Response to weakness 1:

We would like to emphasize that the main contribution of our work is not merely the use of a regularization loss to encourage the model to prune, but rather the development of a learning framework that can automatically learn a pruning ratio embedding during the training process of a 3DGS model. This is achieved through our proposed Gumbel-Sigmoid mask pruning technique (Equation 8) and our ensemble loss function (Equation 10), all with only one round of training. We kindly request that the reviewer refer to our summary of technical contributions in the Common Question.

For the setting of $\lambda_{m}$, we used a value of 5e-4 in all the reported experimental results. We have further conducted an ablation study on this hyperparameter based on the Room scene, as detailed below.

$\lambda_{m}$ have little impact on the final rendering quality.

Response to weakness 2:

Figure 2 is based on the vanilla 3DGS, with our proposed pruning method highlighted within the green box. This box indicates that a mask is learned for each Gaussian to determine whether it should be pruned during the training process. The detailed technique described mathematically relates to Equations 7, 8, and 10.

For Figure 3, since the pruning process involves a binary decision (i.e., whether to prune or not prune), it is preferable for the learned mask distribution to naturally reflect a binary distribution—where values close to 0 indicate the Gaussian should be pruned, and values close to 1 indicate it should not be pruned.

Response to weakness 3:

Since we need to report the training time for an apples-to-apples comparison, it is essential to use the same GPU hardware setup. We have open-sourced all our code and experimental results on GitHub for peer researchers to verify and benchmark.

For 3DGS and Mini-Splatting, we reported results using their open-source codes on our GPU hardware. Additionally, because RadSplat is not open-sourced, we re-implemented and benchmarked their method ourselves. It is important to note that all our self-implemented results are consistent with the rendering quality reported in the original papers. To further clarify, we have provided a comparison of the original paper results for 3DGS and our self-implemented results as follows:

Response to weakness 4:

Please refer to the Common Question about the term "optimal" pruning for additional details.

Dear reviewer, We hope our rebuttal can address your comments and concern. As the deadline for the author-reviewer discussion is approaching, would you please share your thoughts on our rebuttal and reconsider the review score? thanks a lot

Dear authors,

I do agree that "our learned pruning ratio is always 'better' than random chosen pruning ratio, since there is no 'better' pruning ratio in model pruning problem". However, in my previous comment, I only expect that your method is better in over 80% scenes. I do not require an absolute better or optimal. At least a number larger than 50% can prove that the proposed method is effective. I do not think a successful rate of 80% is a high standard for a publishable paper. Learnable and automatic method is not preferred if just a random manual number setting can solve the problem comparably. This is the most important point in my opinion. If the proposed method can not demonstrate enough effectiveness, I can only regard it as an engineering trick in GS reconstruction.

As for the comment "such 0.5 to 0.8 pruning ratio interval "found in this work" is based on the vanilla and default 3DGS training setup, ... Different 3DGS training hyperparameter settings...". I want to clarify that: If you want to prove your method is effective on other variants of GS, please just show evidence. If you want to prove that finding empirical prune ratio of other GS is more difficult compared with adopting your method, please show evidence.

Finally, as for your comment "We kindly recommend the reviewer to check our discussion with the reviewers o7SR and fvLt above. Both agreed the community will benefit from our proposed automatic pruning", I do have seen all the discussions above and I will keep my independent thinking. I've spent lots of time on this paper and I am not an outsider of GS. I kindly recommend the authors repect every review equally no matter they are positive or negative.

Response to weakness 1:

Please refer to the Common Question about the term "contribution" for more details.

Response to weakness 2:

Please refer to the Common Question about the term "optimal" pruning for more details.

Response to weakness 3:

We have comprehensively compared our method with two of the most recent and state-of-the-art (SoTA) 3DGS pruning works, namely RadSplat and Mini-Splatting. The experimental results are reported in Figure 4, Tables 1 and 2, and additional results in the appendix. We appreciate the reviewer’s comment and will include comparisons with the suggested works.

EAGLES primarily modifies the hyperparameters of the densification stage of 3DGS to control the number of Gaussians, which differs from score-based pruning methods such as RadSplat, Mini-Splatting, and LightGaussian. Our method uses a learnable mask on the scores of Gaussian points, making it incompatible with EAGLES.

Scaffold-GS and HAC focus on encoding the attributes of 3DGS to compress the model, which is different from score-based pruning methods. These works use encoding methods rather than pruning, making them incompatible with our approach.

We will cite these works as references and discuss them in the manuscript.

Dear reviewer, We hope our rebuttal can address your comments and concern. As the deadline for the author-reviewer discussion is approaching, would you please share your thoughts on our rebuttal and reconsider the review score? thanks a lot

Response to weakness 1&3:

Please refer to the Common Question about the clarification and summary of our technical contributions.

We would like to highlight that the main contribution of our work is not merely adopting the Gumbel-Sigmoid to prune 3DGS models but rather developing a learning framework that can automatically learn a pruning ratio embedding during the training process of a 3DGS model. This is achieved through our proposed Gumbel-Sigmoid mask pruning technique (Equation 8) and ensemble loss function (Equation 10), with only one round of training.

Regarding the analysis shown in Figure 4, RadSplat/Mini-Splatting calculates their defined importance score for each Gaussian. If the target is to prune 60% of the entire model, the Gaussians are ranked based on this predefined importance score to achieve the 60% pruning. The actual pruning threshold of the importance score value can then be calculated.

To clarify, Figures 1, 4, and 6 display the pruning ratio of the entire model versus the rendering quality, rather than the actual threshold of the importance score. This is why the x-axis ranges from 0 to 1, where each point corresponds to the rendering quality with a pruning ratio ranging from 0.1 to 0.9 in our experiments. Each point represents the result of one round of training with a preset pruning ratio, sweeping from 10% to 90%.

For different pruning methods, such as RadSplat and Mini-Splatting, the results indicate that there always exists a balanced pruning ratio region that achieves a large pruning ratio with almost no or negligible rendering quality drop. In contrast, our method can learn such a pruning ratio with only one round of training.

As observed by the reviewer, our method can automatically learn balanced pruning ratios with respect to different scenes. In other words, our proposed method does not require a pre-defined pruning ratio or threshold like prior pruning works, but instead learns one. We believe this is the greatest contribution of our work, and our method will benefit the community.

Response to weakness 2:

We kindly disagree with the reviewer and believe that we have already provided a comprehensive comparison with two of the most recent and best-performing state-of-the-art (SoTA) works, namely RadSplat and Mini-Splatting. The experimental results are reported in Figure 4, Tables 1 and 2, and additional results in the appendix.

In the paper "RadSplat: Radiance Field-Informed Gaussian Splatting for Robust Real-Time Rendering with 900+ FPS", arXiv:2403.13806, the authors compared their method with LightGaussian, demonstrating better rendering image quality and a smaller model size. Consequently, we adopted RadSplat as our baseline.

We appreciate the reviewer’s comments and, as suggested, we have also included a comparison with LightGaussian as follows:

Regarding the comparison with Mini-Splatting, since our method is a learn-to-prune approach, it exhibits slight variations across different scenes. The average performance is reported in Table 1. It appears that the method performs better with the importance score defined in RadSplat compared to the one defined in Mini-Splatting, given our hyperparameter settings. We will investigate this further in our future work.

Response to weakness 4:

We kindly disagree with the reviewer on this comment. Please refer to our response to the Common Question about technical contributions for a detailed explanation. Our primary goal is not to simply learn a smaller model with improved performance as in previous works, but to develop a learning-to-prune framework that automatically determines a balanced pruning ratio. This approach achieves an excellent compression ratio without compromising rendering quality, addressing a challenge that has not been previously tackled in the 3DGS community.

The overall training cost for our learning-to-prune methodology is significantly lower compared to prior manual pruning works, such as RadSplat, Mini-Splatting, and LightGaussian. Even when compared to Compact3D, our method, which uses a differentiable mask rather than a straight-through estimator (STE) binary mask, demonstrates better performance. Tables 4 and 5 present the ablation study that highlights this improvement.

Response to weakness 5:

We appreciate the comment. We will correct the image.

Dear reviewer, We hope our rebuttal can address your comments and concern. As the deadline for the author-reviewer discussion is approaching, would you please share your thoughts on our rebuttal and reconsider the review score? thanks a lot

Response to W1:

Thank you for the comment. Please refer to the Common Question about the term ``optimal" pruning.

Response to W2:

For Radsplat/Mini-Splatting, they calculate their defined importance score for each Gaussian. If the target is to prune 60% of the entire model, the Gaussians are ranked based on this predefined importance score to prune 60% off. Correspondingly, the actual pruning threshold of the importance score value can be calculated.

To clarify, Figures 1, 4, and 6 show the pruning ratio of the entire model versus the rendering quality, rather than the actual threshold of the importance score. These figures already report what the reviewer asked for. Each point in these figures represents the result of one round of training with a preset pruning ratio, sweeping from 10% to 90%. The results indicate there always exists a balanced pruning ratio region for both Radsplat/Mini-Splatting that can achieve a large pruning ratio with almost no or negligible rendering quality drop. We have never denied this. However, our method can learn such a pruning ratio with only one round of training.

Response to W3:

Following the same training time metric used in most prior 3DGS model pruning works, such as RadSplat, Mini-Splatting, and EAGLES, we also use training time as an important metric for a direct comparison. All experiments we reported are conducted with the same GPU hardware setup, ensuring consistency. Our method learns a smaller 3DGS model, where the saved MACs and FLOPS depend on the learned pruning ratio.

The training time we report is for only one round of training. This is to show that our proposed method incurs almost no training time overhead or even results in shorter training time due to the smaller pruned model in the later stages of the training pipeline. It is important to note that prior manual pruning works require multiple rounds of training to find the balanced pruning ratio setting, making it challenging to simply report the MACs and FLOPS.

Response to MC1:

Thank you for the comments. We conducted further ablation studies on those hyperparameters.

We discovered that the parameter $\tau$ has a very minor effect on the learned pruning ratio.

We also conducted an ablation study on the parameter $\lambda_{m}$ using the Room scene. The experimental results are as follows:

Ablation Study on Parameter $\lambda_{m}$ (Room Scene)

These results indicate that varying $\lambda_{m}$ has a negligible impact on the PSNR, SSIM, and LPIPS metrics in the Room scene.

Regarding the timing of pruning the model, we found that the best time to perform pruning is at the 20,000th epoch. This timing allows the model extra training epochs to fine-tune based on the pruned model. Pruning the model too early is not advisable, as the 3DGS densification stage stops at the 15,000th epoch, and the model requires additional time to learn the scene. We have already reported the complete set of hyperparameter settings and open-sourced our code on GitHub for researchers to reproduce our results. We will include this ablation study in the final manuscript if space permits.

Response to MC2:

Given the rapid development of the 3DGS community, new pruning methods are emerging quickly. We hope our method is compatible with most pruning methods, especially those defining new importance scores for pruning (please refer to the Common Question about Contribution above).

Cosmetic issues:

We appreciate your comments and will address all cosmetic issues in the updated manuscript accordingly.

W1

We appreciate your comment. we will change to learned effective pruning ratio.

W2

We thank you for your further clarification. Please see our detailed response below:

"It’s possible that for one scene a score may be at the 40% pruning ratio whereas that same score may be at the 60% pruning ratio in a different scene."

We totally agree with the reviewer. To justify such, we conducted extra experiments to get the pruning ratio while fixing the pruning threshold of RadSplat defined importance score. The results are listed in below table. It can be clearly seen that the same threshold score for different scenes lead to different pruning ratios for different scenes. This actually further justifies the importance of our work, as it implies that different scenes indeed need the "learned effective pruning ratio'' to optimize tradeoff between compression ratio and rendering quality.
Due to the fact that it is a most common setting in prior works to measure the rendering image quality with target pruning ratios, for fair comparison, we design our framework to learn an effective pruning ratio for each scene, without multiple rounds of manual tuning of pruning ratios as used in prior works

"It’s entirely possible that a specific score may provide a better threshold across scenes because it likely will not corespondent to a specific pruning ratio, but will be a measure of the confidence in a particular Gaussian."

This is a very interesting point. We conducted such experiments on different scenes using different threshold values, based on RadSplat pruning importance score and the results are shown in figure below. It can be seen that, in the kitchen scene, the PSNR starts to drop significantly when the threshold is larger than 0.2. However, in the room scene, such cutoff point is around 0.15. It shows different scenes have different sensitivity against pruning threshold.

Kitchen Room

More interestingly, since our proposed method learns the effective pruning ratio for different scenes during 3DGS model construction, we discovered that it naturally captures such unique cutoff threshold score for different scenes.

MC2

We appreciate you for sharing this valuable insight. In equation 8, if we only use $m_i$, then we could regard opacity ($o$) as the importance score. It is also true that the opacity itself could be one type of importance score. Our objective is to propose a method that is general, learnable, and orthogonal to other prior or potentially future works exploring different definitions of effective 3DGS pruning importance scores.