acceptodds
Under review as a conference paper at ICLR 2027

TDAP-GS: Scale-Adaptive Tile-Based Density Accumulation and Pruning for Highly Efficient 3D Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) achieves high-quality novel-view synthesis with real-time rendering, but densification can produce redundant Gaussians that increase memory use and rendering cost. Existing compression methods may retain primitives in deeply occluded regions or high-frequency floaters, while hard pruning can degrade visual quality and require extensive post-pruning refinement. We propose Tile-Based Density Accumulation and Pruning (TDAP), which uses the tile-based rendering pipeline to estimate a scale-adaptive density for each Gaussian. TDAP then applies quantile-guided opacity modulation to progressively remove low-density candidates while preserving Gaussians in dense scene regions. It requires neither an auxiliary loss nor post-pruning optimization. Experiments show that TDAP reduces the number of Gaussians and rendering cost while maintaining, and in some cases improving, rendering quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting (3DGS), Novel View Synthesis, Pruning, Tile-based Density Evaluation

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_9d21OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Tile-Based Density Accumulation and Pruning (TDAP), which uses the tile-based rendering pipeline to estimate a scale-adaptive density for each Gaussian. TDAP then applies quantile-guided opacity modulation to progressively remove low-density candidates while preserving Gaussians in dense scene regions. The experimental results indicate that TDAP-GS can reduce the number of Gaussians, but cannot maintain comparable rendering performance.

Critical strengths

None

Critical weaknesses
  1. The paper's motivation is unclear. The author did not explain why the idea needs to be done in this way, only stating that it needs to be done in this way. The author lacks evidence to support the feasibility of the idea, and based on the experimental results reported, it is not entirely feasible.
  2. There are many issues with the writing and formatting of the paper. Some of the figures and tables in the text are not cited. There are text errors in some of the images (the vertical axis title of the image on the right side of Figure 5 should be LPIPS). The key figure 3 has no explanation and relevant experimental proof. In Section 4.2 of the text, write "we thoughtfully discuss the causes of this generalization bottleneck in section 4.2." and the limitation without any evidence. Write "our method continuously estimates density for each primitive with near-zero computational overhead" in the conclusion section, but there are no existing conclusions or experimental results to prove it.
  3. There is no introduction to the latest work in the relevant work, and there is no comparison between the recent work in the experimental part.
Questions for authors

See the weakness.

4 comments

We thank the reviewer for highlighting the need for a clearer motivation, stronger empirical justification, and more precise presentation. We will revise the manuscript to make our central contribution explicit and distinguish the evidence supporting it from the limitations of the current method.

Response to Weakness 1: Motivation and Empirical Feasibility

Existing Gaussian simplification methods use various importance criteria, including opacity, gradients, and rendering sensitivity. These criteria capture different aspects of a primitive's contribution. Our work investigates a complementary question: can scale-aware local geometric support provide a useful ranking signal for reducing Gaussian count while retaining reconstruction quality ? Opacity alone does not describe how a primitive relates to neighboring geometry, while positional gradients reflect reconstruction sensitivity rather than local geometric support. Our density term evaluates spatial proximity relative to Gaussian scale and aggregates this information across tile occurrences and training views. The resulting score combines local support with opacity and observation information.

The tile-based implementation follows from this objective. The rasterizer already assigns Gaussians to tiles and sorts them by depth. We reuse these lists to obtain a bounded set of candidate neighbors, then evaluate their spatial relationships without constructing additional spatial data structures or performing global neighbor searches. Tile membership and depth adjacency define the candidate set; they are not themselves treated as evidence of geometric proximity.

The experiments support the practical feasibility of this approach: TDAP-GS substantially reduces Gaussian count and training memory while maintaining competitive rendering quality, particularly on Mip-NeRF 360. However, its advantage is not uniform across datasets, and the configurations yielding the strongest memory reduction and the best rendering quality differ. We will state these distinctions explicitly rather than implying universal superiority or simultaneous gains from a single configuration.

To further examine the contribution of the ranking signal, we evaluated component ablations of the revised score:

MethodsM360 SSIM ↑M360 PSNR ↑M360 LPIPS ↓DB SSIM ↑DB PSNR ↑DB LPIPS ↓TaT SSIM ↑TaT PSNR ↑TaT LPIPS ↓
Full0.83427.8570.1960.90529.5210.2460.84923.7790.188
wo oio_i0.7826.6280.2660.90429.3870.2490.83323.3440.219
wo 1+e−βHi1+e^{-\beta H_i}0.81927.5750.2210.90629.6180.2440.84223.7660.203
wo density itself0.53918.9040.530.86825.3080.3170.60716.4420.478

Removing the local-support term substantially degrades reconstruction quality on all three datasets, indicating that it provides useful information within the evaluated pipeline. Removing opacity also reduces performance, supporting the use of a composite score. The observation factor improves results on Mip-NeRF 360 and Tanks and Temples but does not improve Deep Blending, demonstrating that its benefit is dataset-dependent.

These ablations support the contribution of local support, but do not establish that density alone is an optimal importance measure. We will distinguish these revised-score results from the original formulation and present direct comparisons with simpler ranking criteria under matched Gaussian budgets.

Response to Weakness 2: Presentation and Unsupported Claims

We acknowledge the presentation errors and the insufficient support for several claims. We will make the following corrections:

  • Figures, tables, and cross-references. We will correct the vertical-axis label in the right panel of Fig.~5 to ``LPIPS,'' remove the circular reference in Section~4.2, and ensure that every figure and table is explicitly referenced and explained in the text.

  • Interpretation of Fig.~3. Figure~3 shows the distribution of density scores at iteration 5K. It describes the score distribution at that training stage; it does not, by itself, establish which Gaussians are redundant or prove the effectiveness of pruning. We will explain its axes, measurement setting, and relationship to quantile-based selection, while using the component ablations above to assess the contribution of the score.

  • Limitations and causal explanations. The weaker results on Deep Blending and Tanks and Temples establish a limitation of the current method, but do not establish its cause. We will present explanations involving viewpoint variation or depth-sorted neighborhoods as hypotheses unless supported by controlled experiments.

  • Computational overhead. Our original claim of "near-zero computational overhead" is accurate. By reusing the rasterizer's pre-existing tile lists, hardware profiling (torch.cuda.Event) confirms our density accumulation adds merely ~0.15 ms per iteration. With identical Gaussians, camera views, and rendering settings, the measured times for this comparison are 6.287 ms and 6.432 ms.

MethodAverage Time (ms)
wo Density Accumulation6.287
w Density Accumulation6.432

Response to Weakness 3: Coverage of Recent Work

We agree that the positioning against recent methods needs to be clearer. The submitted manuscript already includes comparisons with LightGaussian, PUP-3DGS, Mini-Splatting, Taming-3DGS, MaskGaussian, and GaussianSpa, but we did not explain their relationship to our contribution sufficiently. We will reorganize the related-work discussion around the ranking signal, pruning stage, and computational requirements of each approach. This will clarify that our contribution is the investigation of density-based ranking through rasterization-native local-support evaluation, rather than a claim that previous methods rely exclusively on opacity or gradients.

We will also review relevant recent work, add missing references, and explain the selection of experimental baselines. Where implementations and evaluation settings permit fair reproduction, we will include additional comparisons during the discussion period. We will distinguish controlled comparisons using the same backbone from comparisons between complete systems.

I thank the authors for their response and helpful clarifications, which have addressed several of my concerns. However, I still feel that the novelty and performance gains are relatively modest. On balance, I will maintain my original score.

We appreciate your continued engagement and confirming that our clarifications have successfully addressed your technical concerns. We sincerely respect your perspective on the novelty and performance improvements. However, we humbly request that please reconsider the magnitude of our contributions within the broader context of recent 3DGS literature.

While the absolute metric improvements might appear modest at first glance, evaluating the performance delta (Δ\Delta) relative to the underlying backbones provides a clearer objective measure of our contribution. Specifically, we contextualize our relative gains against two recently accepted baselines on the Mip-NeRF 360 dataset: GaussianSpa (CVPR'25, relative to Mini-Splatting) and MaskGaussian (CVPR'25, relative to Compact3DGS).

MethodsΔ\DeltaSSIMΔ\DeltaPSNRΔ\DeltaLPIPS
GaussianSpa0.004 (0.821 →\rightarrow 0.825)0.45 (27.40 →\rightarrow 27.85)0.005 (0.219 →\rightarrow 0.214)
MaskGaussian0.006 (0.805 →\rightarrow 0.811)0.12 (27.32 →\rightarrow 27.44)0.007 (0.233 →\rightarrow 0.226)
TDAP-GS0.012 (0.822 →\rightarrow 0.834)0.396 (27.461 →\rightarrow 27.857)0.023 (0.219 →\rightarrow 0.196)

GaussianSpa reports a Δ\Delta PSNR of +0.45, and MaskGaussian reports a Δ\Delta PSNR of +0.12 over their respective backbones. In contrast, our method achieves relative improvement of Δ\Delta SSIM +0.012 ($\sim$2-3x), Δ\Delta PSNR +0.396 ($\sim$0.88-3.3x), and Δ\Delta LPIPS -0.023 ($\sim$3.3-4.6x). In the highly optimized domain of 3DGS, achieving this magnitude of relative gain—without relying on complex auxiliary losses or massive computational overhead—is a non-trivial, rigorously validated advancement.

As Reviewer 8qET explicitly acknowledged in his/her assessment, the core strength and primary contribution of our work lies in the significant decrease in the amount of Gaussians. By achieving up to 80% compression with near-zero (<3<3%) computational overhead, our method provides a highly efficient pruning mechanism. Delivering this magnitude of reduction while maintaining or slightly improving rendering quality is precisely the practical value we aim to contribute to the community. As also acknowledged by Reviewer 8DwB, this streamlined design translates directly into robust, proven effectiveness—most notably reducing training memory by 2 to 3 times compared to unpruned 3DGS—providing a highly efficient and accessible tool for the community.

Given that your initial technical concerns have been resolved, and our method's performance gains are objectively competitive when compared to the deltas of existing baselines, we humbly request that please reconsider your final evaluation, if possible. If you agree that a highly efficient pruning tool holds practical merit for the 3DGS community, we would be deeply grateful if you could consider raising the score.

Regardless of your final decision, we genuinely appreciate your constructive feedback, which has helped us thoroughly strengthen our paper.

Reviewer_8DwBOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

TDAP-GS proposed a tile-based density measure, a depth-ordered, scale-weighted statistic derived from the rasterizer’s tile-level information to rank Gaussian primitives for in-training pruning, and quantile-guided opacity modulation which replaces immediate removal of selected candidates with progressive opacity adjustment, allowing the representation to adapt before pruning. Effectiveness is assessed on primitive count, GPU memory, rendering cost, and novel-view quality, including the separate effects of its density measure and modulation strategy.

Critical strengths
  1. The implementation is elegant and it reuses the computation of existing renderers.
  2. It needs no extra loss tuning or network to work.
  3. Good shown effectiveness, e.g., training memory 2~3x lower than unpruned 3DGS
Critical weaknesses
  1. One of the major proposed novelties, the density score, is not tested on its own, such as its effectiveness and what kind of perspective it can offer. Otherwise the novelty can be incremental. Baseline 3DGS already prunes by opacity, and TDAP's score multiplies density by opacity and then pushes opacity up or down based on that score. After a few rounds, the ranking may be mostly an opacity ranking. Therefore, testing whether that score beats simpler options can offer more insights.
  2. No efficiency number reported other than memory. The paper claims "zero" or "near-zero" overhead and promises training-time and rendering-speed numbers, but none are reported. For memory, it covers one scene and compares only against an unpruned model, not the other pruning methods. It's also unclear whether the memory savings and the best quality numbers come from the same configuration (the authors may provide clarification on this).
  3. TDAP is built on Mip-Splatting while the baselines use vanilla 3DGS, and the unpruned Mip-Splatting result is never reported.
Questions for authors

Plz address the concerns and questions in the weakness section

3 comments

We thank the reviewer for identifying the need to isolate the contribution of the density score, substantiate our efficiency claims, and separate the effects of the rendering backbone from those of TDAP-GS. We address these concerns below.

Response to Weakness 1: Contribution of the Density Score

The motivation for incorporating local support is that opacity does not describe a Gaussian's spatial relationship with neighboring primitives. For example, a high-opacity but spatially isolated floater can receive weak local support, whereas a globally sparse thin structure may still exhibit coherent support among its constituent Gaussians. Nevertheless, local support remains a proxy for rendering importance rather than a definitive criterion. To examine the contribution of individual score components, we evaluated the following ablations of the revised score. Across variants, we made all the methods fixed for post-pruning, modulation policy, and retenting 20% of the final Gaussian counts on Mip-NeRF 360(M360), Deep Blending(DB),and Tanks and Temples(TaT).

MethodM360 SSIM ↑M360 PSNR ↑M360 LPIPS ↓DB SSIM ↑DB PSNR ↑DB LPIPS ↓TaT SSIM ↑TaT PSNR ↑TaT LPIPS ↓
Full0.83427.8570.1960.90529.5210.2460.84923.7790.188
wo oio_i0.7826.6280.2660.90429.3870.2490.83323.3440.219
wo 1+e−βHi1+e^{-\beta H_i}0.81927.5750.2210.90629.6180.2440.84223.7660.203
wo density itself0.53918.9040.530.86825.3080.3170.60716.4420.478

Removing the local-support term reduces PSNR by 8.953, 4.213, and 7.337 dB on Mip-NeRF 360, Deep Blending, and Tanks and Temples, respectively. Removing opacity or observation factor also degrades performance on all three datasets. These component ablations do not, by themselves, establish that the score consistently outperforms opacity-only ranking. To address that question directly, we conducted a controlled comparison between opacity-based ranking, gradient-based ranking, and the original TDAP-GS score. We changed only the ranking criterion while keeping the remaining pipeline fixed, including the initialization, pruning schedule, opacity-modulation policy, and optimization budget.

The below table reports PSNR, SSIM, LPIPS, and the actual retained Gaussian counts on three datasets. Under this matched budget, the original TDAP-GS score reduces PSNR by 0.776, 0.072, and 0.688 dB on Mip-NeRF 360(M360), Deep Blending(DB),and Tanks and Temples(TaT), respectively, relative to opacity-based ranking. Reduce PSNR by 11.68, 9.923, and 9.927 dB respectively relative to gradient-based ranking. We agree that multiplying local support by opacity, followed by repeated opacity modulation, creates a potential feedback effect. These results provide a direct assessment of the proposed score ranking remains distinct from opacity ranking.

MethodsM360 SSIM ↑M360 PSNR ↑M360 LPIPS ↓DB SSIM ↑DB PSNR ↑DB LPIPS ↓TaT SSIM ↑TaT PSNR ↑TaT LPIPS ↓
Gradient-based0.43215.8130.6230.76219.3450.4440.47913.6030.608
Opacity-based0.79226.7170.260.89629.1960.2650.80722.8420.245
Density-based0.82327.4930.2110.89729.2680.2650.83723.5300.200
New Density-based0.83427.8570.1960.90529.5210.2460.84923.7790.188

We additionally investigated whether gradient information provides a complementary signal. We augment the original per-view aggregation in Eq. (5) as follows:

ℓi,mview=1∣Ti,m∣∑t∈Ti,m∇gi,mℓi,m,t,\ell_{i,m}^{\mathrm{view}}=\frac{1}{|T_{i,m}|}\sum_{t\in T_{i,m}}\nabla g_{i,m}\ell_{i,m,t},

where ∇gi,m\nabla g_{i,m} is the gradient of the training loss with respect to the view-space position of Gaussian GiG_i in view mm. We retain the original tile averaging and use the same subsequent aggregation and scoring rules in Eqs. (6)–(7). This variant combines local support with a signal of reconstruction sensitivity, allowing Gaussians with larger positional gradients to receive higher scores than they would under the original formulation. We report the gradient-weighted variant separately in the above table. It reduces PSNR by 0.364, 0.253, and 0.249 dB on three datasets, relative to the original TDAP-GS score. The original-score comparison directly addresses the reviewer’s request, while this additional ablation examines whether gradient information complements our local-support metric.

Response to Weakness 2: Efficiency, Memory, and Configuration

Runtime and overhead. We thank the reviewer for requesting quantitative proof. Our original claim of "near-zero computational overhead" is accurate. By reusing the rasterizer's pre-existing tile lists, hardware profiling (torch.cuda.Event) confirms our density accumulation adds merely ~0.15 ms per iteration. With identical Gaussians, camera views, and rendering settings, the measured times for this comparison are 6.287 ms and 6.432 ms.

MethodAverage Time (ms)
wo Density Accumulation6.287
w Density Accumulation6.432

Peak training memory. The existing single-scene comparison is insufficient to establish consistent memory savings across datasets or superiority over other pruning methods. We will expand the evaluation to report per-scene peak allocated GPU memory, final Gaussian count, and reconstruction quality for TDAP-GS and the compared methods under a common hardware and measurement protocol.

Peak GPU Usage (GB)M360DBTaT
Mip-Splatting10.1899.1115.696
TDAP-GS5.9545.2223.085

Relationship between memory savings and reconstruction quality. The strongest memory reduction and the best reconstruction quality correspond to different pruning configurations, and we will make this distinction explicit:

  • Pre-Pruning operates during densification and targets peak training memory by limiting primitive growth. Because cloning and splitting remain active, the final Gaussian count is not determined solely by the pruning ratio.
  • Post-Pruning begins after densification ends at iteration 15,000, allowing the retained Gaussian count to be controlled more directly. This configuration produces the best reconstruction quality shown in Fig.~1, but cannot reduce the memory already consumed during the preceding densification phase.

We will report memory and quality for each configuration separately, rather than presenting them as benefits achieved simultaneously by a single setting.

Response to Weakness 3: Backbone-Matched Evaluation

The reviewer is correct that TDAP-GS uses Mip-Splatting as its backbone. Consequently, comparisons against vanilla 3DGS alone cannot separate the contribution of TDAP-GS from that of Mip-Splatting. An unpruned Mip-Splatting model must also be identified explicitly rather than labeled as vanilla 3DGS.

We will include a backbone-matched comparison of vanilla 3DGS, unpruned Mip-Splatting, and Mip-Splatting with TDAP-GS, reporting reconstruction quality, Gaussian count, peak training memory, and rendering speed. The difference between unpruned Mip-Splatting and Mip-Splatting with TDAP-GS will quantify the effect of our pruning framework. Comparisons with methods using other backbones will be labeled separately as comparisons between complete systems.

Number of Gaussians(M)M360DBTaT
3DGS3.1102.7801.830
Mip-Splatting3.9493.4842.359
TDAP-GS0.7420.6980.403

Dear Reviewer 8DwB,

As the author-reviewer interaction period is ending, please respond to the authors' feedback ASAP.

Best, AC

Reviewer_8qETOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

3D Gaussian Splatting (3DGS) suffers from redundant Gaussian primitives during densification, leading to high memory consumption and computational costs. To address this issue, the authors propose an in-training pruning method named TDAP-GS that estimates Gaussian density using depth-sorted tile lists and scale-adaptive spatial distances. The method further introduces quantile-guided opacity modulation and progressive stepwise pruning to remove redundant Gaussians. Experimental results show that TDAP-GS significantly reduces Gaussian counts and GPU memory usage while maintaining competitive rendering quality, particularly on Mip-NeRF 360.

Critical strengths

The proposed density evaluation reuses the existing tile-based rasterization pipeline, avoiding additional spatial structures or global neighbor searches. The method achieves significant reductions in Gaussian count and memory usage while maintaining competitive rendering quality, especially on Mip-NeRF 360.

Critical weaknesses

The method assumes that Gaussians with lower local density are less important and should be pruned. However, low-density Gaussians may represent important fine structures, while high-density regions may contain redundant Gaussians. Moreover, the ablation study does not directly compare the proposed density metric with simpler pruning criteria, such as opacity-based or gradient-based pruning.

Questions for authors

Could the authors provide a comparison between the proposed density metric and simpler pruning criteria, such as opacity-based or gradient-based importance, under comparable Gaussian budgets? If author could address this concern, I will consider to change the score.

3 comments

Response to the concern about low- and high-density Gaussians

Thank you for raising this concern. We agree that geometric density alone cannot fully determine rendering importance: sparse Gaussians may represent fine structures, whereas densely populated regions may contain redundant primitives. The motivation for our metric is that opacity alone does not capture whether a Gaussian is supported by neighboring geometry. In particular, an isolated floater can retain high opacity while receiving weak local geometric support, making it difficult to identify through opacity-based pruning alone.

Our score therefore incorporates scale-aware local support rather than simply counting nearby Gaussians. It evaluates spatial proximity relative to Gaussian scale, averages support across tiles and training views (Eqs. (4)–(6)), and combines this information with opacity and an observation-count factor (Eq. (7)). This distinction matters for fine structures: a thin structure may be globally sparse while still exhibiting coherent local support among its constituent Gaussians. Consequently, global sparsity does not necessarily translate into a low score under our metric. Conversely, high opacity alone does not ensure a high score when local support is weak.

Nevertheless, local support is a proxy rather than a definitive measure of rendering importance. An isolated but important Gaussian may still receive a low score, and redundant Gaussians within a well-supported region may receive high scores. Quantile-guided opacity modulation and progressive stepwise pruning are designed to mitigate reconstruction disruption and allow the remaining primitives to adapt between pruning events; they do not guarantee the preservation of every fine structure. We will clarify this distinction and the resulting limitation in the revised manuscript.

Comparison with opacity- and gradient-based criteria

To directly address the reviewer’s question, we conducted a controlled comparison between opacity-based ranking, gradient-based ranking, and the original TDAP-GS score. We changed only the ranking criterion while keeping the remaining pipeline fixed, including the initialization, pruning schedule, opacity-modulation policy, and optimization budget. All variants retained the same final Gaussian count per scene, corresponding to pruning 80% of the primitives, which means, retent 20% at the end.

The below table reports PSNR, SSIM, LPIPS, and the actual retained Gaussian counts on three datasets. Under this matched budget, the original TDAP-GS score reduces PSNR by 0.776, 0.072, and 0.688 dB on Mip-NeRF 360(M360), Deep Blending(DB),and Tanks and Temples(TaT), respectively, relative to opacity-based ranking. Reduce PSNR by 11.68, 9.923, and 9.927 dB respectively relative to gradient-based ranking. These results provide a direct assessment of the proposed score against simpler alternatives within the same pruning pipeline.

MethodsM360 SSIM ↑M360 PSNR ↑M360 LPIPS ↓DB SSIM ↑DB PSNR ↑DB LPIPS ↓TaT SSIM ↑TaT PSNR ↑TaT LPIPS ↓
Gradient-based0.43215.8130.6230.76219.3450.4440.47913.6030.608
Opacity-based0.79226.7170.260.89629.1960.2650.80722.8420.245
Density-based0.82327.4930.2110.89729.2680.2650.83723.5300.200
New Density-based0.83427.8570.1960.90529.5210.2460.84923.7790.188

As for the New Density-based method, please review the below section.

Additional ablation: incorporating gradient information

Motivated by the reviewer’s observation that local support may not fully capture the importance of fine structures, we additionally investigated whether gradient information provides a complementary signal. We augment the original per-view aggregation in Eq. (5) as follows: ℓi,mview=1∣Ti,m∣∑t∈Ti,m∇gi,mℓi,m,t,\ell_{i,m}^{\mathrm{view}}=\frac{1}{|T_{i,m}|}\sum_{t\in T_{i,m}}\nabla g_{i,m}\ell_{i,m,t}, where ∇gi,m\nabla g_{i,m} is the gradient of the training loss with respect to the view-space position of Gaussian GiG_i in view mm. We retain the original tile averaging and use the same subsequent aggregation and scoring rules in Eqs. (6)–(7). This variant combines local support with a signal of reconstruction sensitivity, allowing Gaussians with larger positional gradients to receive higher scores than they would under the original formulation. We report the gradient-weighted variant separately in the above table. It reduces PSNR by 0.364, 0.253, and 0.249 dB on three datasets, relative to the original TDAP-GS score. The original-score comparison directly addresses the reviewer’s request, while this additional ablation examines whether gradient information complements our local-support metric.

We appreciate the reviewer’s suggestion, which motivated this controlled comparison and helped us clarify the scope and limitations of the proposed score.

I appreicate the author's detailed reponse. The new comparisons have addressed my concern. I acknowledge the proposed method provides pratical benefits. However, 3DGS optimization has been extensively studied, I still find the novelty and significance of this work somewhat limited. I consider this work to have some merit, and I would be willing to slightly increase my score if a finer-grained rating were available. However, given the current rating scale, I will maintain my original score.

We appreciate your continued engagement and valuable time in reviewing our rebuttal. We are truly encouraged by your statement that our new comparisons have successfully addressed your concerns, and we are deeply grateful for your acknowledgment of the practical benefits and merit of our proposed method.

While the absolute metric improvements might appear modest at first glance, evaluating the performance delta (Δ\Delta) relative to the underlying backbones provides a clearer objective measure of our contribution. Specifically, we contextualize our relative gains against two recently accepted baselines on the Mip-NeRF 360 dataset: GaussianSpa (CVPR'25, relative to Mini-Splatting) and MaskGaussian (CVPR'25, relative to Compact3DGS).

MethodsΔ\DeltaSSIMΔ\DeltaPSNRΔ\DeltaLPIPS
GaussianSpa0.004 (0.821 →\rightarrow 0.825)0.45 (27.40 →\rightarrow 27.85)0.005 (0.219 →\rightarrow 0.214)
MaskGaussian0.006 (0.805 →\rightarrow 0.811)0.12 (27.32 →\rightarrow 27.44)0.007 (0.233 →\rightarrow 0.226)
TDAP-GS0.012 (0.822 →\rightarrow 0.834)0.396 (27.461 →\rightarrow 27.857)0.023 (0.219 →\rightarrow 0.196)

GaussianSpa reports a Δ\Delta PSNR of +0.45, and MaskGaussian reports a Δ\Delta PSNR of +0.12 over their respective backbones. In contrast, our method achieves relative improvement of Δ\Delta SSIM +0.012 ($\sim$2-3x), Δ\Delta PSNR +0.396 ($\sim$0.88-3.3x), and Δ\Delta LPIPS -0.023 ($\sim$3.3-4.6x). In the highly optimized domain of 3DGS, achieving this magnitude of relative gain—without relying on complex auxiliary losses or massive computational overhead—is a non-trivial, rigorously validated advancement.

We completely respect your feedback on novelty. However, the novelty of our approach lies in its highly elegant implementation that uniquely reuses the computation of existing renderers, allowing it to function effectively without the need for complex auxiliary loss tuning or additional networks. As also acknowledged by Reviewer 8DwB, this streamlined design translates directly into robust, proven effectiveness—most notably reducing training memory by 2 to 3 times compared to unpruned 3DGS—providing a highly efficient and accessible tool for the community. We humbly request that please reconsider the magnitude of our contributions within the broader context of recent 3DGS literature.

Regardless of your final decision, we would like to thank you for your constructive feedback, which has significantly helped us improve the quality of our paper.

Comments on the paper

2 comments
Area_Chair_AU3COpenReview ↗

rebuttal and feedback

Dear Reviewers,

Thank you for your efforts so far. The authors have responded to your reviews, please check them carefully (you may also check other reviewers' comments) and provide your feedback.

Best, AC

AuthorsOpenReview ↗

General Response

We thank the reviewers for their comments. All feedback has undeniably strengthened our manuscript. Below we list all the concerns and answers of our proposed method, and provide a short overview of this paper.

  1. Proposed Method: The motivation for incorporating local support is that opacity does not describe a Gaussian's spatial relationship with neighboring primitives. Based on this idea and motivated by the reviewer’s observation that local support may not fully capture the importance of fine structures, we additionally investigated whether gradient information provides a complementary signal. We augment the original per-view aggregation in Eq. (5) as follows: ℓi,mview=1∣Ti,m∣∑t∈Ti,m∇gi,mℓi,m,t,\ell_{i,m}^{\mathrm{view}}=\frac{1}{|T_{i,m}|}\sum_{t \in T_{i,m}}\nabla g_{i,m}\ell_{i,m,t}, where ∇gi,m\nabla g_{i,m} is the magnitude of the gradient of the training loss with respect to the view-space position of Gaussian GiG_i in view mm. Furthermore, to validate the design rationale of our method works, we add new figures that visualize the pruned points and remained points in the training step to the supplementary material. We hope the visualization can prove our concept of removing floater gaussians, and keep the fine structure ones to maximize the correctness.

  2. Our Strengths and Resolved Concerns: We are highly encouraged that these targeted empirical efforts have successfully cleared the technical hurdles. Throughout the discussion phase, reviewers have explicitly confirmed that these new comparisons have addressed their concerns. Furthermore, reviewers have acknowledged the practical benefits and solid merit of our method, specifically highlighting the significant decrease in the number of Gaussians (achieving 80% compression with <3<3% overhead) as a core strength. Building on this extreme efficiency, our rigorous controlled ablations further demonstrate that our geometric scale-aware local support fundamentally outperforms standard Opacity- and Gradient-based pruning heuristics. Additionally, our method achieves highly competitive relative performance gains that significantly exceed the relative improvements reported by recently accepted baselines (e.g., $\sim$2-3x Δ\Delta SSIM and $\sim$3.3-4.6x Δ\Delta LPIPS performance gain on Mip-NeRF 360 compared with the baselines). Our approach provides a rigorously validated and highly extensible solution for 3DGS, entirely without relying on complex auxiliary losses or massive computational overhead.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.