Optimized Minimal 3D Gaussian Splatting

We sincerely appreciate the reviewers’ thoughtful comments and efforts in reviewing our manuscript. In the following, we address each comment individually, providing detailed explanations and clarifications as required.

Significance of our contributions

Our method achieves a 40-50% reduction in storage compared to the current state-of-the-art, LocoGS (to put this into perspective, for a large-scale and ultra-high resolution 3D digital twin of a city, this would be from 1 TB to 500 GB). To further contextualize this gain: in the field of video coding, each major standard over the past two decades, such as H.264 (2003), H.265 (2013), and H.266 (2020), has achieved approximately a 30–50% reduction per generation, typically at the cost of significantly increased computational complexity. Put differently, our method achieves a decade’s worth of progress without any sacrifices in other important factors, including rendering quality, rendering speed, and training time. We would also like to highlight our training efficiency (LocoGS: 1 hour vs. ours: 20 mins). Even when compared to MiniSplatting, a method not optimized for compression, we introduce only a 1-minute training time overhead. While our method has only been tested on standard benchmarks (Mip-NeRF360, Tanks and Temples, and the Deep Blending dataset), we believe the improvements demonstrated on these datasets remain highly relevant and reflect the current interests of the research community. Last but not least, as we already provided the source code in the supplementary materials, the codes and the model weights will be publicly available.

[Q4] Generalization ability

We have applied OMG to 3DGS-MCMC [1], a method well-known for effective densification. The following result highlights the broader applicability of OMG.

Furthermore, OMG leverages per-Gaussian features to represent the irregularity of Gaussians, making it inherently scalable and well-suited also for complex environments. To demonstrate this, we tuned the original 3DGS-MCMC for a large-scale scene, the NYC scene from the Zip-NeRF [2] dataset, and applied OMG. Our method achieves superior performance while drastically reducing storage requirements. We believe large-scale scenes presents a compelling challenge to other compression approaches that rely on compressibility and locality, and OMG’s scalability opens new possibilities for efficient modeling of extremely large scenes.

[Q3] Clarification of L. 232-233

Lines 232-233 do not suggest that "the existence of multiple Gaussians with near-identical contributions is inherently redundant." Rather, the intended point is that redundant Gaussians may exist, but cannot be effectively pruned if their contribution-based importance scores all exceed the pruning threshold. This issue often arises because Gaussians in close proximity tend to exhibit similar blending contributions. To address this, we introduce LD scoring, which differentiates their importance by assessing the distinctiveness of their appearance—increasing the importance of Gaussians with distinct appearance and decreasing it for those that are less distinctive. This enables the pruning of less distinctive Gaussians, whose contributions can be compensated by the spatial extension of nearby Gaussians during training. The effectiveness of this approach is demonstrated in Table 4 and Figure 5.

[Q2] SVQ configuration

OMG requires four types of vectors: a 3-dimensional scale vector, a 4-dimensional rotation vector, and two 3-dimensional appearance vectors. As demonstrated in Table 5, although these vectors are already short, applying naive vector quantization still demands significant computation due to the need for large codebooks. Through empirical analysis, we found that using 2-dimensional sub-vectors offers a favorable trade-off, avoiding excessive computational overhead while achieving high performance. Accordingly, we apply 2-dimensional S-VQ to the rotation vector and the unified appearance feature, while the 3-dimensional scale vector, due to its odd dimensionality, is quantized using scalar quantization.

[Q1] Existing methods leveraging neural fields

Although we have introduced existing methods that leverage neural fields in the related work section (Lines 139–146), we will provide a more detailed technical discussion in Section 3.0.

[1] Kheradmand, S., et al. 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems, 2024.

[2] Barron, J. T., et al. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023.

Dear Reviewer MiKC,

Once more, we genuinely express our gratitude for your insightful feedback on our manuscript. We gently remind you that the discussion phase will conclude in a couple of days. We believe we have effectively addressed your inquiries, concerns, and recommendations through the outcomes of our supplementary experiments. We are pleased to update our manuscript accordingly.

Should you have any additional concerns, queries, or suggestions, please feel free to reach out to us.

Best regards,

The Authors

Thank you for your response and for raising your evaluation toward acceptance. We are pleased that we were able to address your concerns and will ensure that the final version is updated accordingly.

Once again, we sincerely appreciate the time and effort you dedicated to reviewing our work.

Best regards,

The authors

We sincerely appreciate the reviewers’ thoughtful comments and efforts in reviewing our manuscript. In the following, we address each comment individually, providing detailed explanations and clarifications as required.

Significance of our contributions

Our method achieves a 40-50% reduction in storage compared to the current state-of-the-art, LocoGS (to put this into perspective, for a large-scale and ultra-high resolution 3D digital twin of a city, this would be from 1 TB to 500 GB). To further contextualize this gain: in the field of video coding, each major standard over the past two decades, such as H.264 (2003), H.265 (2013), and H.266 (2020), has achieved approximately a 30–50% reduction per generation, typically at the cost of significantly increased computational complexity. Put differently, our method achieves a decade’s worth of progress without any sacrifices in other important factors, including rendering quality, rendering speed, and training time. We would also like to highlight our training efficiency (LocoGS: 1 hour vs. ours: 20 mins). Even when compared to MiniSplatting, a method not optimized for compression, we introduce only a 1-minute training time overhead. While our method has only been tested on standard benchmarks (Mip-NeRF360, Tanks and Temples, and the Deep Blending dataset), we believe the improvements demonstrated on these datasets remain highly relevant and reflect the current interests of the research community. Last but not least, as we already provided the source code in the supplementary materials, the codes and the model weights will be publicly available.

[Q1] Sensitivity of λ and K

We have tested sensitivity of hyperparameters for LD scoring. To ensure a fair comparison in this experiment, we strictly matched the number of Gaussians by applying Top-K sampling to the importance scores, aligning it exactly with our model. Since the effect of LD scoring is more pronounced in smaller models, we adopt OMG-XS as our baseline for evaluation.

The following results show similar performance across different values of λ and K, with our chosen configuration (λ, K = 2) yielding the best results. Although these values were determined empirically, we demonstrate that this setting consistently shows strong performance and generalizes well across a variety of scenes.

[Q5] FPS on a weaker device

We have measured the FPS of our method on a low-end GPU, NVIDIA GTX 1080Ti, compared to LocoGS (with COLMAP initialization for a fair comparison). Thanks to the minimal number of Gaussians, OMG-XS achieves over 100 FPS even on this low-end hardware.

[Q4] Effectiveness of neural field

The neural field to represent the space feature is constructed by positional encoding and a tiny MLP with 6K weight parameters, The quantitative impact of space feature represented by the neural field is reported in Table 4.

[Q2] Training strategy of SVQ

As referenced in L.197, SVQ is motivated by PQ, sharing the core principle of dividing the original vector into multiple subvectors and applying vector quantization independently to each. However, we introduce SVQ into the 3DGS literature with a novel design that offers scalability across multiple attributes and an efficient training strategy. Specifically, SVQ is applied once, followed by short 1K fine-tuning steps while keeping the quantization indices fixed, which significantly simplifies training. In contrast, prior works using vector quantization, such as CompGS [1] and Compact-3DGS [2], suffer from the high complexity of learning codebooks with VQ iterations, leading to significantly increased training times compared to the original 3DGS, even though they utilize only about one-third of the Gaussians. Our SVQ approach avoids such overhead and enables efficient training, as demonstrated in Table 3 of the main paper.

[Q3] OMG for dynamic scenes

The OMG attribute representation is applicable to all 3DGS-based methods, as it flexibly represents attributes with spatial locality (via space features and SVQ) as well as those without (using SVQ alone). This generalizability is supported by the following results, where OMG is applied to another baseline, 3DGS-MCMC [3]. We believe the same principle can extend to dynamic scenes; however, the process of obtaining a minimal set of Gaussians for such scenes remains an open question and we left this for future work.

[W2] Architectural novelty

As noted by Reviewer WdNJ, the hybrid architecture for attribute representation is a clever design that elegantly balances the need for per-primitive specificity with the efficiency of continuous representation. This design makes OMG inherently scalable and well-suited not only for sparse Gaussians but also for complex environments. To validate this, we tuned the original 3DGS-MCMC [3] for a large-scale scene, the NYC scene from the Zip-NeRF [4] dataset, and applied OMG. Our method achieves superior performance while drastically reducing storage requirements. We believe large-scale scenes presents a compelling challenge to other compression approaches that rely on compressibility and locality, and OMG’s scalability opens new possibilities for efficient modeling of extremely large scenes. In this regard, we argue that the OMG architecture represents a distinct and novel contribution compared to existing approaches.

[W3] Generalization test

Although our experiments focus on the rendering task, the significance of our contributions extends beyond it. When 3DGS models are used as inputs for downstream tasks, the efficiency of the representation, particularly in achieving high fidelity with a small number of Gaussians and attribute parameters, is directly tied to training and inference complexity. And this issue is exacerated in large-scale scenes, while we prove OMG's generalization ability to large-scale scenes above. We believe that the effective minimal-set Gaussian representation proposed in this paper can benefit a wide range of downstream applications, representing a promising direction for future research.

[1] Navaneet, K. L., et al. Compgs: Smaller and faster gaussian splatting with vector quantization. In European Conference on Computer Vision, 2024.

[2] Lee, J. C., et al. Compact 3d gaussian representation for radiance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.

[3] Kheradmand, S. et al. 3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems, 2024.

[4] Barron, J. T., et al. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023.

We sincerely appreciate the reviewers’ thoughtful comments and efforts in reviewing our manuscript. In the following, we address each comment individually, providing detailed explanations and clarifications as required.

Significance of our contributions

Our method achieves a 40-50% reduction in storage compared to the current state-of-the-art, LocoGS (to put this into perspective, for a large-scale and ultra-high resolution 3D digital twin of a city, this would be from 1 TB to 500 GB). To further contextualize this gain: in the field of video coding, each major standard over the past two decades, such as H.264 (2003), H.265 (2013), and H.266 (2020), has achieved approximately a 30–50% reduction per generation, typically at the cost of significantly increased computational complexity. Put differently, our method achieves a decade’s worth of progress without any sacrifices in other important factors, including rendering quality, rendering speed, and training time. We would also like to highlight our training efficiency (LocoGS: 1 hour vs. ours: 20 mins). Even when compared to MiniSplatting, a method not optimized for compression, we introduce only a 1-minute training time overhead. While our method has only been tested on standard benchmarks (Mip-NeRF360, Tanks and Temples, and the Deep Blending dataset), we believe the improvements demonstrated on these datasets remain highly relevant and reflect the current interests of the research community. Last but not least, as we already provided the source code in the supplementary materials, the codes and the model weights will be publicly available.

[W1-5/2-2/3-c] Already Provided Results

We respectfully note that the following key points have already been addressed in the main manuscript or the appendix:

• [W1-5] Importance of the space feature: The space feature F, obtained via positional encoding (PE) and a tiny MLP, is one of the core components of OMG. It is designed to capture local continuity among sparse Gaussians, as elaborated in Lines 169–175 of the main paper. Since this feature encodes the basic structural information of Gaussians, the per-Gaussian features can concentrate more on representing fine details. Importantly, we do not claim that the space feature should be implemented using PE+MLP; rather, we demonstrate that even an extremely small neural field (with 6K weight parameters) is sufficiently effective. The quantitative impact of this space feature is reported in Table 4.

• [W2-2] OMG-XS performance on the Mip-NeRF 360 dataset: The performance of OMG-XS is reported in Table 1, the main table in our manuscript. OMG-XS achieves comparable rendering quality to the previous state-of-the-art method, LocoGS, while requiring nearly half the storage.

• [W3-c] Effectiveness of post-processing: We have analyzed the impact of post-processing techniques in Appendix 4, with corresponding quantitative results shown in Table 3 of the appendix. The results show that applying post-processing achieves an overall storage reduction of approximately 30–32%.

[W1-1] Other methods' FPS with RTX 4090

We have measured the FPS of LocoGS (with COLMAP initialization for a fair comparison) on an NVIDIA RTX 4090 and a lower-end GPU, NVIDIA GTX 1080 Ti. As shown in the results, OMG achieves significantly faster rendering speeds.

[W1-2/1-3/1-4] Clarification of LD scoring

To ensure a fair comparison in LD scoring experiments, we strictly matched the number of Gaussians by applying Top-K sampling to the importance scores, aligning it exactly with our model.

• [W1-2] Usage of Morton order approximation instead of KNN The table below presents a comparison between Morton-order approximation and KNN-based scoring on 9 scenes from the Mip-NeRF 360 dataset. The results are nearly identical, demonstrating that Morton order effectively approximates KNN while offering significantly faster runtime.

  Mip-NeRF 360	PSNR	SSIM	LPIPS	Time (ms)
  OMG-XS	27.06	0.807	0.243	4100
  Morton -> KNN	27.07	0.808	0.241	28

  While we agree with the reviewer that using KNN is reasonable in our current setup (LD scoring once at 20K with fewer than 1M Gaussians), our decision to adopt Morton order is motivated by the need for scalability and generalizability. In more complex or larger-scale scenes, the number of Gaussians can increase significantly during earlier training stages. In such cases, KNN becomes impractical due to its high memory usage and computational complexity of $O(N^2)$. Morton order, by contrast, provides a well-approximated yet efficient alternative that scales better with the number of primitives.

• [W1-3] LD scoring and SVQ LD scoring is not applied after the primitives have been quantized by SVQ. Importance-based pruning with LD scoring is performed at the 20K iteration (as noted in Line 25 of the appendix), whereas SVQ is applied from 29K iteration (during the final 1K iterations). Therefore, SVQ introduces no bias into the pruning process.

• [W1-4] LD scoring with RGB The static appearance feature T is used to represent both RGB values (the DC component of spherical harmonics) and opacity. Since T shares similar characteristics with RGB, LD scoring based on RGB also yields comparable but slightly less performance.

  Mip-NeRF 360	PSNR	SSIM	LPIPS
  OMG-XS	27.06	0.807	0.243
  T → RGB	26.99	0.807	0.243
  OMG-M	27.21	0.814	0.229
  T → RGB	27.19	0.814	0.229

[W1-3-ii] Including the effect of SVQ in Table 4

S-VQ achieves substantial storage reduction while preserving the original performance. We appreciate the reviewer’s suggestion and will include this result in Table 4 in the final version of the paper.

[W2-1-i] Comparison with Taming-3DGS

Taming-3DGS is primarily designed to accelerate training and achieves faster training times compared to our method. However, it does not incorporate any attribute compression, and yet it shows inferior rendering quality compared to OMG. In contrast, OMG achieves substantially better rendering quality while also requiring significantly less storage.

[W2-1-ii] Sensitivity to initialization

Since OMG is built upon Mini-Splatting, it suffers from performance degradation without initialization. However, OMG can also be applied to other baseline methods such as 3DGS-MCMC, which exhibit stable performance regardless of COLMAP initialization. The following result highlights the broader applicability of OMG.

[W2-3/W3-a] Clarification of M and L

As noted by Reviewer WdNJ, SVQ is proposed to strike a better balance among computational cost, storage efficiency, and representational precision for Gaussian attributes. OMG requires four types of vectors: a 3-dimensional scale vector, a 4-dimensional rotation vector, and two 3-dimensional appearance vectors. As demonstrated in Table 5, although these vectors are already short, applying naive vector quantization still demands significant computation due to the need for large codebooks. Through empirical analysis, we found that using 2-dimensional sub-vectors offers a favorable trade-off, avoiding excessive computational overhead while achieving high performance. The scale vectors, having an odd dimension (3), are instead quantized using scalar quantization.

Furthermore, by dividing the vector into sub-vectors, the value distribution within each sub-vector becomes less complex. This simplifies the statistical characteristics of the data, allowing Huffman encoding to exploit redundancy more effectively. As a result, the final storage size remains stable across varying bit allocations, making the method less sensitive to quantization bit-width choices and more robust to distributional variations.

[W3-b] Quantization and rendering speed

Quantization does not affect rendering speed, as rendering in this field is commonly performed on GPUs using floating-point operations. Therefore, the improved speed of OMG results from the reduced number of primitives rather than from quantization.

Dear Reviewer wotF,

Once more, we genuinely express our gratitude for your insightful feedback on our manuscript. We gently remind you that the discussion phase will conclude in a couple of days. We believe we have effectively addressed your inquiries, concerns, and recommendations through the outcomes of our supplementary experiments. We are pleased to update our manuscript accordingly.

Should you have any additional concerns, queries, or suggestions, please feel free to reach out to us.

Best regards,

The Authors

Thank you for your response, and we appreciate you increasing the rating. We are glad that our clarification sufficiently addressed the reviewer’s concerns. We would like to appropriately include the provided materials into the final version.

Once again, thank you for your valuable review.

Best regards,

The authors

it was it, sorry. You answered my questions with enough clarity and I have read the answers to other authors. I will raise my score to BA.

We sincerely appreciate the reviewers’ thoughtful comments and efforts in reviewing our manuscript. In the following, we address each comment individually, providing detailed explanations and clarifications as required.

Significance of our contributions

Our method achieves a 40-50% reduction in storage compared to the current state-of-the-art, LocoGS (to put this into perspective, for a large-scale and ultra-high resolution 3D digital twin of a city, this would be from 1 TB to 500 GB). To further contextualize this gain: in the field of video coding, each major standard over the past two decades, such as H.264 (2003), H.265 (2013), and H.266 (2020), has achieved approximately a 30–50% reduction per generation, typically at the cost of significantly increased computational complexity. Put differently, our method achieves a decade’s worth of progress without any sacrifices in other important factors, including rendering quality, rendering speed, and training time. We would also like to highlight our training efficiency (LocoGS: 1 hour vs. ours: 20 mins). Even when compared to MiniSplatting, a method not optimized for compression, we introduce only a 1-minute training time overhead. While our method has only been tested on standard benchmarks (Mip-NeRF360, Tanks and Temples, and the Deep Blending dataset), we believe the improvements demonstrated on these datasets remain highly relevant and reflect the current interests of the research community. Last but not least, as we already provided the source code in the supplementary materials, the codes and the model weights will be publicly available.

[Q1/2] Performance improvement over LocoGS

• [Q1] OMG-XS is 48.6% smaller than LocoGS-S: We believe this gain is not trivial. For instance, in the field of video coding, each generation has achieved approximately a 2× improvement in compression ratio: H.265 (2013) roughly doubles the efficiency of H.264 (2003), and H.266 (2020) further doubles that of H.265, while imposing significanly increased computation. While 3DGS compression continues to be highly optimized, we would like to argue that the additional compression gains achieved by our method are substantial, especially when considering the accompanying improvements in rendering speed.

  Mip-NeRF 360	PSNR	Size (MB)	FPS (3090)
  LocoGS-S	27.04	7.90	310
  OMG-XS	27.06	4.06 (-48.6%)	350

  Furthermore, while a 4 MB reduction in absolute terms may appear trivial, it becomes highly significant in large-scale scenes. OMG leverages per-Gaussian features to represent the irregularity of Gaussians, making it inherently scalable and well-suited for complex environments. To demonstrate this, we tuned the original 3DGS-MCMC  for a large-scale scene, the NYC scene from the Zip-NeRF  dataset, and applied OMG. Our method achieves superior performance while drastically reducing storage requirements. We believe large-scale scenes presents a compelling challenge to other compression approaches that rely on compressibility and locality, and OMG’s scalability opens new possibilities for efficient modeling of extremely large scenes.

  NYC	#G	PSNR	SSIM	LPIPS	Size
  Zip-NeRF	-	28.42	0.850	0.281	607 MB
  3DGS-MCMC	3 M	27.77	0.857	0.295	675 MB
  MCMC+OMG	3 M	27.82	0.853	0.291	29 MB

• [Q2] MB per Gaussian is not the key factor: To reduce the number of Gaussians in LocoGS, we incorporate Mini-Splatting’s densification strategy while retaining LocoGS’s attribute representation. Although this approach successfully reduces storage by decreasing the number of Gaussians, LocoGS fails to maintain high-quality rendering. This limitation arises because a single neural field struggles to represent multiple attributes of sparse Gaussians effectively. A similar challenge is observed in Compact-3DGS as well, as demonstrated in Table 1 of the appendix.

  Mip-NeRF	#G	PSNR	SSIM	LPIPS	Size
  MS+Loco-S	946 K	27.04	0.804	0.225	9.1 MB
  MS+Loco-S	537 K	26.52	0.789	0.257	6.2 MB
  MS+Loco-L	923 K	27.35	0.815	0.214	14.2 MB
  MS+Loco-L	519 K	26.76	0.799	0.247	11.3 MB
  OMG-XS	427 K	27.06	0.807	0.243	4.1 MB
  OMG-XL	727 K	27.34	0.819	0.218	6.8 MB

[Q3] Clarification of L. 261-262

Lines 261–262 refer to the results on the Mip-NeRF 360 dataset. For a more comprehensive evaluation, we assessed overall performance on Mip-NeRF 360 (9 scenes), Tank & Temples (2 scenes), and Deep Blending (2 scenes). OMG-M achieves higher PSNR and SSIM than LocoGS-S, while reducing storage by 37%. Furthermore, OMG-XL achieves superior SSIM compared to LocoGS-L, with a smaller storage footprint than LocoGS-S. We will revise the final version accordingly to clarify this point.

[W4/5/6] Clarification of LD scoring

To ensure a fair comparison in LD scoring experiments, we strictly matched the number of Gaussians by applying Top-K sampling to the importance scores, aligning it exactly with our model. Since the effect of LD scoring is more pronounced in smaller models, we adopt OMG-XS as our baseline for evaluation.

• [W4] Usage of Morton order approximation instead of KNN The table below presents a comparison between Morton-order approximation and KNN-based scoring on 9 scenes from the Mip-NeRF 360 dataset. The results are nearly identical, demonstrating that Morton order effectively approximates KNN while offering significantly faster runtime.

  Mip-NeRF 360	PSNR	SSIM	LPIPS	Time (ms)
  OMG-XS	27.06	0.807	0.243	4100
  Morton -> KNN	27.07	0.808	0.241	28

  As our current setup applies LD scoring only once at the 20K iteration with fewer than 1 million Gaussians, KNN is affordable and potentially more reliable in this context. However, our decision to adopt Morton order is motivated by the need for scalability and generalizability. In more complex or larger-scale scenes, the number of Gaussians can increase significantly during earlier training stages. In such cases, KNN becomes impractical due to its high memory usage and computational complexity of $O(N^2)$. Morton order, by contrast, provides a well-approximated yet efficient alternative that scales better with the number of primitives.

• [W5] Design choice for LD scoring We conducted an ablation study on the design choices for LD scoring and found that the results remain nearly unchanged. We would like to include this result in the final version of the paper.

  Mip-NeRF 360	PSNR	SSIM	LPIPS
  OMG-XS	27.06	0.807	0.243
  T -> T&V	27.06	0.807	0.243
  L2	27.05	0.806	0.244

• [W6] LD scoring and the most dominant contributor As the reviewer pointed out, LD scoring is applied to Gaussians that have been the dominant contributor for at least one ray. This design reflects our objective to construct a minimal yet effective set of Gaussians. Given the limited number of Gaussians, further pruning typically leads to performance degradation. However, LD scoring enables the identification of less distinctive Gaussians whose contributions can be compensated by learning (generally extending) nearby Gaussians over the following iterations. The effectiveness of this approach is validated in Table 4 and Figure 5.

[W7] Hybrid model for geometry

Considering that the geometry of sparse Gaussians exhibits weak spatial continuity, we chose not to fuse the space feature into the geometry representation. We agree that an ablation study on this aspect would be valuable. However, incorporating geometry predicted by an MLP is not feasible in our setting, as it cannot be properly initialized during training (at 15K iteration). Randomly initialized geometry leads to out-of-memory issues, especially due to the scale of each Gaussian. Despite considerable effort, we found it challenging to train the model under such conditions.

[W1/2/3/8] More results

We sincerely appreciate the reviewer’s suggestion and will include the following results in the final version of the paper.

• [W2/8] Qualitative results compared to other methods Due to the strict rebuttal guidelines of NeurIPS, we are unable to include qualitative results here. However, we will certainly incorporate them in the final version.

• [W3/8] LocoGS-L in Table 2 and Mini-Splatting in Table 1,2

  Tank&Temples	PSNR	SSIM	LPIPS	Size	FPS
  LocoGS-L	23.84	0.852	0.161	12.34	311
  Mini-Splatting	23.41	0.846	0.180	67.6	(1095)

  Deep Blending	PSNR	SSIM	LPIPS	Size	FPS
  LocoGS-L	30.11	0.906	0.243	13.38	297
  Mini-Splatting	30.04	0.910	0.241	124.9	(902)

  Mip-NeRF 360	PSNR	SSIM	LPIPS	Size	FPS
  Mini-Splatting	27.39	0.822	0.216	119.5	(601)

Thank you for your response and for updating your rating toward acceptance. We would be happy to include the provided quantitative results, with the additional qualitative examples, in the final version or supplementary material.

Once again, thank you for your thoughtful feedback and efforts in reviewing our manuscript.

Best regards,

The authors