MVSplat360: Feed-Forward 360 Scene Synthesis from Sparse Views

We introduce MVSplat360, a feed-forward approach for 360{\deg} novel view synthesis (NVS) of diverse real-world scenes, using only sparse observations. This setting is inherently ill-posed due to minimal overlap among input views and insufficient visual information provided, making it challenging for conventional methods to achieve high-quality results. Our MVSplat360 addresses this by effectively combining geometry-aware 3D reconstruction with temporally consistent video generation. Specifically, it refactors a feed-forward 3D Gaussian Splatting (3DGS) model to render features directly into the latent space of a pre-trained Stable Video Diffusion (SVD) model, where these features then act as pose and visual cues to guide the denoising process and produce photorealistic 3D-consistent views. Our model is end-to-end trainable and supports rendering arbitrary views with as few as 5 sparse input views. To evaluate MVSplat360's performance, we introduce a new benchmark using the challenging DL3DV-10K dataset, where MVSplat360 achieves superior visual quality compared to state-of-the-art methods on wide-sweeping or even 360{\deg} NVS tasks. Experiments on the existing benchmark RealEstate10K also confirm the effectiveness of our model. The video results are available on our project page: https://donydchen.github.io/mvsplat360.

Paper

References (77)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 3TXW5/10 · confidence 4/52024-06-13

Summary

This paper proposes a framework of generalizable 3DGS for 360 degree of NVS. The framework comprises of two main components, one is similar to MVSplat that combines multi-view context image information, the other uses a pre-trained diffusion model for post-processing, which is conditioned on the rendered features from the MVS model and outputs detailed images. They also condition the diffusion model on the CLIP features of input images to integrate high-level semantics. Experiments show improvements on both DL3DV and RealEstate10K datasets, especially in qualitative results where the model can generate plausible contents that are not observed in the context images.

Strengths

1. Originality: the paper proposes a novel setting of scene-level wide-sweeping representation through feed-forward. The setting is practical and with potential. 2. Clarity: The paper is well-written and readers can easily understand the proposed method.

Weaknesses

1. Contribution: One limitation of this paper is the technical contribution. The combination of generative model and genalizable 3DGS, and conditioning it on rendered features, are similar to latentSplat [1], and the generative process is directly taken from video diffusion models [2]. And the multi-view image feature fusion is taken from MVSplat [3]. The key contribution is somewhat ambiguous. 2. Experiments: the reported quantitative improvements over previous SOTA methods are limited on both datasets. PSNR of $17.00\sim 18.21$ and SSIM of $0.496\sim 0.553$ is far from satisfactory. The authors mentioned that the refinement module does not guantee improvements on pixel-wise metrics, but the improvements on LPIPS is also minor. 3. Efficiency: another key limitation is on the rendering efficiency. Comparing to latentSplat, integrating a heavy diffusion model for post-processing leads to significantly slower rendering speed ($>100FPS\rightarrow\sim 1.75FPS$). The major problem is that applying similar post-processing as NeRF-based works like ReconFusion [4] largely hinders the real-time rendering of 3DGS. [1] Wewer, Christopher, et al. "latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction." arXiv preprint arXiv:2403.16292 (2024). [2] Blattmann, A., Dockhorn, T., Kulal, S., Mendelevitch, D., Kilian, M., Lorenz, D., Levi, Y., English, Z., Voleti, V., Letts, A., et al.: Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127 (2023) [3] Chen, Yuedong, et al. "MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images." arXiv preprint arXiv:2403.14627 (2024).

Questions

1. To compare with latentSplat more thoroughly, is it possible to evaluate the performance using MVSplat+latentSplat Decoder? The results could better evaluate the necessity of applying a heavy diffusion model for decoding. 2. PSNR results are missing in Table 3(b), and the performance is still not satisfying when using 7 views. Is it possible to provide ablation study results with PSNR using more input views ($>7$) during inference? The results can evaluate the generalization of the model to credibly reconstruct (nearly) fully observed scenes, and also strengthens the claimed 360$^{\circ}$ NVS that can possibly be applied to large scene reconstruction. 3. In Table 2, the extrapolation results of pixelSplat and latentSplat on re10k seem to be adopted from Table 2 of latentSplat, but latentSplat reported interpolation PSNR of 24.32dB for pixeSplat (instead of 25.89dB) due to a different training strategy (supervising on both interpolation and extrapolation views). Is it possible to report latentSplat results and pixelSplat extrapolation results on re10k using the same training strategy as the other methods? The results would form more fair comparisons. 4. In Table 2, MVSplat360 improves only 0.12dB PSNR on extrapolation results over MVSplat. However, there should be major cases where the non-generative methods render nothing in the unobserved regions in extrapolation views (eg. Figure 4 pixelSplat and MVSplat results) due to lack of post-processing. Theoretically, MVSplat360 should improves more than only 0.12dB over MVSplat.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

See the weaknesses and questions above.

Authorsrebuttal2024-08-13

We thank the reviewer for the follow-up comments. We are very grateful for your recognition of our key contribution: good visual quality in an under-explored setting of 360 scene synthesis. Below, we summarise information to justify the running efficiency, in case you, other reviewers and AC want to refer to it during the internal discussion.   * **The task of feed-forward 360 scene synthesis from sparse views is very challenging.** * Sparse views (e.g., 5 views in our experiments) provide only limited coverage of a scene, resulting in many unseen or occluded regions when rendering from 360-degree viewpoints. This causes significant challenges to previous feed-forward methods like pixelSplat, MVSplat and latentSplat. Prominent failures can be observed in all these methods, as demonstrated in our supplementary video. * Another line of research, like ReconFusion and CAT3D, relies on additional per-scene optimization to achieve plausible results, which is an order of magnitude slower than our method (minutes vs. seconds (Ours)). * In this paper, we demonstrate that high-quality feed-forward 360 scene synthesis can be achieved from sparse views, without any additional per-scene optimization. * **The efficiency of video diffusion models can be potentially improved with the latest techniques since video diffusion models are actively developed and advanced.** * The main efficiency bottleneck of our method lies in the video diffusion model, which is slow since it requires the multi-step denoising process. * However, we note that speeding up video diffusion models is an active topic and advanced techniques are emerging. For example, the recent SF-V model speeds up SVD with improved *one-step* denoising. Our model can benefit from such advancements.   Overall, we provide a viable solution to the under-explored task of feed-forward 360 scene synthesis from sparse views, we believe our efficiency can be further improved with more advanced video diffusion models. Feel free to let us know if you would like us to clarify anything further. Many thanks.

Reviewer pb1U5/10 · confidence 4/52024-07-07

Summary

This paper aims to advance 360° novel view synthesis from sparse observations in wild scene scenarios. The key idea is to utilize the improved MVSplat for coarse geometry, refined by a stable video diffusion model to enhance appearance. This differs from prior work that, due to sparse viewpoint inputs, resulted in the ambiguous rendering of wide-sweeping or even 360° novel views. The proposed MVSplat360 method has been evaluated on challenging datasets, showing a marked enhancement over existing state-of-the-art techniques.

Strengths

1. Originality and Significance The concept of combining the generalizable GS with the SVD model is eminently logical; this innovative method effectively addresses the current limitations of feed-forward neural novel view synthesis techniques in 360° scene scenarios. By establishing a new benchmark for 360° scene reconstruction from sparse views, this paper proposes a novel direction for future research in the field. 2. Experimentation and Evaluation The authors constructed a new benchmark using the challenging DL3DV dataset and performed extensive comparisons on the RealEstate10K dataset. The results consistently show that MVSplat360 outperforms existing state-of-the-art methods in both qualitative and quantitative assessments. 3. Presentation The paper is overall well-written and the limitations have been sufficiently discussed in the supplementary.

Weaknesses

1. From L193-201 and the provided demo, it becomes apparent that for complex scenes with occluded and invisible parts, influenced by the generative model, MVSplat360 may struggle to achieve stable multi-view consistency solely through "Multi-view conditions." 2. The article addresses the issue of oversaturated colors through a simple post-processing method; however, the effectiveness of this solution is not validated in the experimental and ablation sections.

Questions

1. In L193-201, the "Multi-view conditions" exhibit limitations in maintaining multi-view consistency across scenes. Firstly, can CLIP embedding tokens of the original visible views guarantee a scene-level global description that guides an accurate denoising process? Additionally, for occluded and invisible parts (as depicted by the black pixels in the rendered results in Fig. 4), are the GS-rendered features filled with background values for these areas? How is multi-view consistency ensured in these regions? It is advisable for the authors to include visual results or analyses of rendered views for invisible parts from multiple perspectives of the same scene in the experiments or appendices. 2. The process of “Appearance Refinement” appears to be influenced solely by the Gaussian spherical harmonics parameters. In cases where the geometry inferred by MVS is inaccurate, what impact would this have on the denoising process of SVD? 3. The distinction between a) and b) in Fig.5 is unclear; it is advisable to revise the figure captions accordingly.

Rating

5

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

The authors have thoroughly outlined the limitations of their work as well as the potential negative societal impacts.

Reviewer RhQM5/10 · confidence 3/52024-07-13

Summary

This paper proposes MVSplat360, a generalized sparse-view novel view synthesis method. MVSplat360 utilizes the Stable Video Diffusion model to guess the novel views besides input sparse views.

Strengths

1. MVSplat360 achieves better novel view synthesis with sparse input views by introducing the stable video diffusion. 2. The paper is well-written.

Weaknesses

1. Most of the success is due to the stable video diffusion, and this paper provides mainly engineering works.

Questions

1. I would like to see what the contributions are besides introducing the diffusion model.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors have discussed the limitations.

Reviewer 4RiL5/10 · confidence 5/52024-07-26

Summary

The paper proposes MVSplat360, a method for wide-sweepign or 360-degree novel view synthesis on general scenes from sparse input views. It extends an existing state-of-the-art approach MVSplat to render 3D feature Gaussians as conditioning for a refinement network in form of a pre-trained video diffusion model, which is fine-tuned jointly. The approach is evaluated on RealEstate10k, an existing benchmark of home walkthrough videos, and on a newly proposed benchmark for 360-degree novel view synthesis, leveraging an existing dataset of diverse scenes. Quantitative and qualitative results show improvements over regression-based and generative, splatting-based baselines and plausable completions in case of incomplete observations.

Strengths

- The paper tackles a challenging and interesting task of generalizable 360-degree novel view synthesis from sparse views on diverse real-world scenes. - The proposed method combines the strengths of splatting-based generalizable 3D reconstruction and large-scale pre-trained video diffusion models as a genrative refinement network: - pixelSplat [1] and MVSplat [2] have shown strong performance in view interpolation. - latentSplat [3] has shown the advantages of using a decoder generative network. - Video diffusion models like Stable Video Diffusion [4] have been trained on massive data and learn a strong prior useful for plausible view extrapolation and scene completion. - The problem definition and main approach explained well. - The evaluation validates the effectiveness of MVSplat360 - The proposed benchmark on the DL3DV dataset [5] is challenging and supports the claims of wide-sweeping or 360-degree novel view synthesis on large real-world scenes. - The approach outperforms all baselines in almost all metrics quantitatively on DL3DV and RealEstate10k. - Qualitative results show high-quality novel views with less artifacts than baselines for both datasets with plausible completions for extrapolation on RealEstate10k. - The supplement provides more convincing qualitative results including a video. - The paper includes an ablation study validating all the proposed design choices and showing desired performance scaling w.r.t. increasing numbers of input views. 1. pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction, CVPR 2024 2. MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images, ECCV 2024 3. latentSplat: Autoencoding Variational Gaussians for Fast Generalizable 3D Reconstruction, ECCV 2024 4. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets, arxiv 2023 5. DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision, CVPR 2024

Weaknesses

- Confusing title: Given the contributions, the title should rather focus on the proposed method. - The title of the paper suggests a focus on benchmarking existing methods. - The paper still proposes a new approach MVSplat360. - Evaluation is done on one existing benchmark (RealEstate10k) and one newly proposed one leveraging the already existing DL3DV dataset. - The problem of generalizable novel view synthesis given sparse views is not novel (see baselines [1], [2], [3]) such that the benchmark creation only consists of the definition of input and target views. - Insufficient contextualization relative to prior work: - Inaccurate use of the term 'concurrent work': - In line 60, ReconFusion [4] and latentSplat [3] (again in line 244) are referred to as concurrent work. - This is not really the case anymore, because - ReconFusion [4] appeared on arxiv in the beginning of December 2023 and was presented at CVPR 2024. - latentSplat [3] appeared on arxiv end of March 2024, 3 days after MVSplat [2], which is one of the building blocks of the proposed method. - Missing references to related work regarding conditioning via feature rendering and CLIP embeddings: - In lines 163-168 and 196-201, the authors propose to render latent features as spatial conditions for a generative model (Stable Video Diffusion), which is later ablated in lines 296-299. - This idea is not novel and the paper is missing important references in this context: - GeNVS [5] proposed to use rendered pixelNeRF features as conditoning for a 2D diffusion model. - ReconFusion [4] adopts this to condition a text-to-image latent diffusion model. - latentSplat[3] also renders feature maps that are decoded to a novel view in a GAN setup. - In lines 193, the authors describe the use of CLIP image embeddings of the input views as global conditioning. - ReconFusion [4] desribes a very similar procedure. - Lack of clarity regarding the proposed method: - From the paper and supplement, it is not clear, for which loss terms the architecture is optimized during training. - Since the paragraph about "Multi-frame diffusion model" starting in line 174 mainly describes SVD [6], it is not clear, whether the loss in equation (2) is also the only loss for MVSplat360. - In line 167f., the authors claim that joint training of MVSplat and SVD can further enhance geometry through the feature conditioning. - Is that the only source of gradient signals to the structural parameters of the 3D Gaussians (location, scale, rotation...)? - ReconFusion [4] and latentSplat [3] both additionally optimize direct RGB renderings to regress the ground truth for better geometry gradients. - The role of the video diffusion model compared to a single-view diffusion model is not clear. - The paper misses to explain, whether and if so how novel views are refined jointly (see questions). - The proposed approach is incremental. - Compared to ReconFusion, MVSplat360 uses MVSplat instead of pixelNeRF and a video instead of an image diffusion model. - Compared to latentSplat, it replaces the pixelSplat encoder with the improved MVSplat and a small CNN-based decoder with SVD [6]. - Unclear conclusion of benchmarking: - latentSplat builds upon pixelSplat, which is shown to be worse in depth estimation than MVSplat. - Regarding this baseline, it is difficult to conclude, how much the reconstruction and the refinement modules contribute to the performance. - If the goal of the paper is benchmarking, a component-wise evaluation (backbone feature extractors, depth estimation, refinement networks) would be more insightful. 1. pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction, CVPR 2024 2. MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images, ECCV 2024 3. latentSplat: Autoencoding Variational Gaussians for Fast Generalizable 3D Reconstruction, ECCV 2024 4. ReconFusion: 3D Reconstruction with Diffusion Priors, CVPR 2024 5. GeNVS: Generative Novel View Synthesis with 3D-Aware Diffusion Models, ICCV 2023 6. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets, arxiv 2023

Questions

- Are the structural parameters of the Gaussians purely trained via the SVD loss through the features? - Does the for features extended 3D Gaussians rasterizer compute gradients through the features to the structural parameters of the Gaussians? - If so, what is the intuition why this apparently works, while baselines (latentSplat and ReconFusion) use auxiliary losses? - Are novel views refined jointly by the video diffusion model? - If yes, how do you deal with long videos? - Is temporal ordering leveraged or is the model agnostic to temporal permutations of target camera poses? - What are implications regarding 3D consistency? - How is the training done on RealEstate10k? - Do you train two different models (also for baselines) for inter- and extrapolation?

Rating

5

Confidence

5

Soundness

3

Presentation

2

Contribution

2

Limitations

The authors addressed limitations and societal impacts in the appendix.

Authorsrebuttal2024-08-12

Further response to follow-up comments (1/2)

We are grateful to see the follow-up comments. We address the additional concerns below. Feel free to let us know if you would like us to clarify anything further. > **A2: Discussions of related “concurrent” work** **Description of GeNVS** We agree that the Matterport3D dataset used by GeNVS is similar to RealEstate10K, and we will correct its description to “GeNVS mainly works on 360-degree object-centric scenes *or nearby viewpoint scene-level datasets*”. Nonetheless, it does not alter the previous conclusion as it still belongs to the two existing scenarios summarized in L28-L30 and Fig. I of the one-page PDF. The key contribution of MVSplat360 is that it focuses on an *unexplored* setting: feed-forward *360-degree* scene synthesis. The *majority of experiments are conducted on DL3DV* (rather than RealEstate10K), which significantly verifies MVSplat360’s effectiveness in handling 360-degree NVS from sparse views. **Discussions with ReconFusion** * We refer the reviewer to the project page of ReconFusion for the results from the diffusion model. In particular, as shown in the last video entitled “ReconFusion distils a consistent 3D model from inconsistent samples” on ReconFusion’s project page, views sampled purely from its diffusion module are far from consistent, showing obvious jittering from frame to frame. Our supplementary video shows that MVSplat360 renders multi-view novel views with much higher consistency. * Note that those inconsistent demos provided by ReconFusion mainly consist of forward-facing or *constraint* orbital camera trajectories, while our consistent demo contains way more challenging rendering trajectories, including different types of *unconstrained* trajectories. * It is necessary for ReconFusion to apply per-scene optimization in order to get satisfactory consistent novel views, while our MVSplat360 renders 3D consistent novel views in a feed-forward manner thanks to our effective design summarised in “Contributions of MVSplat360 (Method)” in the global response to all reviewers.   > **A3: Training objectives** We will update this potentially ambiguous claim regarding "enhancing geometry" in the paper. * It is correct that the reconstruction loss is the only source of gradients for the structural parameters of the Gaussians. * Our initial intention is that joint training can help “enhance the backbone features”. Since the features and other Gaussian parameters are from different heads but share the same backbone, we assume that enhancing the backbone feature will lead to a better reconstruction module and, hence better reconstructed coarse geometry. To avoid unintentional overclaim, we will tone down L168 to “the SVD loss can further optimize the Gaussian features, further *enhancing the reconstruction backbone*”.   > **A4: SD vs SVD** **Ablation of SD** In the updated version, we will make it clearer that we denoise all novel views jointly. We will also add visual comparisons with the SD-based ablation model, which shows inconsistent novel views compared to our default design, as observed in our experiments. **When novel views are far away from each other** We did not observe obvious limitations regarding this issue. This is probably because although the viewpoints range from 180 to 360 degrees for each scene in DL3DV, it still belongs to one scene and might not contain significant “far away” novel views. **Test time trajectories** * In our experiments, the set of novel views is always a plausible trajectory. In particular, for quantitative evaluation, we use the camera trajectories captured by the initial video. For the video demo, we apply a Gaussian filter to the initial captured camera path to obtain 6DoF stabilization results. * Each trajectory contains 56 frames, which are uniformly sampled from the initial video that contains around 300 frames. **Cross-view attention vs. concatenation** In typical single-view or multi-view diffusion-based NVS models, their input features come from the *source/observed viewpoints*. In this case, they are required to align/correct those source viewpoint features to match with the novel view cameras, hence it might be better to achieve via cross-view attention. In contrast, in MVSplat360, the features provided to the SVD are those rendered from *target/novel viewpoints*, containing coarse but correct geometry information for the target viewpoint, hence it is reasonable to achieve via concatenation. It might be helpful to add *additional source/observed view* features to the SVD refinement module via cross-view attention, keeping the current concatenation for *target/novel views*. We will explore this strategy in our further experiments. Thanks for the insightful suggestion!     *(see next comments for more response, thanks.)*

Authorsrebuttal2024-08-12

Further response to follow-up comments (2/2)

*(see previous comments for more response, thanks.)*     > **A7: More details of the refinement module SVD** **Temporal convolution** We do not apply similar operations to temporal convolution. We observe that applying temporal attention to all 56 frames leads to oversmoothness, so we change it to local attention with a window size of 14. In contrast, convolution is operated in local regions and has no such issues. **Discussions with CAT3D** Sorry for the unintentional confusion regarding “permutation.” Our initial intention was to emphasize that our refinement module takes *unconstrained* camera trajectories. We assume that the trajectory is plausible, as detailed in the above clarification regarding A4 (Test time trajectories), but we do not require it to be a meticulous design such as an orbital trajectory. In contrast, CAT3D does assume that its camera trajectory should meet several strict requirements, as detailed in its paper Sec. 3.2 and Appendix C.   *Feel free to let us know if you would like us to clarify anything further. Many thanks.*

Reviewer 4RiL2024-08-13

Thanks again for the discussion. The rebuttal addresses most of my concerns such that I would like to increase my rating to 5: borderline accept. Reasons for that are: - I agree with the significance and difficulty of the 360° NVS setting from sparse views on scene level and the strong performance of the proposed method on the DL3DV benchmark. - The discussion resolved the lack of clarity in the paper. Additionally, I would like to give the following suggestions for a final version: - The joint generation of novel views using a video (SVD) instead of an image diffusion model (e.g. SD) conditioned on features rendered from a 3D representation and the resulting 3D consistency is a very important part of the method and should be highlighted and further evaluated. - Regarding multi-view diffusion models as a recent trend, 3D consistency of the generated views seems to be the main bottleneck. - It would be very interesting to see how view-conditioning via features rendered from a 3D representation compete against alternatives like Plücker coordinates used in multi-view diffusion models. - For a more convincing evaluation of 3D consistency (compared to the SfM results in the rebuttal PDF), I would recommend a similar approach of mesh reconstruction from generated novel views as done in latentSplat. The incremental nature of the proposed method compared to previous works is the main reason that prevents me from giving an even higher rating.

Authorsrebuttal2024-08-13

We thank the reviewer for the thoughtful discussions. We are more than grateful for your recognition of our key contributions: our strong performance on the under-explored and challenging 360 scene synthesis setting. We will make the writing clearer following all of our discussions. We will also start working on your follow-up suggestions regarding highlighting and further evaluating SVD in terms of 3D consistency, the comparison with Plücker coordinates conditions and mesh reconstruction. Thanks again for your insightful suggestions for making this work more solid.

Reviewer 4RiL2024-08-11

Rebuttal Questions

I thank the authors for their clarifications regarding my concerns and providing additional results. Follwing the rebuttal, I still have some questions and concerns: > **A2: Discussions of related “concurrent” work** The discussion regarding related work is not completely accurate. GeNVS is not limited to object-centric scenes, but the paper includes results for the Matterport3D dataset, which is very similar to RealEstate10k. Moreover, ZeroNVS, ReconFusion and CAT3D do not necessarily require per-scene optimization of a NeRF, but can also be used directly for possibly inconsistent novel view synthesis. However, this is exactly comparable to the proposed MVSplat360. For obtaining a consistent 3D representation that allows fast rendering, the output novel views would need to be fused via optimization, e.g., of 3D Gaussians or a NeRF. > **A3: Training objectives** Could you elaborate more on my first question (and sub-points) in the questions section of my review? In line 167f. of the paper, you claim that joint training of MVSplat and SVD can further enhance geometry through the feature conditioning. Is that really the case, if the auxiliary reconstruction loss is the only source of gradients for the structural parameters of the Gaussians? > **A3: SV vs SVD** To me, the fact that you denoise novel views jointly was not clear from the paper. The comparison with SD sounds like an interesting insight that would be very nice to have in the paper as an ablation. However, I have some follow-up questions regarding this: - As you use concatenation along the channel dimension processed by a 3D UNet of SVD, I would expect problems if novel views are too far away from each other w.r.t. the camera pose. Is that the case? - How do you handle this at test time? - Does a set of novel views always have to be a plausible camera trajectory? - How long are these trajectories? - Would cross-view attention as done in multi-view diffusion models be a more suitable alternative than concatenation in channel dimension? > **A7: More details of the refinement module SVD** 1. SVD does not only consist of temporal attention, but also temporal convolutions. Is temporal consistency in the camera trajectory considered for this fact? 2. I am confused about this. If I am not mistaken, CAT3D conditions on individual camera poses, but does not assume any consistent temporal camera trajectory for this. SVD however is build for videos, e.g., by using temporal convolutions. Therefore, the architecture is tailored for temporally consistent camera poses, making it not agnostic to temporal permutations. Please correct me, if I am wrong.

Reviewer 3TXW2024-08-13

Thank you for your detailed rebuttal, which addressed my concerns on the experiments. I think overall this is a technically solid paper on an under-explored direction with good qualitative results. Since I still have concerns about the efficiency of leveraging a video diffusion model on the rendered Gaussian features, I will first raise my rating to 5-borderline accept, and discuss further with the other reviewers and AC in the later phase.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC