4DGCPro: Efficient Hierarchical 4D Gaussian Compression for Progressive Volumetric Video Streaming

Thank you for your positive feedback on our paper and for highlighting its strengths, including "the first hierarchical 4D Gaussian compression approach for progressive volumetric video streaming", "a well-designed architecture that adapts to the heterogeneous capabilities of client devices", "compact and consistent temporal representation" and "efficiency and scalability". We will now address each of the weaknesses and questions you have raised one by one, and we commit to incorporating the following revisions into the final version of the paper.

Unseen Regions (W1)

Although we have not designed a process for adding Gaussians, our motion-aware adaptive grouping strategy enables the modeling of emerging content. When new regions appear in the scene, the motion of Gaussians becomes intense, triggering the switch to a new keyframe as the subsequent reference. While this strategy does not directly add new Gaussians, it models emerging regions by reperforming keyframe modeling when new objects appear for accurate representation. As evidenced by the results on the N3DV dataset, our method still maintains competitive performance, achieving better reconstruction quality than 3DGStream under storage conditions of less than 1/12.

QP setting (W2)

We have supplemented ablation experiments with different QP settings on the 4DGCPro dataset, and the results are shown in the table below. Combined with the experimental phenomena, it can be seen that different QP settings have a significant impact on the performance of 4DGCPro: Low QP (0, 10) preserves fine Gaussian parameters, with reconstruction quality close to lossless but large data volume (+1.05MB for QP=0, +0.61MB for QP=10), which is suitable for scenarios requiring high image quality; Medium QP (20) achieves a balance between quality and compression ratio with less information loss, making it the default choice; High QP (30) leads to excessive information loss during compression due to excessive quantization intensity, resulting in a significant decline in quality (-0.97dB). By adjusting the QP value, 4DGCPro can provide multiple discrete bitrates, adapting to streaming scenarios with fluctuating bandwidth.

Coding Methods (Q1)

Through experiments, we found that the encoding and decoding time consumption of H.265 is greater than that of H.264: encoding takes approximately 3.1 times longer, and decoding takes about 1.7 times longer. More importantly, H.265 causes more information loss (-3.4 dB) during the compression of Gaussian features. In contrast, deep learning-based encoding methods significantly increase decoding latency, making real-time applications impossible, while also being difficult to integrate into existing pipelines. We will incorporate the above analysis into the final paper.

Rotation Entropy (Q2)

We have conducted additional ablation experiments to illustrate the rationality of the training scheme selection. The reconstruction of keyframes does not rely on motion information from other frames. Performing entropy estimation for rotation features in keyframes can effectively reduce the size of Gaussians (-0.03MB) with almost no impact on reconstruction quality. However, introducing this in motion modeling will affect the accuracy of rigid transformation estimation, leading to a decrease in reconstruction quality (-0.2dB).

Data Process (Q3)

Following the setting in HiFi4G, we first performed 2x downsampling for subsequent reconstruction.

Thank you for acknowledging our work as "dividing entire frames into optimal sub-groups" and "SOTA 4D Gaussian representation". Going forward, we will respond to each of the weaknesses and questions you have raised one by one, and we commit to incorporating the following revisions into the final version of the paper.

Technical contributions (W1&Q4)

While our framework shares the high-level goal of compressing dynamic 3D Gaussian representations with V³, our method introduces several substantial technical innovations that differentiate it from V³ in terms of motion modeling, grouping, bitrate scalability, and entropy coding:

(1) Motion-aware Adaptive Gaussian Grouping. Unlike V³ which adopts a fixed group size, we propose a motion-aware adaptive grouping strategy that dynamically adjusts group boundaries based on scene motion magnitude. This allows better temporal consistency and compression efficiency across varying motion intensities. As shown in the middle part of Table 4, our method significantly improves RD performance compared to fixed-size grouping, achieving a minimum BD-PSNR gain of +0.25dB.

(2) Motion Decomposition. Unlike V³, which estimates motion solely via position offset and subsequently fine-tunes other attributes, our method explicitly decomposes motion into rigid transformations (position + rotation) and residual deformations (scale, opacity, color, etc.). This dual-level modeling allows better handling of complex dynamics and improves reconstruction fidelity. To verify this, we replaced our motion strategy with that of V³ in our pipeline, which led to a PSNR drop from 31.64dB to 31.50dB on the N3DV dataset as shown in the below table, confirming the effectiveness of our fine-grained motion decomposition.

(3) Beyond V³: Single-Model Bitrate Flexibility. Our framework introduces a Perceptually-Weighted Hierarchical Gaussian Representation with layer-wise RD supervision and is the first to support variable bitrates within a single unified model for 4D Gaussian compression. This progressive structure enables smooth adaptation to bandwidth variations and device constraints during volumetric video streaming. In contrast, V³ lacks this flexibility, as it requires training and storing separate models for each target bitrate, leading to higher storage overhead and limited adaptability in practical streaming scenarios.

(4) Attribute-Specific Entropy Modeling. Unlike V³, which does not incorporate entropy supervision for keyframes during training, our method adopts an RD-aware, attribute-specific entropy coding strategy. Specifically, we employ KDE-based modeling for keyframes to better capture their complex and diverse attribute distributions, and Gaussian-distribution modeling for inter-frame residual attributes. This tailored approach enables more accurate entropy estimation, thereby improving overall RD performance. To assess its effectiveness, we conducted an ablation study by removing entropy supervision for keyframes, i.e., adopting V$^3$'s setting, and observed a clear performance drop: the BD-PSNR on the HiFi4G dataset decreased 0.20dB compared to our full method. This result highlights the critical role of entropy supervision in our design.

Clarification and Analysis of Hierarchical Representation (W2&Q1)

Thank you for your insightful comment. We provide the requested clarifications below, focusing on our initialization strategy, layer-wise supervision, and empirical analysis with varying numbers of hierarchy levels.

(1) Initialization Strategy. We adopt a principled approach to initialize the hierarchical structure. Specifically, after pre-training a full-resolution 3DGS model (see Lines 209–213), we sort all Gaussians in descending order using our significance metric defined in Eq. 3, which balances geometric visibility and photometric opacity. These Gaussians are then evenly divided into $L$ layers to form the initial hierarchy. To validate this design, we additionally experimented with an alternative metric that replaces the summation in Eq. 3 with multiplication. This variant significantly underweights Gaussians with low volume but high opacity, resulting in blurred reconstructions and a -1.33dB BD-PSNR drop, confirming the effectiveness of our proposed initialization scheme.

(2) Layer-wise Rate-Distortion Supervision. As described in Subsection 3.3, we design an end-to-end layer-wise RD supervision strategy that explicitly optimizes each level in both Keyframe and Inter-frame stages. For each layer, we compute the entropy of attributes and the photometric loss independently, ensuring that each layer contributes non-redundant, high-value information. As shown in the right part of Table 4, disabling this layer-wise supervision and applying global supervision to all layers will cause a 61.21% increase in BDBR and a 2.89dB drop in BD-PSNR, demonstrating that our layer-wise scheme is crucial for maintaining compression efficiency and reducing redundancy.

(3) Analysis with Varying Number of Layers. To analyze the impact of the number of layers $L$, we conducted experiments with different settings (see table below). Using $L$=6 as baseline, we found that fewer layers (e.g., $L$=4) degrade BD-PSNR by up to 0.87dB, while increasing $L$ beyond 6 yields marginal gains but increases training time significantly (e.g., +1.2 min/frame). We therefore select $L$=6 as the best trade-off between efficiency and RD performance.

Comparison with Existing Significance Metrics (Q2)

We have conducted additional experiments comparing our proposed significance metric with those adopted by recent compact 3D Gaussian approaches, including LightGaussian, EAGLES, Mini-Splatting and Taiming-3DGS. The results are summarized in the table below.

• LightGaussian computes a dynamic, view-dependent importance score considering volume, opacity, and light transmittance. It offers a +0.07 dB BD-PSNR gain, but at the cost of +1.1 min/frame pre-training time.
• EAGLES uses static rendering weights as significance. It is efficient but less adaptive to motion and occlusion, resulting in a -0.15 dB drop in BD-PSNR.
• Mini-Splatting adopts blending weights but lacks generalization to dynamic scenes (BD-PSNR: -0.13 dB, +3.6 min/frame overhead).
• Taiming-3DGS combines geometric and perceptual saliency terms. It improves quality slightly (+0.03 dB), but adds high training complexity due to multi-view evaluations.

In contrast, our proposed metric, based on a simple combination of volume and opacity, is lightweight, scene-agnostic, and efficient, achieving comparable or better RD performance without incurring extra training cost. By avoiding dependence on rendering feedback or multi-view saliency, our method is particularly suitable for scalable 4D Gaussian streaming under constrained resources.

Clarification of Spatial Volume Definition (Q3)

In the context of Gaussian representations, the "spatial volume" refers to the 3D volume that a Gaussian occupies in the 3D space, which is calculated as $\dfrac{4}{3}\pi abc$, where $a$, $b$, and $c$ correspond to the scale parameters (standard deviations) along the three principal axes of the Gaussian.

Effectiveness of Motion-Aware Adaptive Grouping (Q5)

We agree that validating the motion-aware adaptive grouping strategy is important. In fact, Table 4 of our manuscript already provides an ablation study comparing our adaptive grouping method against multiple fixed group sizes (5,10,15,20,25). Across these settings, our adaptive strategy consistently demonstrates better RD performance, indicating its effectiveness under various scene dynamics.

To further strengthen this analysis, we have conducted an additional experiment comparing our adaptive grouping to a degenerate case of GOP size = 1 (i.e., treating every frame independently with no temporal grouping). This setting performs even worse, with a BD-PSNR degradation of -0.96dB, which underscores the importance of appropriately grouping frames with coherent motion.

We also explored the sensitivity of our method to the grouping threshold hyperparameter $\tau_u$, by testing values of 0.002 and 0.003. Both resulted in noticeable RD performance drops (−0.07dB and −0.10dB BD-PSNR, respectively), suggesting that our chosen default ($\tau_u$ = 0.0025) provides a good trade-off between motion sensitivity and compression performance.

Minor Typo (Q6)

Thank you for pointing this out. We will correct the duplicated word "are" in the revised version.

Writing Quality (W3)

Thank you for your suggestion. We will improve the overall writing in the revised version. Specifically, we will enhance the clarity of technical descriptions, reorganize the methodology section for better logical flow, and polish the language throughout the paper to improve readability and presentation quality.

(1) More technical details. We will provide a more detailed introduction to progressive rendering. Additionally, we will elaborate further on the initialization strategy based on significance metric sorting, layer-wise rate-distortion supervision, and the selection of hyperparameter $L$ that balances RD performance and training efficiency.

(2) More terminology explanations and typo fixes. We will clarify the definitions of spatial volume and interp(·) in the main text. Meanwhile, we will remove the redundant "are" in Line 147.

Thank you for recognizing our work as "interesting and show promise for progressive streaming", "superiority" and "more compatible with existing hardware and supports faster decoding". We will now address each of the weaknesses and questions you have raised one by one, and we commit to incorporating the following revisions into the final version of the paper.

Clarity (W1&Q1)

Thank you for your valuable feedback. We appreciate your concerns regarding the clarity of certain parts of the paper, and we will address each of your points in detail:

(1) Spatial Volume in Equation 3. In the context of Gaussian representations, the "spatial volume" refers to the three-dimensional volume that a Gaussian occupies in the 3D space, which is calculated as $\dfrac{4}{3}\pi abc$, where $a$, $b$, and $c$ are the three dimensions of the Gaussian’s scale. This volumetric measure quantifies the spatial influence of a Gaussian, contributing to its visual significance in the scene.

(2) Why is Equation 3 Formulated as a Weighted Sum? Eq.3 is formulated as a weighted sum of spatial volume $S$ and opacity $\alpha$ because these two attributes contribute independently to the visual significance of a Gaussian. Specifically, $S$ reflects the spatial extent of a Gaussian in 3D space, while $\alpha$ quantifies its visual prominence in rendering. Both factors influence the final reconstruction quality, but in different and complementary ways.

A weighted sum is used to preserve their respective contributions while allowing flexible trade-offs. In contrast, a multiplicative formulation would overly suppress Gaussians with small volume but high opacity, leading to loss of fine details. To balance the influence of $S$ and $\alpha$, we introduce a tunable trade-off parameter $\lambda_{\Psi}$.

We further conducted ablation studies to validate this design choice. As shown in the table below, using a multiplicative form or removing $\lambda_{\Psi}$ altogether leads to significant performance degradation, with BD-PSNR drops of -1.33 dB and -2.06 dB, respectively. These results confirm the necessity of the weighted formulation and the role of $\lambda_{\Psi}$ in achieving robust performance.

(3) How Gaussians Are Partitioned into Layers. The partitioning of Gaussians into layers is based on their importance scores, computed using a significance metric that jointly considers spatial volume $S$ and opacity $\alpha$. After sorting all Gaussians in descending order by their importance scores, we uniformly divide them into $L$=6 hierarchical layers. This strategy ensures that each layer captures a progressively finer level of scene detail, facilitating scalable reconstruction.

We chose $L$=6 based on empirical validation: as shown in the table below, using fewer layers degrades RD performance (e.g., BD-PSNR drops of -0.87dB and -0.38dB for smaller $L$), while increasing $L$ beyond 6 offers marginal gains at the cost of increased training complexity (+0.6/+1.2 min per frame). This confirms that our choice strikes a favorable trade-off between rate-distortion efficiency, visual quality, and training cost.

Experimental Concerns (W2&Q2)

We thank the reviewer for their insightful comments. Below, we address each of the points raised:

(1) Ablation Study for Residual Modeling (Group=1). We acknowledge that an important ablation study without residual modeling was missing in our initial submission. To address this, we have now conducted the experiment by setting Group=1, which removes the residual modeling step. This condition reflects a scenario where each frame is modeled independently, without the temporal consistency enforced by residual deformations. The results are shown in Table below. Specifically, when comparing this configuration with our full method, we observe a 48.37% increase in BDBR and a -0.96dB drop in BD-PSNR. This demonstrates the significant redundancy in Gaussian information when residual modeling is not used, which in turn leads to a noticeable degradation in compression efficiency and visual quality. These results highlight the importance of residual modeling in preserving temporal consistency and reducing redundancy.

(2) Comparison with QUEEN. We have compared our method with QUEEN under the same experimental settings. The results are shown in the table below. As can be seen from the comparison, our method achieves better reconstruction quality while maintaining a significantly lower bitrate. Specifically, for the N3DV dataset, our method outperforms QUEEN in both PSNR and SSIM, while reducing the file size from 0.75 MB (QUEEN) to 0.73 MB. For the Immersive dataset, we observe a similar trend, where our method achieves a PSNR of 30.23dB and SSIM of 0.947, compared to QUEEN’s PSNR of 29.40dB and SSIM of 0.917, while also reducing the file size from 1.79 MB to 1.34 MB. Additionally, a key advantage of our method is its ability to render multi-quality reconstruction results from a single model. This capability enables flexible quality scaling based on network conditions and computational resources, which is particularly beneficial for adaptive streaming scenarios.

Inconsistency in claims (W3)

We apologize for any confusion caused by the repeated use of the term “large-scale.” To clarify, in Line 161, "large-scale complex motions" refers to intense, complex object motions within a scene, where our method excels. Thanks to our precise motion modeling and adaptive grouping, 4DGCPro performs particularly well with highly dynamic motions. In contrast, "underperforms in large-scale scenes" in Line 315 refers to spatially extensive scenes, such as city-scale environments, where the method currently lacks optimizations to efficiently handle the large scene size. This is a limitation we aim to address in future work.

Qualitative Result (W4)

Thank you for your valuable feedback. We would like to clarify that all images, including those for 3DGStream, are from the same timestamp (the 30th frame). The apparent differences arise from the limitations of the 3DGStream method in handling complex motions, such as large displacements or rapid movements. Specifically, 3DGStream only models rigid motion, and it struggles to capture fast or large motions, such as head rotation, sleeve movement, and umbrella spinning, which results in motion artifacts like trajectory fragmentation and residual errors from earlier frames. These issues cause visual differences, which are not due to different timestamps but are a result of the method's motion modeling constraints. We will revise the figure caption in the manuscript to include this explanation.

Typo Fixes (W5)

Thank you for pointing out these minor issues. We will correct the duplicated word “are” in Line 147 and change “a” to “an” in Line 202. Additionally, we will unify the terminology in Line 220 and Eq.9 to use PMF consistently throughout the manuscript.

Ablation Studies (Q3)

Thank you for your insightful suggestion. We have incorporated a comprehensive set of ablation studies that target the core design modules of our framework. Specifically, we evaluate (1) the necessity of temporal grouping, (2) the formulation of the significance metric, (3) the impact of the number of hierarchy levels, (4) the effectiveness of multi-layer coding, and (5) the role of attribute-specific entropy modeling. The results and analyses are summarized below:

(1) Group Size = 1 (No Temporal Grouping). We simulate a frame-by-frame modeling baseline by setting the group size to 1, which disables residual modeling. Compared to our full method, this leads to a 48.37% increase in BDBR and a -0.96 dB drop in BD-PSNR, highlighting the importance of motion grouping and temporal modeling. This result is discussed in our response to W2.

(2) Significance Metric Design (Eq. 3). To verify the necessity of our weighted sum formulation, we compared it against alternatives such as multiplicative fusion and omission of the trade-off weight λΨ. These result in BD-PSNR drops of -1.33dB and -2.06dB, respectively, demonstrating that our proposed metric more effectively captures Gaussian importance. Detailed discussion can be found in our response to W1.

(3) Number of Hierarchical Levels. We conducted additional experiments varying the number of hierarchy levels $L$. The results show that fewer layers (e.g., $L$=3) lead to performance degradation, while increasing $L$ beyond 6 provides negligible gains at the cost of higher training time. Thus, we set $L$=6 as a balanced choice. Please refer to our response to W1 for quantitative results.

(4) Single-Layer vs. Multi-Layer Coding. To verify the effectiveness of our residual-aware hierarchical design, we compare our 4DGCPro (multi-layer) against a single-layer coding baseline. Despite introducing layered transmission, our method achieves comparable quality (29.47dB at 1.31MB vs. 29.50dB at 1.27MB), confirming that our hierarchical representation avoids inter-layer redundancy. This is enabled by our residual learning and Gaussian significance-based sorting.

(5) Attribute-Specific Entropy Modeling. We also ablated our entropy coding strategy by removing KDE-based supervision from keyframes. The result is a BD-PSNR drop of 0.20dB, affirming the benefit of attribute-specific modeling for capturing distributional variations across key and inter-frames.

Thank you for acknowledging our work as "outperforms the included baselines" and for providing numerous helpful suggestions. Next, we will respond to each of the weaknesses and questions you have raised one by one, and we commit to incorporating the following revisions into the final version of the paper.

Novelty and Contributions beyond Conventional Coding Strategies (W1&Q1)

We sincerely thank the reviewer for this insightful comment. While some high-level concepts such as intra-/inter-frame coding and progressive transmission are common in video compression, we would like to clarify that our work introduces novel and tailored adaptations specifically for 4D Gaussian compression, which differ significantly from traditional pixel-domain compression. Our key innovations include:

(1) Joint Representation-Compression Framework for 4D Gaussians. Traditional video compression techniques operate in the 2D pixel domain, while our work directly compresses spatiotemporal 3D Gaussian representations. This requires rethinking how motion, hierarchy, and entropy modeling are designed. Unlike conventional pipelines, we propose a deeply coupled framework that unifies representation learning and compression-aware optimization of 4D Gaussians. This is a challenging and relatively underexplored direction, with few studies providing systematic solutions.

(2) Fine-Grained Motion Modeling at the Representation Level. Unlike traditional video codecs that model motion at the pixel or block level, our method explicitly decomposes motion on the Gaussian primitive level into rigid transformations (position and rotation) and residual deformations (scale, opacity, color, etc.). This dual-level modeling better captures complex dynamic behaviors and leads to improved reconstruction fidelity. To validate its effectiveness, we conducted an ablation study replacing our motion model with that of V³[1] within our pipeline. This substitution resulted in a PSNR drop from 31.64dB to 31.50dB on the N3DV dataset, confirming the advantage of our motion decomposition design.

(3) Motion-Aware Adaptive Grouping for Enhanced Compression Efficiency. In contrast to traditional video coding that often uses a fixed GOP size, we propose a motion-aware adaptive grouping strategy that dynamically clusters Gaussian primitives based on the estimated motion magnitude in 3D space. This adaptive grouping dynamically adjusts group boundaries according to scene motion intensity, enabling better temporal consistency and compression efficiency across frames with varying motion. As demonstrated in the middle section of Table 4, our method achieves a significant improvement in rate-distortion performance over fixed-size grouping, with a minimum BD-PSNR gain of +0.25dB.

(4) Single-Model Bitrate Scalability with Minimal Redundancy. Unlike traditional scalable video coding, which often suffers from inter-layer redundancy due to repeated information across layers, our method introduces a Perceptually-Weighted and Residual-Aware Hierarchical Gaussian Representation with layer-wise RD supervision. This design ensures that each layer contributes non-overlapping and perceptually significant details, minimizing redundancy and enhancing coding efficiency. This is the first framework that supports multiple bitrates within a single unified model for 4D Gaussian compression, eliminating the need for model retraining and excessive storage typically required in prior works. We validate our multi-layer design by comparing it with a single-layer baseline. 4DGCPro achieves comparable RD performance (29.47dB at 1.31MB vs. 29.50dB at 1.27MB), confirming that our progressive scheme retains high quality while offering scalability.

(5) Attribute-Specific RD-Aware Entropy Modeling. We further go beyond standard entropy coding by tailoring distribution-specific estimators for each Gaussian attribute. For example, we use KDE-based entropy modeling for keyframe Gaussians (with high variance) and Gaussian-based modeling for inter-frame Gaussians. This novel design leads to improved RD efficiency and is verified via ablation studies (Section 4.2), where removing this entropy design causes BD-PSNR to drop by 0.60dB.

Usage of the term "variable rate" (W2&Q2)

We understand your concern regarding the use of the term "variable-rate model." To clarify, our model allows for flexibility in bitrate selection, but rather than offering truly continuous bitrate variability, it provides a set of discrete bitrate options based on the multi-layer coding structure. Specifically, the model uses a 6-layer structure that corresponds to different quality/detail levels, and within each layer, we adjust the QP parameters (ranging from 1 to 25) during the compression phase to allow for fine-grained bitrate variation. This enables flexible bitrate choices across these fixed levels, effectively offering a 6x25 range of possible bitrates.

We agree that "variable-rate" may imply continuous bitrate variability, which is not the case here. Therefore, we will revise the terminology in the paper to "multiple bitrate" to more accurately describe the nature of the model’s bitrate flexibility.

Progressive Rendering Details (W3)

We appreciate your concern regarding the details of progressive rendering. Our process begins with the base layer ($l$=1), which contains essential scene information. As additional layers are progressively decoded, we merge them with the previously decoded layers to refine the visual quality. Specifically, each layer adds extra detail, and only a small amount of additional data (Gaussians) is needed for the next layer ($l$+1), enhancing the rendering quality.

By dynamically adjusting the level of detail based on computational and bandwidth resources, we maintain a balance between quality and efficiency. This approach enables scalable, high-fidelity visualization while ensuring real-time performance within resource constraints. We will expand on these details in Subsection 3.1 of the revised manuscript, particularly on how layers are combined and rendered.

interp(·) (W4)

In our work, interp(·) refers to the hash grid interpolation operation, which is used to map points in the hash grid to corresponding values during the encoding process. We will include this definition in the revised manuscript.

Single-layer and multi-layer coding results (W5&Q3)

We have conducted experiments comparing single-layer coding with our 4DGCPro (multi-layer coding) on our dataset. Our results show that 4DGCPro achieves comparable RD performance to single-layer coding, with only a slight difference in quality (29.47dB at 1.31MB vs. 29.50dB at 1.27MB). This is primarily due to our compact hierarchical Gaussian representation, which effectively reduces redundancy both between layers and across frames. In particular, we apply a residual learning-based representation to model the residuals between layers and across frames, ensuring that each layer captures unique, non-redundant information. This minimizes the redundant information that typically arises in multi-layer coding, thus maintaining high compression efficiency.

Additionally, we introduce layer-wise RD supervision to refine the optimization process, ensuring that each layer contributes optimally to the overall representation. As a result, despite the typical trade-offs associated with multi-layer coding, our approach significantly reduces redundancy and maintains high efficiency, achieving performance similar to single-layer encoding.

We will include these detailed comparisons and further explanations on compression efficiency in the revised manuscript to provide a clearer understanding of the trade-offs involved.

Experimental setup (W6)

Thank you for your suggestion. The details of our experimental setup, including the hardware and platform used, are provided in Line 264 and Lines 533-534. We will ensure that these details are more clearly highlighted in the revised manuscript.

H.264 configuration (W7&Q4)

We thank the reviewer for pointing this out. The H.264 encoder was configured using the x264 library with the following settings: I/P-frames only (no B-frames), 3 reference frames, color space in YUV4:4:4, and preset set to "medium." We will include these details in the main paper to ensure reproducibility.

[1] V³: Viewing Volumetric Videos on Mobiles via Streamable 2D Dynamic Gaussians, SIGGRAPH Asia 2024