FlexWorld: Progressively Expanding 3D Scenes for Flexible-View Exploration

We sincerely thank you for the thorough, insightful, and highly constructive review. Your detailed feedback has been incredibly helpful in strengthening our manuscript. We have carefully considered all your comments and questions, and we provide our point-by-point responses below.

Weakness 1: Writing Clarity

We sincerely thank you for your valuable and detailed feedback on the manuscript's writing and structure. We agree that clarity and conciseness are paramount for conveying our work effectively. Based on your constructive suggestions, we will undertake a thorough revision of the manuscript.

Specifically, we will perform a full review to shorten overly long sentences and improve overall readability. The specific sentences you highlighted will be revised. For example, the sentence on Lines 117–119 will be revised from "Unlike most methods that generate solely from given images..., our approach integrates current scene information into newly generated content." to a more direct statement, such as: "Our approach integrates existing scene information into the generation of new content to ensure geometric consistency." Similarly, the sentence on Lines 127–129 will be revised from "While deriving 3D structures from sequential frames and maintaining consistency... presents significant challenges..., we propose an effective empirical approach." to focus on our solution, for example: "To maintain global consistency, we propose an effective empirical approach for integrating 3D structures from sequential frames."

Furthermore, as you correctly pointed out, we will explicitly reference Figure 2 at the beginning of Section 3.2 to provide readers with a clear, high-level overview of our framework from the outset. We are grateful for these actionable comments and are confident these revisions will substantially improve the clarity of our paper.

Weakness 2: Support for Flexible Camera Control

We thank you for this valid and constructive criticism. We agree that our initial manuscript claimed "flexible camera control" without providing a dedicated analysis to substantiate it. Your feedback has helped us significantly improve the validation of this key strength of our method.

To rectify this, we will add a new subsection to the paper (e.g., Section 4.5) titled "Analysis of Flexible Camera Control." In this new section, we will consolidate and reframe existing evidence from our paper to explicitly support our claim:

• Performance on Diverse and Complex Trajectories: We first argue that our model's superior performance on challenging real-world datasets like RealEstate10K and Tanks and Temples is strong evidence of its flexibility. These datasets usually feature highly varied, complex, and non-linear camera paths. Our quantitative results on these datasets, presented in Table 1, demonstrate that our model robustly handles these diverse trajectories compared to other baselines.
• Robustness to Large Camera Variation: We now explicitly draw the reader's attention to figures that specifically test the limits of camera motion. We highlight that Figure 3 and Appendix Figure 9 showcase our model's ability to maintain high quality even under large camera variations, a key aspect of flexible control.
• Qualitative Showcase of Trajectory Diversity: Finally, we point to the gallery of results in Appendix Figure 11, which provides extensive qualitative evidence of our model's performance across a wide array of different camera paths, further underscoring its versatility.

By consolidating these results into a focused analysis, we believe we have now provided clear and compelling support for our claim of flexible camera control. We thank the reviewer again for prompting this important addition to our paper.

Weakness 3: Discussion of Dataset Generation Pipeline

We thank the reviewer for this constructive suggestion. We agree that our data generation process is a crucial aspect of our work that warrants a more detailed discussion. Our data generation methodology is guided by the principle of maintaining consistency between training and inference conditions. During inference, our model is designed to handle incomplete views that are rendered from an optimized 3DGS scene with well-posed geometry. Adopting a simpler approach, such as directly generating training data from a single-shot depth estimator, would introduce significant artifacts (as shown in Figure 6 of our paper), creating a domain gap between the training data and the inference conditions. To avoid this discrepancy, our pipeline first creates a high-quality 3DGS scene from real views. From this clean 3DGS, we then extract accurate depth maps, which are back-projected to form incomplete 3D scenes. Finally, we render "incomplete" views from the scene. This process ensures that our training data closely aligns with the conditions our model will encounter during inference.

Nevertheless, we wish to emphasize the importance of the other components in our framework. As quantitatively demonstrated in our newly added quantitative ablation study (see response to reviewer v55n), our trained V2V model and our unique framework design each contribute significantly and independently to the final performance. For instance, the comparisons between configurations (1) & (4) and (2) & (4) in the table isolate and confirm the substantial impact of our V2V model and framework, respectively.

We also acknowledge the valuable suggestion regarding the experimental analysis of the dataset generation. While we are keen to conduct this analysis, the rebuttal period does not provide sufficient time to construct a new training dataset using a single-shot depth estimator like MASt3R and complete the subsequent V2V model training. Therefore, we are currently unable to include this specific experiment in the ablation study. We sincerely apologize for this omission and commit to incorporating a thorough analysis of this aspect in the final version of the manuscript.

Question 1: Camera Trajectory as Input

Thank you for your sharp observation and for pointing out this confusing statement. We sincerely apologize for the lack of clarity.

The camera trajectory c is used exclusively to render the incomplete video y (as described in Ln 164), which then serves as the input for the V2V model. The V2V model itself does not take c as a separate, direct input.

The statement on Ln 122 was poorly phrased and was the source of this misunderstanding. We intended to convey that the control over the camera path is implicit in the input video, rather than an explicit condition.

We will correct this in the revised manuscript. The sentence on Ln 122 now reads: "The V2V model enables flexible control over camera trajectories hidden within input incomplete videos to generate novel views."

Question 2: Only One Camera Trajectory?

Thank you for your question, which helps us clarify the flexibility of our method. We apologize if the wording on Line 124 caused any misunderstanding.

The trajectory mentioned in Line 124 is not the only one used across all experiments. The trajectory described on Line 124 is a specific example for generating 360-degree scenes.

Our method is designed to be general and is not constrained to any single type of camera path. As shown in other experiments, particularly in Figure 4 and Figure 5, our model successfully handles more diverse and arbitrary camera trajectories to complete generation tasks.

Question 3: Figure 4 Output and "Novel View Synthesis"

Thank you for pointing out this ambiguity regarding “novel view synthesis”. We sincerely apologize for the confusion. Figure 4 shows the direct output of our proposed V2V (video diffusion) model, not the views rendered from the optimized 3DGS scene. Our use of the term "novel view synthesis" was intended to be consistent with prior work, specifically ViewCrafter [1], where it describes the task of generating new views as a video sequence.

To resolve this ambiguity, we will revise the caption of Figure 4 to state that the results are videos generated by our video diffusion model to avoid misunderstanding.

[1] Yu W, Xing J, Yuan L, et al. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis[J]. arXiv preprint arXiv:2409.02048, 2024.

Question 4: How the Incomplete Video y is Obtained

Thank you for your question. To clarify, your understanding that y and x share the same camera trajectory is correct. However, the 3D structure used to render y is built using only the information from the first frame of the video, as stated on Ln 172. We create a sparse point cloud from this single frame, which naturally lacks information about parts of the scene outside the initial view. The "incomplete video" y is the result of rendering this sparse structure along the full camera path.

Consequently, y is guaranteed to contain significant missing regions and artifacts compared to the complete ground truth video x, making it a suitable input for our V2V task.

We sincerely appreciate your time and insightful feedback. Your comments have been invaluable in improving our manuscript. We have carefully addressed all your concerns through additional analyses and clarifications, which we believe will strengthen the paper. While some suggested experiments will require more time than is available for this rebuttal, we are confident the current modifications have substantially improved our work.

We sincerely thank you for your time and careful review of our manuscript. Your insightful feedback has been invaluable in enhancing our work. We will carefully incorporate all of your suggestions in the revised manuscript. Our point-by-point responses to your comments follow below.

Weakness 1: Novelty of the Pipeline

We thank you for your feedback. We would like to respectfully clarify how our framework's architecture represents a significant and novel departure from established approaches.

The core distinction lies in our use of a persistent 3D scene, as detailed in Sec 3.2. This allows us to leverage rich geometric and color information from the existing scene to guide subsequent expansions into a final, flexible-view 3D scene. This stateful, 3D-aware approach is a fundamental departure from frameworks like ViewCrafter, which discard intermediate geometry after each step and instead construct temporary 3D structures using only historical RGB information. This framework modification leads to superior final results, as demonstrated in Fig. 8 by the comparison between setups $c$ and (d).

We have conducted a new quantitative experiment inspired by your and reviewer v55n's suggestions to further validate the superiority of our framework design. We carefully isolated the framework's contribution by comparing two setups using the same V2V model and no image refinement: (1) a ViewCrafter-style framework and (2) our proposed framework. We generated 360° 3D scenes for 155 test cases (due to time constraints, we selected the first 5 out of 20 total categories and the first 31 out of 100 images from each, totaling 155 scenes for evaluation) under five key configurations, rendered 360° videos from these scenes, and then evaluated the videos using the official WorldScore protocol. The results below demonstrate that our framework architecture consistently outperforms the alternative across all relevant metrics.

(Note: Higher is better for all metrics. The Camera Control metric is not applicable (N/A) as our final scene camera movements are more complex than the WorldScore defaults.)

Furthermore, this persistent 3D structure uniquely endows our framework with the ability to perform scene extrapolation, as detailed in Sec. 4.4, allowing our model to seamlessly extend an existing 3D scene, a capability not present in prior works like ViewCrafter or See3D. We believe these supplementary explanations demonstrate that our framework design is, in itself, a novel and impactful contribution. We will add this quantitative study to the ablation section in our revised manuscript.

Weakness 2: Justification for Using 3DGS

Thank you for this important question. Our choice of 3D Gaussian Splatting (3DGS) over simpler point-cloud representations from methods like DUSt3R is motivated by two fundamental advantages essential to our framework: rendering quality and functional capability.

First, regarding quality, a raw point cloud is merely a geometric scaffold. Rendering it directly often produces sparse results with holes and lacks photorealism. In contrast, 3DGS is a complete scene representation where each Gaussian splat models not just geometry but also local appearance, enabling the synthesis of high-fidelity, anti-aliased novel views that capture complex view-dependent effects. Furthermore, our method operates with known camera trajectories, making direct reconstruction into a high-fidelity 3DGS representation a more natural and effective choice compared to methods using unknown camera trajectories like VGGT and DUSt3R.

Most critically, however, 3DGS is inseparable from our progressive scene expansion process. A 3DGS scene is composed of explicit, editable primitives. This property is foundational to our method, as it allows our framework to iteratively fuse new information and seamlessly expand the 3D scene. This dynamic, progressive construction is a core novelty of our work and is fundamentally different from the static, holistic point-cloud outputs generated by methods like DUSt3R.

In summary, 3DGS is not merely a post-processing choice for better visuals; it is a foundational component that enables the core functionality of our method. We are very grateful for your suggestion which prompted this discussion, and we will add this explanation to the revised manuscript.

Weakness 3: Evaluation Metrics

We sincerely thank you for this excellent and highly constructive suggestion. Adopting a 3D-centric metric is indeed crucial for a comprehensive assessment. Following your advice, we have conducted a new evaluation using the official WorldScore benchmark.

Due to the time constraint, we performed a head-to-head comparison between our method and the baseline, ViewCrafter, and evaluated a substantial subset of 300 scenes from the WorldScore static test dataset (2000 total). The subset was created by systematically selecting the first 15 scenes from each of the 20 categories to ensure diversity (including indoor/outdoor, stylized/photorealistic scenes). The evaluation followed the official WorldScore protocol, with the generation methods for both the test videos and 3D scenes remaining consistent with Sections 4.2 and 4.3 in the original paper, respectively. Due to the time limitation, we chose the first 100 scenes to do a 3D scene generation evaluation. The detailed results are presented in the table below (note: higher is better for all metrics):

Quantitative comparison of camera‐controlled video generation:

Quantitative comparison of underlying scene quality:

The results from the WorldScore evaluation show that our model achieves better overall scores, which confirms its strong performance and aligns with the findings in Tables 1 and 2 of our original manuscript. Notably, our method still shows improvement even though the camera movement in the WorldScore benchmark is relatively simple. Moreover, our approach demonstrates a more significant advantage in scenarios with large camera variations, as illustrated in Figure 3 and 9 of the manuscript.

In conclusion, this new, more comprehensive 3D-centric evaluation confirms the superiority of our method over the baseline. We will conduct these new experiments on the complete benchmark and add the results to the experimental section of our revised manuscript.

Weakness 4: Accumulated Error

Thank you for this critical question regarding error accumulation during progressive scene expansion. While it is common for accumulated errors to arise in multi-round iterative generation frameworks, our framework is specifically designed to mitigate this issue.

As mentioned in our response to Weakness 1, the core of our approach is the maintenance of a persistent 3D structure. At each expansion step, the existing 3D scene (including its geometry and color) serves as a strong, coherent anchor for the generation of the next video segment. This ensures that newly generated content is well-aligned with the established scene, fundamentally reducing the kind of accumulated error that can occur in frameworks that lack this persistent geometric memory.

By the way, we recognize that depth estimation inaccuracies may introduce accumulated errors, though improved estimators (e.g., VGGT) could reduce them. This limitation will be further discussed in the final manuscript.

Thank you again for your constructive comments. We believe the revisions and additional experiments, undertaken in response to your feedback, have thoroughly addressed your concerns and will strengthen the overall quality of our manuscript.

Thanks for your reply. I got some other questions.

1. Upon reviewing Table 1, I noticed that the results for PSNR, SSIM, and LPIPS are significantly inferior compared to those of Wonderland, which also utilized Cogvideo as their base model. What could be the reasons for this inconsistency?

2. How did you ensure scale consistency between the original pose in Re10K and the estimated point clouds? During inference, when given a single reference image, you should first build the initial point cloud and then render the video as conditioning. However, it seems that the scale of the estimated point cloud is considerably smaller than that of the ground truth pose.

3. Since worldScore do not provide ground-truth camera pose, how the camera control metric is computed?

We sincerely appreciate the time and effort you have devoted to reviewing our manuscript and providing insightful feedback. Your thoughtful comments have significantly contributed to improving the quality of our work. We have thoroughly addressed each of your suggestions and will implement the necessary revisions in the final version of the manuscript. Below, we provide detailed responses to each of your points.

Weakness 1: Limited Technical Novelty

We thank you for your feedback. While we agree that our method leverages a strong base model and an effective data construction strategy, we would like to respectfully clarify that our core technical novelty lies in the architectural design of our generation framework itself.

Our primary contribution is a novel framework that maintains and iteratively refines a persistent 3D structure. (as detailed in Lines 146-148). This allows us to leverage rich geometric and color information from the existing scene to guide subsequent expansions into a final, flexible-view 3D scene. This stateful, 3D-aware approach is a fundamental departure from frameworks like ViewCrafter, which discard intermediate geometry after each step and instead construct temporary 3D structures using only historical RGB information. This framework modification leads to superior final results, as demonstrated in Fig. 8 by the comparison between setups $c$ and (d), and is also supported by the table in the ablation study conducted at your suggestion (presented in the response to Weaknesses 2 & 3 & Question 1). Furthermore, this persistent structure enables a unique capability not present in previous methods: 3D scene extrapolation, allowing our model to seamlessly extend a scene beyond its initial boundaries (Sec. 4.4).

We believe these supplementary explanations demonstrate that our framework design is, in itself, a novel and impactful contribution. We will add this new quantitative study and its corresponding analysis to our revised manuscript.

Weaknesses 2 & 3 & Question 1: Pipeline Complexity and Lack of Quantitative Ablation

We sincerely thank the reviewer for this insightful and constructive feedback. We agree that a quantitative ablation study is crucial for clarifying the contributions of each component in our framework and for providing a clearer comparison with strong baselines (i.e., ViewCrafter).

First, for the sake of clarity, we wish to address the concern about the refinement module's impact on our main quantitative comparisons. As stated in the original manuscript (Lines 229-231), the results in Table 2 follow the experimental protocol of ViewCrafter, where metrics are calculated from 3D scenes reconstructed from videos generated without FLUX refinement. Therefore, the use of FLUX for refining the 360° flexible-view 3D scene shown in Fig. 1b does not influence those specific results.

Nevertheless, we fully agree that a quantitative ablation study is crucial for delineating the contribution of each component. Following your suggestion, we have conducted a new quantitative ablation study. Measuring the quality of 360° scenes is challenging; fortunately, a recently released WorldScore benchmark [1] (April 2025, also suggested by Reviewer GLFg) provides a way for this. For this study, we generated 360° 3D scenes for 155 test cases (due to time constraints, we selected the first 5 out of 20 total categories and the first 31 out of 100 images from each, totaling 155 scenes for evaluation) under five key configurations, rendered 360° videos from these scenes, and then evaluated the videos using the official WorldScore protocol. The five configurations are:

1. ViewCrafter's V2V Model + Our Framework (with super-resolution, without refinement)
2. Our V2V Model + ViewCrafter's Framework (with super-resolution, without refinement)
3. Our Framework (without super-resolution or refinement)
4. Our Framework (with super-resolution, but without refinement)
5. Our Full Framework (with super-resolution and refinement)

The results are presented below. (Note: Higher is better for all metrics. The Camera Control metric is not applicable (N/A) as our scenes involve complex camera paths.)

This comprehensive ablation provides several key insights that quantitatively support some of the qualitative results in Figure 8:

• Impact of our V2V Model: Comparing (1) and (4) demonstrates the superiority of our trained V2V model. It significantly boosts performance across nearly all metrics, which aligns with the qualitative improvements shown between Fig. 8a and Fig. 8d.
• Impact of our Framework: Comparing (2) and (4) highlights the effectiveness of our core framework design. When both use our V2V model, our framework substantially outperforms the ViewCrafter-style framework, confirming the observations from Fig. 8c and Fig. 8d and supporting the claim in the responses to Weakness 1.
• Impact of Super-Resolution: The difference between (3) and (4) is minimal, indicating that the Real-ESRGAN super-resolution component is an optional step for potentially enhancing visual detail and not a primary contributor to our framework.
• Impact of FLUX Refinement: Finally, comparing (4) and (5) directly addresses your question about the refinement step. As we elaborate in the Appendix (Line 492), the refinement step is used solely to enhance the visual quality of the generated scenes (which is consistent with the qualitative results in Fig. 8d and 8e), and it even leads to a slight decrease in comprehensive 3D metrics like WorldScore.

We believe this new study thoroughly clarifies the contribution of each component and quantitatively validates the ablations presented in Figure 8. We will incorporate a full quantitative ablation study and our analysis into the "Ablation study" subsection in the revised manuscript.

[1] Duan H, Yu H X, Chen S, et al. Worldscore: A unified evaluation benchmark for world generation[J]. arXiv preprint arXiv:2504.00983, 2025.

Question 2: Handling of Unobserved Regions

Thank you for this insightful question regarding our training data. In our current implementation, we do not apply any special handling (e.g., masking or inpainting) for the black, unobserved regions in the rendered training videos. This design choice was made to simplify the data framework and use a standard I2V model for our V2V task with minimal architectural changes. The model learns to interpret these unobserved areas as part of the input condition alongside the visible scene content. We will add sentences to the implementation details section of our manuscript to clarify this point.

Question 3: Dataset Filtering and Camera Trajectories

Thank you for asking for these important details.

Dataset Filtering: We performed a rigorous, multi-step filtering process on the DL3DV-10K dataset. First, we discarded scenes with low image resolutions (below 540x960) or incomplete camera annotations. Next, as a critical quality control step, we verified consistency between the provided DL3DV camera parameters and those from a COLMAP reconstruction, removing any samples with significant discrepancies. This yielded our final training set of 10,253 high-quality scenes.

Camera Trajectories: The trajectory choice was straightforward. As described in the manuscript, for each scene, we simply sample one continuous camera trajectory of a fixed length (49 frames) at random from the scene's original sequence.

We will add these explicit details to the revised manuscript. To fully support reproducibility, we also affirm our commitment to releasing the data construction code upon the paper's publication.

We are truly grateful for your time and expertise in reviewing our work. Your comments were invaluable in guiding our revisions, and we hope that the additional clarifications and experiments now fully address your concerns while improving the overall quality of the paper.

We would like to express our sincere gratitude for the valuable time you have dedicated to the peer review process and for your thorough examination of our work. Your constructive comments have been instrumental in enhancing the quality of our manuscript. We have carefully considered all suggestions and will incorporate the corresponding revisions in the final version of our manuscript. Below is our point-by-point response to your comments.

Weakness 1: Incremental Novelty

We sincerely thank you for your insightful feedback and for acknowledging the effort involved in our data preparation. We appreciate this opportunity to clarify our core novelty, which lies not only in combining existing components but more importantly, in the architectural design of our scene generation framework and the unique capabilities it enables.

First, our primary contribution is a novel framework centered around a persistent 3D structure. This allows us to leverage rich geometric and color information from the existing scene to guide subsequent expansions into a final, flexible-view 3D scene. This stateful, 3D-aware approach is a fundamental departure from frameworks like ViewCrafter, which discard intermediate geometry after each step and instead construct temporary 3D structures using only historical RGB information. Our framework leads to superior final results, as demonstrated in Fig. 8 by the comparison between setups $c$ and (d).

We have conducted a new quantitative experiment to further validate the effectiveness of our main contribution, as suggested by reviewers v55n and GLFg. We carefully isolated the framework's contribution by comparing two setups using the same V2V model and no image refinement: (1) a ViewCrafter-style framework and (2) our proposed framework. We generated 360° 3D scenes for 155 test cases (due to time constraints, we selected the first 5 out of 20 total categories and the first 31 out of 100 images from each, totaling 155 scenes for evaluation) under five key configurations, rendered 360° videos from these scenes, and then evaluated the videos using the official WorldScore protocol. The results below demonstrate that our framework architecture consistently outperforms the alternative across all relevant metrics.

(Note: Higher is better for all metrics. The Camera Control metric is not applicable (N/A) as our final scene camera movements are more complex than the WorldScore defaults.)

Furthermore, this persistent 3D structure uniquely endows our framework with the ability to perform scene extrapolation, as detailed in Sec. 4.4, allowing our model to seamlessly extend an existing 3D scene, a capability not present in prior works like ViewCrafter or See3D. We believe these supplementary explanations demonstrate that our framework design is, in itself, a novel and impactful contribution. We will add this quantitative study to the ablation section in our revised manuscript.

Weakness 2: Computational Cost

We thank the reviewer for this important point. While the 30-minute computation time per scene is a current limitation, our method's runtime for generating a 3D scene is comparable to that of ViewCrafter under the same settings.

Moreover, this issue will be addressed in future work without altering the core components of our approach. Our analysis indicates the primary bottlenecks are the video diffusion model's generation time and the iterative 3D Gaussian Splatting optimization. However, this can be substantially mitigated by leveraging rapid advancements in foundational models. For instance, recent progress in video model distillation [1] is dramatically reducing inference time, while emerging feed-forward 3DGS methods [2] can replace the costly optimization step. These technologies can be integrated into our framework in the future to make it significantly more efficient. To ensure full transparency, we will add a detailed discussion of this limitation and these potential future improvements to the "Limitations" section of our revised manuscript.

[1] Yin T, Zhang Q, Zhang R, et al. From slow bidirectional to fast causal video generators[J]. arXiv e-prints, 2024: arXiv: 2412.07772.

[2]Jiang L, Mao Y, Xu L, et al. AnySplat: Feed-forward 3D Gaussian Splatting from Unconstrained Views[J]. arXiv preprint arXiv:2505.23716, 2025.

Question 1: "3D Scene Generation" vs. "Novel View Synthesis"

We thank the reviewer for highlighting this lack of clarity and apologize for the confusion. We will explicitly define these terms in our final manuscript. In our work, we use them to distinguish two sequential stages:

1. Novel View Synthesis (NVS): This refers specifically to the output of the video diffusion model, which generates a video clip simulating camera movement through a scene. It is fundamentally a 2D video generation task, creating a sequence of new views.
2. 3D Scene Generation: This refers to the subsequent process where we use the generated video to construct an explicit and renderable 3D scene representation (i.e., 3D Gaussian Splatting). This final 3D asset can be rendered from any arbitrary viewpoint.

This distinction is adopted from the influential ViewCrafter paper [3], where our "Novel View Synthesis" is analogous to their video generation step (Fig. 3), and our "3D Scene Generation" corresponds to their "scene reconstruction" stage (Fig. 4). We will add these clarifications to the introduction of our revised manuscript.

[3] Yu W, Xing J, Yuan L, et al. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis[J]. arXiv preprint arXiv:2409.02048, 2024.

Question 2: Number of Training Samples

Thank you for this question. We trained our model on 10,253 successfully reconstructed 3D scenes from the DL3DV dataset, a widely-used collection of 3D data. This number was reached after a rigorous filtering process where we first excluded scenes with low image resolution (< 540x960) or missing camera parameters. Subsequently, we cross-validated the camera poses between the official DL3DV annotations and a COLMAP reconstruction cache, discarding any scenes with significant discrepancies to ensure high-quality data. We will update the manuscript to include these specific details about our dataset preparation. To further enhance reproducibility, we also commit to making our data processing scripts publicly available upon publication.

Question 3: Computational Cost for Fitting Gaussians

Thank you for your question regarding the reconstruction cost. For each of the 10,253 scenes in our training set, we fit the 3D Gaussians using the official implementation, running the optimization for 7,000 steps. This process takes approximately 5 minutes per scene on a single NVIDIA A800 GPU. By parallelizing the workload on a server equipped with 8 NVIDIA A800 GPUs, we were able to complete the full dataset reconstruction in approximately 4-5 days. We will add these computational specifics to the implementation details section of our manuscript for transparency and reproducibility.

Thank you again for your valuable comments and suggestions. We believe that the clarifications and supplementary experiments, made in response to your feedback, have appropriately addressed your concerns and further strengthened the persuasiveness of our paper.