reviewer for this comment. Our evaluation follows the standard benchmark protocol of prior works (Instruct-GS2GS, DGE, EditSplat, and DreamCatalyst). In the revised manuscript, we have expanded our benchmark to 10 scenes and 40 editing tasks across five standard datasets, surpassing the scale of existing 3D editing benchmarks:

MethodVenueScenesTasks
CUESplat-1040
DreamCatalystICLR 2025840
CoreEditorTVCG 2026720
EditSplatCVPR 2025816
ViP3DEditAAAI 2026510
DGEECCV 2024310

The 12 newly added tasks include 2 large-scale urban street edits on 2 new scenes from the Waymo dataset, along with 2 large-geometry object replacements, 4 local appearance edits, and 4 global stylizations on existing scenes.

Evaluated under identical camera views across all 11 methods, CUESplat ranks first on every metric as shown in the table below:

MethodCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
CUESplat0.26970.17390.76223.080.2991
DGE0.23540.08900.74012.670.3231
GaussianEditor0.24250.10700.75451.830.3348
DreamCatalyst0.24140.10990.75231.920.3212
TrAME0.23450.08850.74422.250.3163
GaussCtrl0.24500.11030.73052.830.3286
CoreEditor0.22280.06640.72502.080.3254
Instruct-GS2GS0.23420.12860.71211.330.4808
EditSplat0.24240.12520.74162.420.3238
ViP3DEdit0.21920.05040.70721.580.3441
Omni-3DEdit0.22150.08330.71482.170.3569

Across both large-geometry replacements (turning the bear statue into a stone elephant) and street-scale Waymo edits (converting residential streets into snowy winter and autumn scenes), CUESplat effectively suppresses floating artifacts and preserves structural coherence across foreground targets and distant street geometry. By anchoring edited depths to source Gaussians and enforcing cross-view support, SEI preserves unedited background structures while localizing updates to supported regions. We have specified the dataset source of every scene in Table 1 and included all 40 tasks and visual comparisons in the revised manuscript.

Q2: Addition and removal metrics and ablations.

A2: We thank the reviewer for this suggestion. We have tested all baselines on both object addition and removal tasks:

MethodCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
CUESplat0.24200.11130.70501.830.2530
DGE0.21140.07260.67921.420.3586
GaussianEditor0.23210.08220.68491.330.2745
DreamCatalyst0.20970.06940.67781.330.3228
TrAME0.21080.06990.67931.420.3189
GaussCtrl0.21110.07150.68181.330.2967
CoreEditor0.20830.07220.68041.420.2880
Instruct-GS2GS0.20100.06530.67901.250.4133
EditSplat0.21250.07080.68911.420.3298
ViP3DEdit0.20070.06470.67251.330.3962
Omni-3DEdit0.22530.07080.66471.500.2869

Existing methods usually are forced to fit fixed source primitives to 2D image changes, either leaving unwanted geometry of removed objects or stretching background Gaussians to mimic added objects.

To further show how our three core modules resolve these structural bottlenecks, we conduct an ablation study on the object addition and removal tasks:

MethodCLIP-S ↑CLIP-D ↑DINO ↑Gemini ↑MEt3R ↓
CUESplat (Full)0.24200.11130.70501.830.2530
w/o Multi-view Edit Coupling (MEC)0.22500.10090.69331.500.2733
w/o Structural Edit Inference (SEI)0.22330.09670.69521.330.2849
w/o Selective Gaussian Adaptation (SGA)0.22150.08920.69311.500.2658

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 10.3% and reducing MEt3R by 7.4%), SEI extracts paired depth evidence to separate true structural changes from 2D appearance shifts (boosting Gemini quality by 37.6% and CLIP-D by 15.1%), and SGA executes local structural updates by removing selected Gaussians, reusing revealed surfaces, and initializing missing geometry (improving Gemini quality by 22.0% and CLIP-D by 24.8%).

We have updated Table 1(b) and included this structural ablation analysis in the revised manuscript.

Q3: Abstract clarity and Figure 1 visual example.

A3: We thank the reviewer for this suggestion. Existing render-edit-fit pipelines optimize Gaussians directly against 2D edited images without verifying 3D geometric support, which leaves floating unwanted geometry during structural edits.

In contrast, our core insight is a render-edit-verify-fit pipeline, which verifies whether a proposed 2D edit is geometrically supported across views before updating the 3D scene. Compared with existing pipelines, CUESplat can eliminate floating unwanted geometry and produce clean surfaces. We have revised the Abstract and updated Figure 1 and its caption accordingly in the revised manuscript.

Q4: Significance of metric improvements and multi-view Gemini evaluation.

A4: We thank the reviewer for the careful inspection. Because local 3D edits share unedited backgrounds, feature cosine metrics such as DINO operate within a narrow dynamic range where margins between leading methods are typically around 0.003, as reported in TINKER [1].

All evaluations use fixed random seeds for reproducible optimization, and to reduce viewpoint sampling variance in Gemini evaluation, we doubled the evaluated views from 4 to 8 across all 11 methods:

MethodGemini, 4 ViewsGemini, 8 Views
CUESplat3.133.25
TrAME2.313.06
GaussCtrl2.692.63
DGE2.692.56
CoreEditor2.132.50
EditSplat2.252.13
DreamCatalyst1.882.00
Omni-3DEdit1.881.94
GaussianEditor1.631.69
Instruct-GS2GS1.501.56
ViP3DEdit1.311.50

With 8 views per task, CUESplat improves from 3.13 to 3.25 and ranks first among all baselines, confirming consistent visual quality across viewpoints. We have included the 8-view evaluation in the revised manuscript.

[1] Zhao et al., TINKER: Diffusion’s Gift to 3D: Multi-View Consistent Editing From Sparse Inputs without Per-Scene Optimization, ICLR 2026.

Thank the authors for the rebuttal and the added experiments. After reviewing the responses, my concerns are not fully addressed, and I keep my current rating.

The fixed-seed setting supports reproducibility, but no mean ± std over multiple runs and no significance check are provided. With several gains still this small — for example DINO 0.7745 vs 0.7707 and ablation differences below 0.01 — it remains unclear whether the margins are real improvements or run-specific variation. The narrow dynamic range of DINO explains why gaps are small, but it does not replace statistical evidence. This concern about improvement significance still stands.

The addition/removal results are not consistent across tables: the rebuttal reports CLIP-D 0.1113 and Gemini 1.83 for the full model, while Table 1(b) gives 0.1721 and 3.17. The explanation about different task subsets is noted, but without per-task logs and a unified task scope, the ablation and the main table cannot be compared directly. The concern about result consistency and the reliability of the module contributions is therefore not removed.

Large-scene evidence is still limited to two Waymo street tasks, which is not enough to support a claim that the method generalizes to large scenes. The added experiments remain mostly object-centric, so performance on large-scale scenes is still under-supported.

For these reasons, the rebuttal does not fully address my concerns on soundness and evaluation reliability. I maintain my score of Weak rejection (2).

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

This manuscript presents CUESplat that addresses text-driven editing of 3D Gaussian scenes by using inferred geometric changes to guide scene updates. It proposes three main modules Structural Edit Inference (SEI), Selective Gaussian Adaptation (SGA), and Multi-view Edit Coupling (MEC). SEI compares source and edited depths and checks cross-view support; SGA then removes affected primitives, reuses existing surfaces, and initializes missing geometry; MEC exchanges editing increments using evolving geometric correspondences. It evaluates various editing tasks across scenes and reports improvements in semantic alignment and multi-view consistency.

Critical strengths
  1. The paper identifies that when fitting Gaussians to edited images, visually plausible supervision does not necessarily specify which existing surfaces should be preserved, removed, or reconstructed. This study is well motivated and sound.
  2. The proposed SEI and SGA reusing revealed surfaces before generating completion geometry is a sensible design for preserving existing content.
  3. Evaluations against prior art on visual and semantic metrics are thorough.
Critical weaknesses

The method itself is sound and looks nice; I have only two questions.

  1. In Figure 4, while the foreground object is edited, the background color is also enhanced (1st row, 4th row). However, the prompt does not mention editing the background.
  2. From Table 1-3, the advantages over existing methods and the ablations are marginal.
Questions for authors

See the ''Critical Weaknesses'' above.

3 comments
AuthorsOpenReview ↗

Response to Reviewer eaRV

Q1: Background color variations in Figure 4.

A1: We thank the reviewer for this observation. In CUESplat, SEI and SGA keep unselected background Gaussians fixed to prevent unintended primitive additions or removals. For appearance updates, SGA uses evidence-weighted fitting to concentrate optimization on editable foreground primitives while allowing slight appearance adjustments on retained boundary Gaussians with geometry fixed, keeping ambient tone consistent with the edited foreground without harsh seam artifacts.

Following the PIE-Bench evaluation protocol on held-out views, we evaluate all 11 methods without external SAM masks during editing:

MethodDiffusion BackbonePSNR-BG ↑MSE-BG ×104\times 10^4 ↓SSIM-BG ×102\times 10^2 ↑LPIPS-BG ×103\times 10^3 ↓CLIP-Edited ↑CLIP-Whole ↑
CUESplatDiT-based24.398121.591.34115.30.25130.2580
CoreEditorUNet-based23.374149.079.55118.60.22710.2354
DGEUNet-based22.753125.590.23129.50.22990.2401
DreamCatalystUNet-based22.667133.190.81117.60.22060.2372
EditSplatUNet-based23.071126.089.01119.30.23830.2483
GaussCtrlUNet-based23.420122.687.48116.60.22880.2421
GaussianEditorUNet-based23.850153.788.47117.30.22340.2329
Omni-3DEditDiT-based20.238156.789.51124.40.21310.2198
TrAMEUNet-based22.629134.289.35126.50.23150.2479
ViP3DEditUNet-based20.195167.589.74119.20.22730.2409
Instruct-GS2GSUNet-based17.228300.374.86204.20.22690.2390

CUESplat ranks first across all six metrics, achieving the highest background fidelity alongside the strongest foreground editability. We have included this background preservation evaluation in the revised manuscript.

Q2: Magnitude of quantitative improvements.

A2: We thank the reviewer for this question. Across eight representative 3D editing papers, relative CLIP-D gains over the best prior baseline average 14.9%, whereas CUESplat achieves a 33.0% improvement over the runner-up Instruct-GS2GS.

On the remaining primary metrics, CUESplat improves Gemini visual quality by 16.4% over the runner-up GaussCtrl and DGE, and reduces MEt3R geometric error by 1.1% over the runner-up Omni-3DEdit and by 48.8% over the closest render-edit-fit baseline EditSplat. On DINO, where shared backgrounds compress cross-view similarity, our 0.0038 gain over the runner-up DreamCatalyst matches typical 0.003 SOTA margins.

Integrating SEI and SGA into DGE and EditSplat also reduces MEt3R error by 25.1% and 32.5%, while ablating our core modules causes clear drops up to 14.1% in Gemini quality and 13.8% in CLIP-D.

Thank you for the clarification and the additional evaluation. My concern is not about boundary seams. In Figure 4, the appearance of background regions changes visibly: the tree and grass in the first row, and the leaves and grass in the fourth row. These changes occur away from the foreground object and are not requested by the prompts.

Keeping background Gaussian geometry fixed does not establish that their appearance attributes remain unchanged. Likewise, aggregate background metrics do not address these specific examples. I encourage the authors to clarify which background appearance attributes may be updated and to acknowledge unintended background color changes as a limitation, rather than relying on comparisons with other methods to resolve this concern.

We sincerely thank the reviewer for this constructive feedback. Below we analyze the mechanism that causes these background appearance updates and incorporate this discussion and visual analysis into the Limitations section of our revised manuscript.

In CUESplat, retained source Gaussians keep their geometric attributes (center μg\mu_g, scale, and rotation) fixed, whereas primitives whose compositing-weighted RGB difference (HidiH_i d_i) exceeds threshold are selected as editable and allowed to update their appearance coefficients cgc_g and opacity ogo_g. While this design supports valid appearance edits on static surfaces, it introduces a limitation when the 2D candidate editor produces global color shifts on unedited backgrounds, such as the trees and grass in Figure 4 (rows 1 and 4). Because SEI verifies 3D support through depth residuals, static background surfaces with pure color shifts exhibit zero depth displacement and do not trigger unsupported geometric masks. Consequently, when the color drift exceeds the RGB attribution threshold, SGA cannot distinguish these shifts from intended appearance-only edits and updates those background Gaussians during refinement.

Existing methods typically prevent background color drift by directly replacing background regions in the edited views with pre-edit source pixels prior to Gaussian fitting. As shown in Figure 6 (rows 3) in the revison manuscript, when applied to our method, it help to prevent background color drift. However, hard pixel replacement is a heuristic that bypasses the underlying coupling between 2D diffusion editing and 3D scene optimization. Decoupling intended 3D appearance edits from global 2D RGB drift remains a limitation of CUESplat and a direction for future work.

Following your suggestion, we have expanded Appendix F (Limitations and Failure Cases) to analyze this limitation and provide a visual analysis (Figure 6 in Appendix F) of the background color shifts.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.