acceptodds
Under review as a conference paper at ICLR 2027

ELGD-GS: Error-Guided Localization and Gradient Consistency-Aware Density Control for Efficient Dynamic Scene Reconstruction

Abstract

4D Gaussian Splatting (4DGS) delivers high-quality dynamic scene reconstruction with real-time rendering. However, existing 4DGS methods adopt densification strategies that allocate Gaussians inefficiently. They over-populate well-reconstructed regions and ignore the influence of gradient consistency, leading to unnecessary storage overhead and suboptimal reconstruction quality. In this paper, we propose ELGD-GS, an Error-guided Localization and Gradient Consistency-aware Density Control of 4D Gaussian Splatting framework that enables high-quality dynamic scene reconstruction with efficient Gaussian representations. First, we introduce an error-guided Gaussian localization mechanism to identify 4D Gaussians requiring densification based on pixel-wise reconstruction errors. By recording pixel-wise photometric residuals and high-frequency structural discrepancies over time, this mechanism effectively prevents redundant Gaussian growth in well-reconstructed regions. Second, we propose a gradient consistency-aware density control strategy that measures gradient direction consistency across timestamps to regulate Gaussian growth. It facilitates the effective splitting of Gaussians affected by conflicting gradients while suppressing unnecessary densification of Gaussians with highly consistent gradient directions, thereby improving structural fidelity and representation efficiency. Experiments across multiple datasets demonstrate that ELGD-GS achieves substantial improvements in reconstruction quality and highly compact representations, outperforming the state-of-the-art method TAD-GS by a 0.42 dB PSNR gain and a 58% storage reduction on the N3DV dataset. The source code will be released at http://github.com.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
4D Gaussian Splatting, Dynamic Scene Reconstruction, Gaussian Densification

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_rMHyOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper investigates the efficiency of Gaussian densification in dynamic scene reconstruction with 4D Gaussian Splatting (4DGS). Based on the TAD-GS framework of the deformation field route, it proposes a two-stage densification optimization scheme. The first stage is an error-guided Gaussian localization mechanism, which identifies regions requiring densification through reconstruction errors and suppresses redundant Gaussian growth in well-reconstructed areas. The second stage is a gradient consistency-aware density control strategy, which bidirectionally adjusts the cloning and splitting thresholds of Gaussians by calculating the consistency of gradient directions across timestamps, to alleviate the problems of gradient collision and redundant growth from aligned gradients.

From the perspective of technical origin, the core ideas of both modules have research foundations in the field of static 3DGS. The idea of error-guided densification has been discussed in related work at ECCV 2024, and the core logic of gradient consistency-aware density control is similar to that of the GDAGS work at ICLR 2026. The core positioning of this paper is to integrate two independent 3D densification optimization ideas and extend them completely to 4D dynamic scenes for the first time, which represents an application-oriented incremental innovation. It should be clarified that this paper does not belong to the direction of 4DGS model compression, but focuses on densification strategy optimization during training. Its goal is to achieve higher reconstruction quality with fewer Gaussian primitives through more reasonable Gaussian allocation.

Strengths

1, it achieves the first complete extension of two densification paradigms to 4D scenarios. It adapts the two densification optimization ideas, error guidance and gradient consistency, from static 3D to 4D spatiotemporal scenes. Through error accumulation across time windows and gradient consistency statistics across timestamps, it mitigates the problems of densification redundancy and structural blur caused by motion blur and temporal inconsistency in dynamic scenes, filling part of the gap in 4D dynamic densification optimization.

2, it constructs a systematic two-stage serial densification pipeline. It integrates two originally independent optimization ideas in 3D into a two-stage architecture: error region screening followed by gradient consistency adjustment. The error mask first eliminates gradient signals in well-reconstructed regions from the source, greatly reducing invalid computations. Then precise gradient consistency adjustment is performed on the remaining high-error regions. The two modules produce complementary gains, achieving a better balance between efficiency and quality compared with any single module. This systematic integration is a clear increment of this work relative to existing 3D studies.

3, targeted detailed improvements are made to the error module. A Laplacian structural term is introduced into the reconstruction error calculation, which enhances the sensitivity to under-reconstructed edges and high-frequency details, and can locate structurally blurred regions more accurately than pure pixel photometric errors. Meanwhile, a 16×16 patch-level voting error screening mechanism is adopted, which suppresses false triggers caused by isolated pixel noise and improves the reliability of densification localization.

4, the gradient consistency module is adapted to the spatiotemporal dimension. Compared with single-frame spatial consistency calculation in 3D scenes, this paper implements weighted spatiotemporal joint consistency statistics combined with the error mask. On the one hand, it excludes irrelevant gradient interference from low-error regions; on the other hand, it accumulates direction information across time windows, which is more compatible with the motion characteristics of dynamic Gaussians and has a better effect on alleviating gradient collision in dynamic scenes.

5, the experimental verification is relatively sufficient. The paper takes TAD-GS as the direct main baseline, and controls variables such as training schedule, optimizer, and resolution, ensuring fair comparison. Ablation experiments systematically verify the independent contributions of the two modules and their sub-components. On multiple mainstream datasets, it achieves simultaneous improvement in reconstruction quality and reduction in Gaussian count and storage volume, and the results are highly credible.

Limitations and Weaknesses

1, the originality of the core ideas is relatively limited. The basic logic and core formulas of both core modules have been proposed in static 3DGS scenarios. The contributions of this paper are mainly reflected in scenario extension and engineering integration, with relatively weak breakthroughs at the basic theoretical level, which is a typical incremental improvement.

2, the discussion on scenario adaptability is insufficient. The paper does not analyze scenarios with different characteristics. For example, whether the performance gain of the method differs between high-speed motion scenarios and slow-motion scenarios, or between scenarios dominated by static backgrounds and scenarios with a large proportion of dynamic regions. Whether there are regular patterns for the optimal settings of hyperparameters such as error thresholds and consistency thresholds across different scenarios also lacks discussion and verification.

3, the selection of key hyperparameters lacks theoretical support. Key parameters such as the 16×16 patch size, the 0.7 threshold for high-error pixel proportion within a patch, and the start and end values of the progressive quantile are all set empirically, without further theoretical explanation or adaptive adjustment strategies, leading to insufficient support for the generalizability of the method.

4, cross-architecture comparison is insufficient. The comparison methods are mainly concentrated on the deformation field route, lacking horizontal comparison with benchmark methods of native 4DGS, making it difficult to fully reflect the competitiveness of the method in the entire 4DGS field.

Questions and Suggestions

1, regarding the source of comparison data. For all comparison methods in Table 1 except the main baseline TAD-GS, are the metric data obtained by the authors through unified reproduction under exactly the same training and evaluation protocols, or directly cited from the original papers of each method? Please clarify the acquisition approach of the comparison data to confirm the fairness of the horizontal comparison.

2, regarding the selection of structural error operators. The Laplacian operator used in error calculation is relatively sensitive to image noise. In real-world scenarios with acquisition noise, it may cause deviation in error localization. We wonder whether the authors have considered using edge detection operators that are more robust to noise, such as the Sobel operator, to replace the Laplacian operator and improve the stability of structural error detection.

3, regarding the generalizability verification of the densification strategy. The core contribution of the paper focuses on the improvement of the densification strategy, and the experimental verification of this part can be more rigorous. Current comparisons are mostly horizontal comparisons with complete methods of different architectures. Since only the densification component is modified, it is recommended to adopt a control-variable approach: fix all other training settings and model components, and only replace the densification module in different baseline methods with the scheme proposed in this paper, then compare the performance changes. This can more clearly prove that the performance gain indeed comes from the densification strategy itself rather than other architectural differences, and also better reflects the general value of this module.

4, regarding the suggestion on core metric emphasis. The core theme of this paper is densification strategy optimization, and the most direct effect of densification is to optimize the number and spatial distribution efficiency of Gaussians. Among current results, the PSNR improvement of a few tenths of a decibel is relatively heavily affected by factors such as random initialization of training and differences in implementation details, and its supporting strength as the core contribution is limited alone. In contrast, the substantial changes in Gaussian count and storage volume are the most direct and stable reflection of the densification strategy. It is suggested that the authors focus more discussion on the control effect of Gaussian count, and supplement the analysis of Gaussian count changes in different types of scenarios and different regions. This will be more consistent with the core positioning of densification optimization and make the practical value of the method clearer.

Critical strengths

Though the core ideas of both modules are rooted in existing static 3DGS research, this work is the first to integrate them into a sequential densification pipeline and fully extend the framework to 4D dynamic scenes. This constitutes a valuable incremental innovation with notable practical and engineering reference value for the field.

Critical weaknesses

None

Questions for authors
  1. Clarification: For the comparison methods listed in Table 1 except TAD-GS, were all metrics reproduced under exactly the same training and evaluation protocols (identical resolution downsampling, test split, and rendering settings), or were they directly cited from the original papers? Please clarify the source of these comparison results to confirm the fairness of the cross-method comparison.
  2. Suggestion: The Laplacian operator used for structural error detection is sensitive to image noise. It would be worth exploring whether more noise-robust edge operators such as Sobel could serve as an alternative, to improve the stability of error localization in real-world noisy acquisition scenarios.
  3. Requested additional experiment: As the core contribution of the paper lies in the densification strategy itself, please consider adding a control-variable validation experiment. With all other model components and training configurations fixed, replace the densification module in at least one other mainstream baseline (such as Swift4D) with the proposed strategy, and report the resulting performance. This would more firmly verify that the performance gains come from the densification strategy itself rather than other architectural differences, and demonstrate the cross-framework generalizability of the method, which is essential for evaluating the significance of the core contribution.
  4. Suggestion: Given the paper's focus on densification optimization, it would be more aligned with the core theme to center more discussion on Gaussian count and storage reduction rather than the small PSNR improvement, which can be easily affected by random initialization and implementation details. Further analysis of how Gaussian count varies across different scene types and regions would also better highlight the practical value of the densification strategy.
2 comments
AuthorsOpenReview ↗

Response For rMHy

1.We thank the reviewer for raising this important question. We clarify that the results of TAD-GS, Ex4DGS, Swift4D, and 4D-Scaffold in Table 1 are obtained by independently running their official implementations, rather than directly citing results from their original papers. In contrast, the STG results are taken from its original paper. For all reproduced methods, we adopt consistent evaluation protocols, including image resolution, test-view selection, and evaluation metrics, while following each method's official training configuration.

2.We thank the reviewer for this valuable suggestion. To evaluate the effect of the error operator, we replace the Laplacian operator with Sobel on N3DV while keeping all other components and training settings unchanged. The results show that Sobel achieves comparable performance, with only marginal differences in reconstruction quality and Gaussian count.

MethodError OperatorPSNR ↑SSIM ↑Number ↓
ELGD-GS (Ours)Laplacian32.840.949347K
ELGD-GSSobel32.810.949344K

3.We thank the reviewer for this valuable suggestion. To evaluate cross-framework generalizability, we integrate ELGD-GS densification into Swift4D, replacing only its original densification module while keeping all other model components, initialization, and training configurations unchanged. As shown below, our strategy improves reconstruction quality while reducing the Gaussian count, demonstrating its effectiveness beyond the original TAD-GS backbone.

MethodPSNR ↑SSIM ↑LPIPS ↓Gaussian Count ↓
Swift4D (Original)32.120.9440.104354K
Swift4D + ELGD-GS Densification32.380.9460.102296K

4.We agree that Gaussian count and storage reduction are more direct measures of densification efficiency. Accordingly, we have expanded our analysis to emphasize these aspects across datasets and spatial regions. Gaussian Count and Storage Reduction Across Datasets. We report the Gaussian count and storage consumption on N3DV and InterDigital. Compared with TAD-GS, ELGD-GS reduces the Gaussian count by 51.26% on N3DV and 52.87% on InterDigital, demonstrating consistent densification efficiency across datasets.

DatasetMethodNumber ↓Storage (MB) ↓
N3DVTAD-GS712255
N3DVELGD-GS (Ours)347106
InterDigitalTAD-GS1,604575
InterDigitalELGD-GS (Ours)756175

Densification Decisions Across Static and Dynamic Regions. We further analyze cloning and splitting decisions separately in static and dynamic regions across six N3DV scenes. As shown below, ELGD-GS consistently reduces both types of densification decisions in both regions.

RegionTAD-GS Clone (K)Ours Clone (K)Clone Reduction (%)TAD-GS Split (K)Ours Split (K)Split Reduction (%)
Static3,110.982,124.7431.704,350.641,771.5059.28
Dynamic1,257.61911.2727.542,137.72948.0755.65
Total4,368.593,036.0130.506,488.362,719.5858.09

Overall, ELGD-GS reduces cloning decisions by 30.50% and splitting decisions by 58.09%. These consistent reductions across static and dynamic regions further demonstrate the effectiveness of our densification strategy.

Reviewer_rMHyOpenReview ↗

Response to the authors' rebuttal

Thanks, the authors have addressed all my concerns, and I have no further questions.

Reviewer_nQm2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper changes how 4D Gaussian Splatting decides where to add Gaussians during training. Standard densification accumulates positional gradients per Gaussian and clones or splits whatever crosses a threshold. The authors argue that this keeps adding Gaussians in regions that are already well reconstructed and ignores whether a Gaussian's gradients agree with each other over time. So they add two things on top of TAD-GS. First, an error map built from photometric and edge residuals decides which Gaussians are allowed to accumulate gradient at all. Second, a ratio of signed to absolute gradient magnitude, aggregated over timestamps, scales the clone and split scores so that Gaussians with consistent gradients grow less and Gaussians with conflicting gradients split more. On three multi-view dynamic datasets this gives about half a dB over TAD-GS with roughly half the primitives.

Critical strengths

(1) The empirical picture is consistent and useful in practice. Against the same TAD-GS backbone the method halves the primitive count on every scene while PSNR, SSIM and LPIPS all improve, and because the paper also reports PSNR on dynamic regions only, we can see that the gain is where 4DGS methods usually struggle rather than just in the static background.

(2) The method only changes the densification signal and adds no learned module, so it should carry over to other explicit 4D backbones. The ablation also shows the two components are complementary, in that the error gate is responsible for most of the reduction in primitives while the consistency control is responsible for most of the quality gain.

Critical weaknesses

(1) The headline storage reduction is confounded, and although the paper mentions the cause in one sentence it never addresses it. The appendix says that on two of the three datasets a 3D hash table is used to encode some of the Gaussian parameters, which reduces storage on its own. If you divide reported storage by primitive count, it becomes clear the baseline did not get this. TAD-GS stores the same number of bytes per Gaussian on all three datasets, and so does the proposed method on the one dataset without the hash table, whereas on the other two its bytes per Gaussian drop. The same pattern shows up inside the ablation table. Now, a densification rule changes how many Gaussians you have, not how many bytes each one costs, so the extra saving on those two datasets must be the compression codec, applied to one side of the comparison only. On the main dataset something like a sixth of the headline saving comes from the codec rather than from the method. In addition, the FPS and timing columns have no GPU attached, so they cannot be interpreted.

(2) The paper also never shows what the two mechanisms actually do during training, and as a result the gains cannot be attributed to the stated logic rather than to simply growing fewer Gaussians. The densification analysis in Section 5.4 consists of three error mask snapshots of one scene and one picture of Gaussian ellipsoids, and that is all. There is no measurement of what fraction of Gaussians the error gate excludes, no distribution of the consistency score, and no count of how many clone or split decisions the two consistency thresholds actually change, which leaves those thresholds unmotivated. The ablation even hints that one of them does little, because removing the suppression of high consistency Gaussians barely moves PSNR. Moreover, the hyperparameters that set how much of the scene the mechanisms touch, namely the quantile schedule of the error map, the patch threshold, the iteration at which consistency control turns on and its strength, are not ablated at all. And the one control that would settle the question is absent, which is a TAD-GS run with its own densification threshold simply raised until it reaches the same primitive count. Fewer primitives often improve PSNR by themselves, so without that run the quality gain cannot be separated from growing less.

(3) Both components have close precedents in static scene densification. Error guided growth appears in Revising Densification, the signed to absolute gradient ratio with high and low thresholds appears in AbsGS and GDAGS, and in fact the consistency equations here are structurally very close to GDAGS. The paper never says what is new beyond aggregating the ratio over observations, and there is no ablation that ports those static rules into TAD-GS without temporal aggregation to show that the temporal part helps. There is also a mismatch between what the consistency score is said to measure and what it computes. Because it takes the norm of each timestamp's gradient before summing over time, gradients pointing in opposite directions at different times do not cancel, which means it measures within frame agreement averaged over time rather than consistency across time. Finally, the compact dynamic methods the paper itself cites in related work are missing from the tables even though compactness is the central claim.

Questions for authors

(1) Were the baselines trained with the same hash table encoding? Please report storage and FPS without it, and name the GPU used.

(2) Please report the fraction of Gaussians gated out by the error map over training, a histogram of the consistency score, and the number of clone and split events changed relative to the TAD-GS rule, split by static and dynamic regions.

(3) Please add a TAD-GS run with its densification threshold raised to match your primitive count, and ablate the iteration at which consistency control turns on and its strength.

(4) Please add an ablation that applies AbsGS or GDAGS style density control inside TAD-GS using per iteration statistics only, so that the benefit of temporal aggregation is isolated.

(5) Given that the consistency score sums per observation norms, in what sense does it capture consistency across time? Would summing the signed gradients over time first behave differently?

(6) Report the per iteration overhead of the error map, the quantile computation and the absolute gradient accumulation, plus peak memory, at a matched primitive count. The timing columns suggest the overhead is absorbed by the smaller model, but that is not shown.

(7) Were the baseline numbers re-run or copied from the papers? The TAD-GS averages differ slightly between the main table and the appendix.

6 comments
AuthorsOpenReview ↗

Response to Reviewer nQm2 (1/3)

We sincerely thank the reviewer for the insightful and constructive comments and we provide detailed responses to address these concerns point by point below, with additional experiments and clarifications.

Q1 & Weakness1 :Hash Encoding and Rendering Efficiency.

We thank the reviewer for raising this concern. Unlike TAD-GS, our ELGD-GS implementation uses 3D hash-grid encoding for selected static Gaussian attributes. To isolate its effect, we additionally evaluate ELGD-GS without hash-grid encoding while keeping the density-control strategy unchanged:

MethodHash GridNumber↓Storage↓FPS↑
TAD-GS✗712K255MB146
ELGD-GS✗347K124MB171
ELGD-GS✓347K106MB168

To further clarify the storage contribution of hash-grid encoding, we provide a detailed storage breakdown to isolate the contributions of Gaussian attributes, hash-grid parameters, and the decoder. The hash-grid and decoder introduce fixed storage overhead independent of the Gaussian count, explaining the differences in average per-Gaussian storage across scenes.

Storage ComponentWithout Hash GridWith Hash Grid
Hash-grid-replaced Gaussian Attributes66.2 MB0 MB
Remaining Per-Gaussian Attributes57.8 MB57.8 MB
Hash Grid Parameters0 MB48.1 MB
Decoder MLP Parameters0 MB0.09 MB
Total Storage124 MB106 MB

Q2 & Weakness2 : Error Mask Statistics and Density-Control Analysis.

We thank the reviewer for this valuable suggestion. To better understand the proposed densification mechanisms, we conduct additional quantitative analyses on six scenes of N3DV, examining the Error Mask, consistency-score distribution, and resulting densification decisions.

(1) Fraction of Gaussians Gated Out by the Error Mask. We define the Error Gated Ratio as the percentage of valid Gaussian-image observations excluded by the Error Mask:

Rgate=NgatedNcandidate×100%R_{\mathrm{gate}}=\frac{N_{\mathrm{gated}}}{N_{\mathrm{candidate}}} \times 100\%, where $N_{\mathrm{gated}}$ and $N_{\mathrm{candidate}}$ denote the numbers of gated and valid candidate Gaussian-image observations, respectively. The overall gated ratios across the six scenes of N3DV are:

SceneCoffee_MartiniCook_SpinachCut_Roasted_BeefFlame_Salmon_1Flame_SteakSear_SteakMean
Error Gated Ratio (%)30.9337.3036.8733.5935.5435.6834.98

These ratios are computed over the entire recorded training period using the total numbers of gated and valid candidate observations. We further examine the evolution of the gating ratio during training. The following values represent the arithmetic mean across six scenes, computed within each 200-iteration accumulation window:

Iteration2K4K6K8K10K12K14K16K
Mean Error Gated Ratio (%)15.7019.5324.2229.8134.8640.9646.2752.46

We have also added a plot showing the evolution of the gating ratio during training across all six N3DV scenes in Fig. 8 (page 17) of the revised manuscript.

(2) Distribution of the Consistency Score. We collect all valid consistency scores during the active consistency-control period across the six scenes of N3DV. Each observation corresponds to one Gaussian at one densification operation; thus, the counts represent accumulated Gaussian-step observations rather than unique Gaussians. We have also added histograms showing the distribution of consistency scores for each N3DV scene in Fig. 9 (page 17) of the revised manuscript.

Consistency ScoreNumber of Observations (K)Percentage (%)
[0.0, 0.2)3,825.444.09
[0.2, 0.8)82,558.5288.35
[0.8, 1.0]7,063.117.56
Total93,447.06100.00

(3) Changes in Cloning and Splitting Decisions. We compare ELGD-GS with the original TAD-GS under identical experimental settings, changing only the densification strategy. Parent Gaussians are classified into static and dynamic regions based on their projected locations in the test-view masks. The resulting cloning and splitting statistics are reported below (in thousands of Gaussians,K):

RegionTAD-GS CloneOurs CloneClone Decisions ReducedTAD-GS SplitOurs SplitSplit Decisions Reduced
Static3,110.982,124.74986.244,350.641,771.502,582.62
Dynamic1,257.61911.27346.342,137.72948.071,195.85
Total4,368.593,036.011,332.596,488.362,719.583,778.47

Compared with TAD-GS, ELGD-GS reduces cloning and splitting decisions by 30.50% and 58.09%, respectively. The reductions are observed in both regions:

  • Static regions: 31.70% fewer cloning decisions and 59.28% fewer splitting decisions.
  • Dynamic regions: 27.54% fewer cloning decisions and 55.65% fewer splitting decisions.
AuthorsOpenReview ↗

Response to Reviewer nQm2 (2/3)

Q3 & Weakness 2: Density-Matched Comparison and Consistency-Control Ablations.

(1) Density-Matched TAD-GS Baseline. To examine whether the improvements stem merely from reducing the Gaussian count, we conduct experiments on N3DV and increase the TAD-GS densification threshold to obtain a comparable Gaussian count while keeping other settings unchanged. We further add visual comparisons in Fig. 7 (page 9) of the revised manuscript.

MethodNumber↓PSNR↑SSIM↑
TAD-GS712K32.440.947
TAD-GS (Matched)347K32.010.944
ELGD-GS (Ours)347K32.840.949

(2) Activation Iteration. We vary the iteration at which consistency control is activated while keeping its modulation strength and all other configurations unchanged.

Activation IterationPSNR ↑SSIM ↑Number ↓
4,00032.590.947334K
6,00032.760.948342K
8,000 (Ours)32.840.949347K
12,00032.720.947359K

(3) Consistency Modulation Strength. We investigate different modulation strengths by introducing a scaling factor λ\lambda to control the magnitude of consistency-based densification modulation, while keeping the activation iteration and all other configurations fixed.

λ\lambdaPSNR↑SSIM↑Number↓
0.232.630.947356K
0.5 (Ours)32.840.949347K
0.832.670.947362K

Q4 & Weakness 3:Comparison with 3D Density-Control Strategies.

To investigate the effectiveness of temporal aggregation, we implement AbsGS-style and GDAGS-style density control within the TAD-GS framework and conduct additional experiments on N3DV.

(1) GDAGS-style Density Control.

GDAGS originally computes gradient consistency by aggregating pixel-wise sub-gradients across views. We extend this strategy to dynamic 4DGS by computing a consistency score independently at each training iteration using the sampled timestamp tkt_k, and then averaging the per-iteration scores:

Cavg=1K∑k=1K∥∣gi(tk)∥∣2+ϵ∥∣g~i(tk)∥∣2+ϵ.C_{\mathrm{avg}} = \frac{1}{K}\sum_{k=1}^{K}\frac{\||\mathbf g_i^{(t_k)}\||_2+\epsilon}{\||\tilde{\mathbf g}_i^{(t_k)}\||_2+\epsilon}.

In contrast, ELGD-GS first accumulates the L2 norms of signed and absolute sub-gradient sums across observations and computes their ratio afterward Cours=A/BC_{\mathrm{ours}}= A/B, where

A=∑k=1K∥∣gi(tk)∥∣2+ϵA=\sum_{k=1}^{K}\||\mathbf{g}_i^{(t_k)}\||_2+\epsilon,

B=∑k=1K∥∣g~i(tk)∥∣2+ϵB=\sum_{k=1}^{K}\||\tilde{\mathbf{g}}_i^{(t_k)}\||_2+\epsilon.

This aggregation strategy allows observations with larger gradient magnitudes to contribute more significantly to the final consistency estimation, rather than assigning equal importance to the consistency scores of individual observations.

(2) AbsGS-style Density Control. Following AbsGS, we construct an alternative variant by directly accumulating absolute sub-gradient magnitudes for densification, without explicitly computing gradient consistency:

GiAbsGS=1K∑k=1K∥∣g~i(tk)∥∣2.G_i^{\mathrm{AbsGS}} = \frac{1}{K}\sum_{k=1}^{K} \||\tilde{\mathbf g}_i^{(t_k)}\||_2.

For a fair comparison, all variants retain the same TAD-GS backbone, error mask, training configuration, and densification schedule. All experiments are conducted on N3DV. The results are reported below:

MethodPSNR ↑SSIM ↑Number↓
TAD-GS + Error Mask + AbsGS32.280.945706K
TAD-GS + Error Mask + GDAGS32.510.947378K
ELGD-GS (Ours)32.840.949347K

These results demonstrate that our consistency estimation strategy is more effective for dynamic scene reconstruction, achieving better reconstruction quality with fewer Gaussians than both AbsGS-style and GDAGS-style density control.

AuthorsOpenReview ↗

Response to Reviewer nQm2 (3/3)

Q5 & Weakness3: Temporal Interpretation of the Consistency Statistic.

We clarify that our consistency statistic measures within-observation pixel-wise gradient consistency, rather than directional consistency across timestamps. Its temporal aspect comes from aggregating per-observation consistency statistics over multiple timestamps. This follows the standard 4DGS densification formulation (Eq. 4), which computes the L2 norm of positional gradients at each observation before temporal accumulation:

Gi=1K∑k=1K∥∇μi2DLk∥2.G_i=\frac{1}{K}\sum_{k=1}^{K} \left\|\nabla_{\boldsymbol{\mu}_{i}^{2D}}\mathcal{L}_k\right\|_2.

Thus, gradients from different timestamps are not directly summed and cannot cancel each other. Our formulation follows the same principle by evaluating pixel-wise gradient consistency independently at each observation before temporal aggregation. Directly accumulating signed gradients across timestamps would introduce cross-temporal directional relationships Ccross=A′/BC_{\mathrm{cross}}= A'/B, where

A′=∥∣∑k=1Kgi(tk)∥∣2+ϵA' = \||\sum_{k=1}^{K}\mathbf{g}_i^{(t_k)}\||_2+\epsilon,

B=∑k=1K∥∣g~i(tk)∥∣2+ϵB = \sum_{k=1}^{K}\||\tilde{\mathbf{g}}_i^{(t_k)}\||_2+\epsilon.

This can cause gradient cancellation when gradients from different timestamps point in opposite directions, even if their within-observation pixel-wise consistency is high. Consequently, cross-temporal directional differences may be incorrectly interpreted as pixel-wise gradient conflicts, leading to less reliable densification decisions. To verify this distinction, we conduct an additional ablation on N3DV:

Consistency ComputationPSNR↑SSIM↑Number↓
ELGD-GS + CcrossC_{\mathrm{cross}}32.380.946376K
ELGD-GS32.840.949347K

Q6: Computational Overhead and Peak Memory Analysis.

We thank the reviewer for this valuable suggestion. We profile the per-iteration overhead of the three additional components: Error Map Computation, Quantile Computation, and Absolute Gradient Accumulation. The results are reported below:

ComponentTime (ms/iteration)↓
Error Map Computation1.279
Quantile Computation0.395
Absolute Gradient Accumulation3.496
Total Additional Overhead5.173

Furthermore, to isolate the computational and memory overhead from the effect of Gaussian count, we compare ELGD-GS with a density-matched TAD-GS baseline containing the same number of Gaussians (347K):

MethodGaussian Count ↓Iteration Time (ms)↓Peak GPU Memory (GB)↓
TAD-GS (Matched Density)347K97.13.84
ELGD-GS (Ours)347K102.34.23

These results quantify the additional runtime and memory costs of ELGD-GS under matched primitive counts.

Q7: Baseline Reproducibility and Numerical Consistency.

We thank the reviewer for raising this important question. The results of TAD-GS were obtained by re-running its official implementation rather than directly copying the numbers from the original paper. This is because the original paper does not report certain metrics, such as SSIM and the number of Gaussian primitives, which are necessary for a comprehensive comparison. We therefore re-evaluated TAD-GS under our experimental settings to ensure a more complete and fair comparison.

Regarding the slight discrepancies between the main table and the supplementary material, we found that some per-scene PSNR values were rounded to one decimal place before being averaged, resulting in numerical deviations of less than 0.05 dB. To address this issue, we recalculated the averages using the per-scene results reported in the supplementary material and updated the corresponding TAD-GS results in the main table. Specifically, the TAD-GS PSNR on N3DV has been corrected from 32.42 dB to 32.44 dB, while that on InterDigital has been corrected from 34.14 dB to 34.18 dB. We have also carefully checked and revised the related numerical descriptions in the main text to ensure consistency across the paper.

Weakness3: Comparison with Existing Compact Dynamic Reconstruction Methods.

We additionally compare ELGD-GS with 4DGS-1K (Yuan et al.), which reduces redundant Gaussians through spatial-temporal variation-based pruning and temporal filtering. Moreover, we report the average results across all six scenes of N3DV. As shown below, ELGD-GS achieves higher PSNR and SSIM while reducing the Gaussian count by 47.9% and storage by 74.6% compared with 4DGS-1K, demonstrating its effectiveness in achieving compact representations without sacrificing reconstruction quality.

MethodPSNR ↑SSIM ↑GS Count (K) ↓Storage (MB) ↓
4DGS-1K31.880.946667418
ELGD-GS (Ours)32.840.949347106
Reviewer_hSNwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ELGD-GS improves density control for dynamic Gaussian reconstruction using error-guided localization and gradient-consistency modulation on a TAD-GS backend. A photometric and structural error mask selects regions for densification, while signed and absolute pixel-subgradient statistics regulate cloning and splitting. Compact, high-quality dynamic reconstruction is a useful objective, and the reported improvements over TAD-GS on N3DV, InterDigital and VRU Basketball are encouraging. Table 3 provides module and subcomponent ablations, and the appendix adds per-scene results and dynamic-region metrics beyond aggregate scores.

Critical strengths
  1. Clear mechanistic diagnosis.
  2. Consistency statistic is well-motivated for dynamic scenes.
Critical weaknesses
  1. The contribution relative to GDAGS is insufficiently isolated. Section 2.2 already acknowledges GDAGS (Zhou et al., 2026), and Section 4.1 adopts TAD-GS for temporal modeling. However, Table 3 does not establish whether the proposed modulation improves upon adapting GDAGS to the same setting. A controlled comparison of GDAGS with the same TAD-GS backend and error mask would help separate the contribution of the proposed dynamic mechanism from transferring and combining existing strategies (pp. 3-4, 9).
  2. Equation (10) needs a more precise temporal interpretation. As written, the signed and absolute gradient sums are converted to magnitudes at each observation t before temporal accumulation. For two observations at different timestamps, internally aligned gradients can point in opposite directions across the observations while the statistic still equals one. It therefore appears to measure within-observation consistency aggregated over time, rather than relative directions across timestamps. Please clarify this interpretation and its connection to the claimed dynamic mechanism (pp. 5-6, Eqs. 9-10).
  3. The rendering-speed advantage needs attribution and a common measurement protocol. On N3DV, Ex4DGS uses 268K Gaussians at 91 FPS, whereas ELGD-GS uses 347K at 168 FPS (p. 7, Table 1). This is not necessarily inconsistent: active Gaussians per frame, projected coverage, temporal computation, and renderer implementation can all affect speed. Although Section 5.1 specifies half-resolution training and evaluation, please confirm whether all FPS results share that resolution, hardware, and timing boundary, and explain the main sources of the speed advantage.
  4. Storage savings need separate attribution to representation and density. Appendix A uses a 3D hash table for some Gaussian parameters on N3DV and InterDigital. Please clarify whether the baselines use the same representation and separate the storage reduction due to encoding from that due to fewer Gaussians. A comparison with the same density-control configuration and hash encoding enabled versus disabled could help quantify the encoding contribution, counting all parameters required for rendering. A matched-representation comparison or an equivalent breakdown would also address this concern (pp. 6-7, 13).
Questions for authors
  1. How does the consistency statistic interact with the error mask?
  2. What is the mechanism-level difference from GDAGS?
  3. Is the error mask a region selector or a per-Gaussian gate?
  4. Does densification timing interact with the consistency statistic?
  5. What happens under gradient collision without error guidance?
4 comments
AuthorsOpenReview ↗

Response to Reviewer hSNw (1/3)

We sincerely thank the reviewer for the insightful and constructive comments and we provide detailed responses to address these concerns point by point below, with additional experiments and clarifications.

Q1: Consistency Statistic Interact with the Error Mask. In our work, the consistency statistic cic_{i} is a measure of per-Gaussian pixel-wise gradient consistency to guide cloning and splitting, while the error mask ωi(t)\omega_{i}^{(t)} is a densification gate determining which Gaussians are allowed to participate in densification. These two variables alternate with each other in each training iteration. On the one hand, we construct error mask based on the current reconstruction errors to identify 4D Gaussians requiring densification. Consistency statistic is then used to guide the densification process, helping model to suppress unreliable densification and promote more effective Gaussian growth. This improves the representation of high-error regions and progressively reduces their reconstruction errors. On the other hand, based on the updated reconstruction, error mask is subsequently recomputed, driving densification process to focus on the remaining high-error regions. Through this alternating process in training iterations, the error mask and consistency-guided densification continuously drive each other, enabling more effective density allocation and progressively improved reconstruction quality.

Q2 & Weakness 1: Mechanism-level Difference from GDAGS. ELGD-GS and GDAGS both exploit pixel-wise gradient direction information for density control. However, there are three main differences at the mechanism level between them:

(1) The difference in the consistency computation scheme. GDAGS is originally designed for 3DGS static scene reconstruction and computes gradient coherence by aggregating pixel-wise sub-gradients across different views. This scheme of GDAGS to extent into 4DGS would compute gradient consistency at each training iteration kk, based on timestamp tkt_{k} sampled in that iteration. The corresponding consistency score is computed by averaging these per-observation scores: Cavg=1K∑k=1K∥∣gi(tk)∥∣2+ϵ∥∣g~i(tk)∥∣2+ϵC_{\mathrm{avg}} = \frac{1}{K}\sum_{k=1}^{K} \frac{\||\mathbf{g}_i^{(t_k)}\||_2+\epsilon}{\||\tilde{\mathbf{g}}_i^{(t_k)}\||_2+\epsilon}.

In contrast, ELGD-GS separately accumulates the L2 norms of the signed and absolute sub-gradient sums across sampled observations and computes their ratio only after temporal accumulation: Cours=A/BC_{\mathrm{ours}}= A/B, where

A=∑k=1K∥∣gi(tk)∥∣2+ϵA=\sum_{k=1}^{K}\||\mathbf{g}_i^{(t_k)}\||_2+\epsilon,

B=∑k=1K∥∣g~i(tk)∥∣2+ϵB=\sum_{k=1}^{K}\||\tilde{\mathbf{g}}_i^{(t_k)}\||_2+\epsilon.

This aggregation strategy allows observations with larger gradient magnitudes to contribute more significantly to the final consistency estimation, rather than assigning equal importance to the consistency scores of individual observations. We also provide an ablation study on N3DV with by CavgC_{\mathrm{avg}} and CoursC_{\mathrm{ours}}, as shown in following table. These results demonstrate that directly adopting the GDAGS-style consistency averaging strategy is less effective than our temporal aggregation strategy for dynamic scene reconstruction.

Consistency ComputationPSNR↑SSIM↑Number↓
ELGD-GS + CavgC_{\mathrm{avg}}32.510.947378K
ELGD-GS + CoursC_{\mathrm{ours}}32.840.949347K

(2) The difference in density modulation strategy. ELGD-GS introduces a more targeted density modulation strategy. Unlike GDAGS, which employs a single nonlinear weighting function and its inverse to regulate cloning and splitting, ELGD-GS uses two independently controlled mechanisms: high-consistency suppression for cloning and low-consistency promotion for splitting. These mechanisms are governed by separate thresholds, enabling more targeted control over Gaussian growth.

(3) The difference in Error-guided observation selection. ELGD-GS incorporates error-guided observation selection into consistency computation. The error mask acts as an observation-dependent gate on pixel-wise sub-gradients, allowing only gradients associated with under-reconstructed image regions to contribute to the temporal consistency statistics. This enables the density control mechanism to focus on regions where additional Gaussian representation is most needed.

In summary, our ELGD-GS differs clearly from GDAGS in the gradient consistency computation scheme, density modulation strategy, and error-guided observation selection. These differences demonstrate the uniqueness and novelty of our ELGD-GS.

AuthorsOpenReview ↗

Response to Reviewer hSNw (2/3)

Q3: Error Mask is a Per-Gaussian Gate. Error mask ultimately serves as a per-Gaussian gate, while patch-level filtering is introduced to suppress noisy pixel-level errors. For each observation timestamp, each 4D Gaussian is deformed into its 3D state and projected onto the corresponding 2D view. Its mask ωi\omega_i is determined by whether its projected location falls within the selected high-error patches, controlling the retention of associated pixel-wise sub-gradients for densification-statistic accumulation. Thus, 2D error regions serve as intermediate cues, while the final mask operates at the per-Gaussian level.

Q4: Densification Timing and Consistency Statistic. In our implementation, densification is performed every 200 training iterations, and the consistency statistics are accumulated interval between consecutive densification operations. Specifically, consistency statistic is accumulated over training iterations and used to compute the consistency scores when densification is performed. After each densification operation, the accumulated gradient and consistency statistics are reset for the next cycle. We will clarify this implementation detail in the revised manuscript.

Q5: Under Gradient Collision without Error Guidance. Without error guidance, the consistency statistic is computed using gradient contributions from all observed image regions, including those that have already been well reconstructed. When gradient collision occurs, gradients from different regions may interfere with one another, potentially degrading the consistency estimation and leading to less reliable densification decisions. In particular, gradients from well-reconstructed regions may introduce irrelevant contributions that interfere with density control in under-reconstructed regions.
We further compare gradient consistency with and without error guidance under identical configurations. As shown below, removing error guidance (GDonly) decreases PSNR by 0.11 dB while increasing the Gaussian count by 150K (28.7%) and storage by 62 MB, demonstrating that error guidance enables more targeted density control and better reconstruction quality.

MethodPSNR↑SSIM↑LPIPS↓Number↓
GD only32.730.9480.089487K
ELGD-GS (Ours)32.840.9490.088347K

Weakness 2: More Precise Temporal Interpretation of Equation (10).

We thank the reviewer for pointing out this ambiguity. Our consistency statistic measures pixel-wise gradient direction consistency within individual observations, rather than gradient consistency across different timestamps.

We have revised the manuscript and resubmitted the updated PDF. Specifically, we have clarified the computation and interpretation of Eq. (10), with corresponding revisions to the Abstract, Introduction, and other relevant sections. Some examples are provided below:

  • Description of Eq. (10): "Each observation is computed by L2 norms before temporal accumulation, following the procedure in Eq.4, so that this statistic quantifies pixel-wise gradient consistency within each observation.} A larger cic_i indicates higher gradient direction consistency among pixel-wise sub-gradients within individual observations,"

  • Abstract: “Second, we propose a gradient consistency-aware density control strategy that measures gradient direction consistency {over temporal observations} to regulate Gaussian growth.”

Importantly, standard 4DGS densification (Eq. 4), i.e., Gi=1K∑k=1K∥∣∇μi2DLk∥∣2G_i=\frac{1}{K}\sum_{k=1}^{K} \||\nabla_{\boldsymbol{\mu}_{i}^{2D}}\mathcal{L}_k\||_2, already computes the L2 norm of the positional gradient at each observation before temporal accumulation. Thus, gradients from different timestamps are not directly summed, and their directions cannot cancel each other. Following this principle, our method focuses on within-observation pixel-wise consistency.

Directly accumulating signed gradients across timestamps before computing their norms would conflate within-observation pixel-wise consistency with cross-timestamp directional relationships. We formulate this alternative as: Ccross=A′/BC_{\mathrm{cross}}= A'/B, where

A′=∥∣∑k=1Kgi(tk)∥∣2+ϵA'=\||\sum_{k=1}^{K}\mathbf{g}_i^{(t_k)}\||_2+\epsilon,

B=∑k=1K∥∣g~i(tk)∥∣2+ϵB=\sum_{k=1}^{K}\||\tilde{\mathbf{g}}_i^{(t_k)}\||_2+\epsilon.

This may mask severe pixel-wise gradient conflicts: two observations with substantial internal conflicts could still produce similarly directed aggregated gradients, yielding an artificially high consistency score and preventing necessary splitting. To verify this distinction, we conduct an additional N3DV ablation comparing cross-timestamp gradient accumulation with our formulation under identical settings:

Consistency ComputationPSNR↑SSIM↑Number↓
ELGD-GS + CcrossC_{\mathrm{cross}}32.380.946376K
ELGD-GS + CoursC_{\mathrm{ours}}32.840.949347K
AuthorsOpenReview ↗

Response to Reviewer hSNw (3/3)

Weakness 3: The Rendering-Speed Advantage needs Attribution and a Common Measurement Protocol.

We ensure a fair comparison by conducting all FPS measurements on the same NVIDIA A800 GPU, using half-resolution rendering and the same rasterization backend. GPU synchronization is performed before and after each frame, and timing includes both dynamic attribute computation and rasterization. We further report the average number of active Gaussians per frame. Despite having more Gaussians overall, ELGD-GS activates fewer Gaussians per frame than Ex4DGS, reducing the rendering workload. Additionally, differences in their deformation mechanisms may lead to different dynamic attribute computation costs, further contributing to the FPS gap.

MethodTotal Gaussians ↓Avg. Active Gaussians / Frame ↓FPS ↑
Ex4DGS268K256K91
ELGD-GS (Ours)347K228K168

Weakness 4: Storage Savings need Separate Attribution to Representation and Density.

We thank the reviewer for raising this concern. We clarify that TAD-GS does not use 3D hash-grid encoding in its original implementation, and its storage consists entirely of explicit Gaussian parameters. For N3DV and InterDigital, our method additionally employs a 3D hash grid to replace a subset of static Gaussian attributes. Accordingly, the total storage of ELGD-GS includes the remaining Gaussian parameters, hash-grid features, and decoder MLP parameters.

Compared with TAD-GS, ELGD-GS modifies the densification strategy and additionally introduces 3D hash-grid encoding on N3DV and InterDigital. To ensure a matched representation, we conduct two sets of experiments on both datasets: one without hash-grid encoding and the other with hash-grid encoding enabled for both methods. In this way, the representation is kept identical between the two methods in each comparison, such that the only difference affecting storage is the number of Gaussian primitives. For the hash-grid comparisons, the grid resolution, feature dimensions, and decoder architecture are kept identical across both methods.

N3DVHash GridGaussian Count ↓Storage ↓
TAD-GS✗712K255 MB
ELGD-GS✗347K124 MB
TAD-GS✓712K167 MB
ELGD-GS (Ours)✓347K106 MB
InterDigitalHash GridGaussian Count ↓Storage ↓
TAD-GS✗1604K575 MB
ELGD-GS✗756K271 MB
TAD-GS✓1604K317 MB
ELGD-GS✓756K175 MB

Taking TAD-GS without hash-grid encoding as the baseline, density control reduces storage by 51.4% (255→124 MB) on N3DV and 52.9% (575→271 MB) on InterDigital. Hash-grid encoding further reduces storage by 14.5% (124→106 MB) and 35.4% (271→175 MB), respectively. Under the same hash-grid representation applied to both methods, ELGD-GS still reduces storage by 36.5% on N3DV and 44.8% on InterDigital compared with TAD-GS. The hash-grid configuration, including grid resolution, feature dimensions, and decoder architecture, is kept identical across both methods to ensure a fair comparison.

To further clarify the contribution of hash-grid encoding, we provide a detailed storage breakdown on N3DV and InterDigital, separating the Gaussian attributes replaced by the hash grid, the remaining per-Gaussian attributes, and the additional hash-grid and decoder parameters:

N3DV:

Storage ComponentWithout Hash GridWith Hash Grid
Hash-grid-replaced Gaussian Attributes66.2 MB0 MB
Remaining Per-Gaussian Attributes57.8 MB57.8 MB
Hash Grid Parameters0 MB48.1 MB
Decoder MLP Parameters0 MB0.09 MB
Total Storage124 MB106 MB

InterDigital:

Storage ComponentWithout Hash GridWith Hash Grid
Hash-grid-replaced Gaussian Attributes144.2 MB0 MB
Remaining Per-Gaussian Attributes126.8 MB126.8 MB
Hash Grid Parameters0 MB48.1 MB
Decoder MLP Parameters0 MB0.09 MB
Total Storage271 MB175 MB

These results clearly separate the storage savings from Gaussian density reduction and hash-grid encoding, showing that both contribute to the overall compactness of ELGD-GS.

AuthorsOpenReview ↗

Response to Reviewer hSNw

Dear Reviewer,

We hope this message finds you well.

As the discussion period is nearing its end with less than two days remaining, we would like to ensure that we have addressed all of your concerns satisfactorily. If you find our responses satisfactory, we would greatly appreciate it if you could consider raising your score in recognition of our efforts. If there are any additional points or feedback, please let us know. Your insights are invaluable to us, and we are eager to address any remaining issues to further improve our work.

Thank you for your time and effort in reviewing our paper.

Sincerely,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Authors' Responses

Dear Area Chairs,

We sincerely thank area chairs and reviewers for their valuable time and thoughtful comments. We would like to provide a brief summary of our contributions and the additional work conducted during the rebuttal period.

Contributions

ELGD-GS introduces an Error-guided Localization and Gradient Consistency-aware Density Control framework for 4D Gaussian Splatting. By reconstructing errors to localize under-reconstructed regions and aggregating gradient consistency over temporal observations to guide Gaussian densification, it achieves high-quality dynamic scene reconstruction using fewer Gaussian primitives.

Discussion Status

We sincerely thank all reviewers for their valuable feedback. Reviewer rMHy confirmed that all concerns had been addressed. Reviewer nQm2 raised follow-up questions, which we addressed, but has not responded further. Reviewer hSNw has not yet responded to our rebuttal.

Main Responses and Additional Experiments

[For Reviewer hSNw]

  • Error Mask and Consistency Interaction (Q1, Q3–Q5): We clarified the per-Gaussian gating mechanism, patch-level filtering, consistency-guided cloning and splitting, and the 200-iteration accumulation/reset schedule. We also explained the gradient interference caused by removing error guidance.
  • Differences from GDAGS (Q2, Weakness 1): We clarified the differences in consistency computation, density modulation, and observation selection, with additional controlled comparisons (32.84 vs. 32.51 dB PSNR).
  • Temporal Interpretation of Eq. (10) (Weakness 2): We clarified the within-observation interpretation of gradient consistency and added cross-temporal aggregation ablations (32.84 vs. 32.38 dB PSNR).
  • Rendering Efficiency (Weakness 3): We standardized FPS evaluation using the same NVIDIA A800 GPU, resolution, rasterization backend, and synchronized timing protocol.
  • Storage Attribution (Weakness 4): We added matched-representation comparisons with and without hash-grid encoding on N3DV and InterDigital, together with detailed storage breakdowns to distinguish the contributions of Gaussian density reduction and encoding.

[For Reviewer nQm2]

  • Hash-Grid Encoding and Rendering Efficiency (Q1, Weakness 1): We evaluated ELGD-GS without hash-grid encoding and provided a detailed storage breakdown to quantify the storage costs of Gaussian attributes, hash-grid parameters, and decoder parameters.
  • Error Mask and Consistency Statistics (Q2, Weakness 2): We reported error-gated ratios across six N3DV scenes, consistency-score distributions, and cloning/splitting changes across static and dynamic regions.
  • Density-Matched Comparison and Hyperparameter Ablations (Q3, Weakness 2): We added a density-matched TAD-GS baseline (347K Gaussians), demonstrating a 0.83 dB PSNR improvement. We also evaluated different consistency-control activation iterations and modulation strengths.
  • Alternative Density-Control Strategies (Q4, Weakness 3): We implemented AbsGS-style and GDAGS-style density control within TAD-GS under identical configurations. ELGD-GS achieves the highest PSNR with the fewest Gaussians.
  • Temporal Interpretation of the Consistency Statistic (Q5, Weakness 3; Follow-up Q1): We clarified the interpretation and computation of Eq. (10), added a cross-temporal gradient accumulation ablation, and revised the Abstract, Introduction, and Section 3 accordingly.
  • Computational Overhead and Peak Memory (Q6): We quantified the additional per-iteration overhead (5.173 ms) and compared iteration time and peak GPU memory under matched Gaussian counts.
  • Baseline Reproducibility (Q7): We clarified that TAD-GS results were reproduced using its official implementation and corrected numerical inconsistencies between the main paper and supplementary material.
  • Expanded Compactness Comparisons (Weakness 3; Follow-up Q2): We added comparisons with 4DGS-1K and MoDec-GS across N3DV, InterDigital, and VRU Basketball, incorporating the results into the main experimental tables and expanding the Related Work section.

[For Reviewer rMHy]

  • Baseline Reproducibility (Q1): We clarified baseline sources and evaluation protocols.
  • Structural Error Robustness (Q2): We replaced the Laplacian operator with Sobel and evaluated its impact on reconstruction quality.
  • Cross-Framework Generalization (Q3): We integrated ELGD-GS densification into Swift4D, improving reconstruction quality with fewer Gaussians.
  • Densification Analysis (Q4): We expanded per-scene Gaussian count, storage, and cloning/splitting analyses.

We believe these additional experiments and revisions substantially strengthen the technical contributions and experimental evidence supporting ELGD-GS. We respectfully hope that the AC will consider the reviewers' positive feedback and our revisions in the final assessment.

Thank you very much for your time and support.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.