acceptodds
Withdrawn from ICLR 2027

View4D : Synchronized Multi-View Video Generation in the Open World

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present View4D, a generative method that synthesizes synchronized multi-view videos with moving cameras from a single input video. Video generation models have recently evolved toward the new paradigm of world models, yet they typically generate scenes along a single camera trajectory. While recent methods explore multi-view video generation in the open world, they remain limited to static cameras. To achieve multi-view videos with camera movement, we propose a multi-view attention module that incorporates camera poses as relative positional encodings to ensure cross-view consistency. We further design an optional geometry-injection module to seamlessly leverage the 3D structures extracted from the input video. Finally, a hybrid training strategy is introduced to effectively integrate synthetic multi-view videos with real-world images and videos. Extensive experiments demonstrate that View4D significantly outperforms existing state-of-the-art methods, achieving superior geometric consistency and visual fidelity on complex in-the-wild videos. Overall, our work marks a crucial step in advancing generative world models—transitioning from single-camera trajectories to synchronized 4D observation under continuous camera motion.

Primary area:
generative models
Keywords:
Camera-Controlled Video Generation, Multi-View Video Generation, World Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ksWrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new method, namely View4D, for generating synchronized multi-view video from an input video or a text prompt, along with camera poses. The proposed method adds a multi-view attention module into a pre-trained video diffusion model, while per-frame camera information is injected into the multi-view attention as relative positional encoding. A module named explicit geometric guidance is presented that uses depth maps to first obtain warped target frames and masks, and then fuse them with the original latents via a trainable encoder. In addition, the paper leverages both multi-view videos and multi-view images to train the video diffusion model. Experiments are conducted to show the effectiveness of the proposed method.

Critical strengths
  1. The paper explores an important and interesting problem that generates multi-view videos given moving camera trajectories.

  2. The paper is well organized and easy to follow.

  3. Experimental results show that the proposed method outperforms baselines in terms of geometric consistency and video quality.

Critical weaknesses
  1. The proposed View4D is trained on a newly constructed dataset, while the baselines are not trained or finetuned on the same data. Moreover, it seems that the test data differs from that used in baseline methods. Although the authors provide additional ablation studies in Table 4 to analyze the contributions of the training data and proposed architecture, the main comparisons are still based on models trained with different datasets. It remains unclear how much of the performance improvement comes from the proposed architecture rather than the additional training data. A comparison with baselines trained on the same data would make the results more convincing.

  2. As shown in the ablation results, the new dataset contributes significantly to the performance improvements. Although the authors provide additional details, more information about the dataset construction process, scene diversity, and data distribution would help readers better understand the dataset and reproduce the results. In addition, publicly releasing the dataset would be a valuable contribution to the community and facilitate future research.

Questions for authors
  1. The authors provide additional ablation studies in Table 4 to analyze the contributions of training data and architecture. However, the main baselines are still trained on different datasets. Would it be possible to train or finetune representative baselines on the same constructed dataset for a fairer comparison?

  2. The proposed method relies on off-the-shelf geometry models such as Pi3 to estimate camera parameters and depth maps. These estimates may be inaccurate in complex in-the-wild videos, especially those containing fast motion, occlusions, or low-texture regions. Although the authors provide qualitative results with unaligned depth guidance in Figure 11 and evaluate using different geometry estimators in Table 6, a systematic quantitative robustness analysis is still missing. It would be helpful to evaluate how different levels of errors in depth, camera intrinsics, and relative camera configurations affect the generation quality and cross-view consistency.

2 comments

Thank you for your detailed comments and insight suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:

1. Comparison with baseline trained on the same dataset (in response to Weakness 1 and Question 1)

We fine-tune SynCamMaster [1], a synchronized multi-view video generation method, on our dataset with the same backbone and training conditions. The results below highlight the importance of both our dataset and architecture. Training on our dataset produces more realistic and consistent videos, while our architecture is crucial for achieving cross-view camera accuracy and appearance consistency. We will include this comparison in the revised paper.

MethodFVD ↓KVD ↓RotErr-V ↓TransErr-V ↓PSNR-m ↑SSIM-m ↑MET3R ↓
SynCamMaster405.090.530.780.9312.930.3080.106
SynCamMaster w/ our data230.740.240.530.6013.760.3520.096
Ours215.270.210.110.0817.670.4400.092

[1] SynCamMaster: Synchronizing Multi-Camera Video Generation from Diverse Viewpoints. ICLR 2025

2. Dataset details and release (in response to Weakness 2)

We thank the reviewer for the suggestions. We provide more details about the dataset below:

  • Dataset construction: At each time step, we place 25 cameras uniformly along a full circle around a vertical axis in the scene. The radius is randomly sampled from 0.50–1.00 m for outward-facing cameras and 2.00–9.00 m for inward-facing cameras. We use a pinhole camera model with a horizontal field of view randomly sampled between 50° and 100°. Throughout each sequence, the cameras move together through the scene.
  • Scene diversity: Our dataset covers diverse indoor and outdoor environments, with variations in scene layout, lighting, weather, and object density. It also includes challenging conditions such as occlusions, reflective surfaces, and large depth variations, alongside characters with varied appearances and motion patterns.
  • Data distribution: Our dataset contains roughly equal inward-facing and outward-facing multi-view videos. Walking and running sequences are similarly represented, with a small number of sequences featuring stationary movements. The scenes primarily consist of natural outdoor environments and urban streets, complemented by some indoor settings.

We are very willing to release our dataset and to help advance progress in this area. To ensure the reproducibility of our work, we will provide a detailed document describing how to collect and construct the training data. We are actively working through the open-sourcing process. However, since the dataset is relatively large, the full release may require some additional time. In the meantime, we can first provide a subset of the data so that the community can quickly iterate on experiments and verify the correctness of the dataset.

3. Robustness analysis of geometric guidance (in response to Question 2)

The relative camera configurations {T0→j}j=1n\lbrace T_{0 \to j} \rbrace_{j=1}^{n} are specified by the user rather than estimated by a geometry foundation model, and serve as reference for evaluating cross-view camera accuracy in RotErr-V and TransErr-V.

To assess robustness to errors in estimated depth and camera intrinsics, we separately perturb the depth maps and focal lengths by 5%, 10%, and 20% (Ours_depth_x and Ours_focal_x, where x = 05, 10, or 20). We randomly select 10 video clips and run View4D with the perturbed geometric conditions, comparing all settings against the unperturbed baseline on the same clips. As shown below, our method maintains comparable generation quality and cross-view consistency across the tested perturbation levels, supporting its robustness to errors in depth and camera intrinsics from geometry estimation [2].

MethodFVD ↓KVD ↓RotErr-V ↓TransErr-V ↓PSNR-m ↑SSIM-m ↑MET3R ↓
Ours757.031.460.040.0221.250.5680.134
Ours_depth_05761.501.480.040.0321.220.5710.133
Ours_depth_10746.581.400.040.0420.970.5640.133
Ours_depth_20767.871.520.040.0321.320.5740.129
Ours_focal_05758.051.460.040.0221.310.5680.136
Ours_focal_10748.061.380.040.0420.760.5590.138
Ours_focal_20738.791.360.040.0320.410.5450.141

[2] π³: Permutation-Equivariant Visual Geometry Learning. ICLR 2026

Thank you for the authors' response. As demonstrated in the authors' reply, the dataset itself is a valuable contribution. I encourage the authors to include more details about the dataset construction in the final version and hope they will release the dataset in the future. I will maintain my weak accept score.

Reviewer_WjbVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents View4D, a framework for generating synchronized multi-view videos with moving cameras from a monocular input video. Built on a frozen Wan2.1 backbone, it introduces a parallel multi-view attention module conditioned on camera projection matrices through relative positional encoding. An optional geometry-guidance branch incorporates depth-based reprojections of the input video. Training combines synthetic multi-view videos, real multi-view images, and real driving videos. Evaluation on 150 SpatialVid clips shows improvements over the compared baselines on most metrics, supported by ablation studies and a small user study.

Critical strengths
  1. The paper addresses a useful problem: generating multiple camera-controlled videos while preserving the content and dynamics of an input video. The framework is clearly described, and the qualitative examples suggest improved structural coherence over several compared baselines.
Critical weaknesses
  1. The introduction presents synchronized multi-view video generation with moving cameras as an unaddressed task. However, Plenoptic Video Generation already addresses cross-view synchronization under novel camera trajectories through autoregressive video conditioning. Although the paper acknowledges this work, it distinguishes View4D primarily by its joint generation of all views. Joint versus autoregressive generation is a methodological distinction, it does not establish that the underlying task was previously unaddressed.

  2. The main technical ingredients also have close precedents: inter-view synchronization modules in SynCamMaster, relative projective camera encoding in Cameras as Relative Positional Encoding, point-based geometric guidance in ViewCrafter and TrajectoryCrafter, and mixed static-multiview/dynamic-video training in TrajectoryCrafter. The paper acknowledges several of these foundations but does not convincingly establish what substantial capability or technical insight emerges from their combination. Combining existing components can yield a valuable system, but demonstrating that value requires stronger comparisons against closely related designs under controlled training conditions.

  3. The claimed data gap also requires clarification. ReCamMaster’s MultiCamVideo dataset already contains synchronized videos of dynamic scenes captured by moving cameras, while Syn4D provides synchronized multi-view dynamic scenes with camera motion and dense geometric annotations. The authors should explain the specific limitations of these datasets relative to View4D’s requirements.

Questions for authors
  1. What capability does View4D offer that is absent from Plenoptic Video Generation, beyond the distinction between joint and autoregressive generation? Please clarify the scope of the claimed research gap.

  2. What specific requirements are not met by MultiCamVideo and Syn4D? Can the authors provide a controlled comparison demonstrating the added value of their dataset over existing moving-camera multi-view data?

  3. Can the authors provide an evaluation on held-out synchronized multi-view data with ground-truth geometry? For PSNR-m/SSIM-m, please also clarify which camera parameters are used for reprojection and report valid-mask coverage across methods.

  4. What are the practical limits of scaling the number of synchronized views? The main evaluation uses three target views, while Table 5 evaluates only camera-pose accuracy for two to five target views within a fixed total viewing range. Table 7 reports computational costs for up to eight target views but does not establish their visual quality or dynamic synchronization. Please clarify the maximum number of target views tested and report how cross-view appearance consistency, foreground motion synchronization, and disocclusion consistency change as the number of views and spatial coverage increase.

4 comments

Thank you for your detailed comments and insight suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:

1. Discussion of Plenoptic Video Generation (in response to Weakness 1 and Question 1)

We agree with the reviewer that describing the task as “unaddressed” was an overstatement. We have changed this wording to “underexplored” in the updated version.

The capability of View4D over Plenoptic Video Generation [1] is that our method can achieve better cross-view synchronization through joint prediction. However, a direct comparison is currently infeasible because its publicly available code lacks critical modules and the corresponding checkpoint, as documented in the issue linked in the footnote on page 18 of our submission (page 19 in the updated version). As an alternative, we implement two variants using an autoregressive scheme similar to that of Plenoptic Video Generation to evaluate the benefits of joint multi-view prediction:

  • Neoverse w/ iter. (Table 1): Autoregressive video generation using Neoverse [2], the strongest single-view baseline in our comparison.
  • Single-View w/ iter. (Table 3): Autoregressive video generation using the single-view variant of our model, with the same architecture and training conditions.

We summarize the results below. Our method achieves better overall visual quality and cross-view synchronization. These results support the benefits of joint generation over the evaluated autoregressive variants. We provide detailed analysis and visualizations in the paragraph titled “Analysis on multi-view joint prediction” in Section D of the supplementary material and we have moved this part to Section 4.3 of the main paper in the update version.

MethodFVD ↓KVD ↓RotErr-V ↓TransErr-V ↓PSNR-m ↑SSIM-m ↑MET3R ↓
Ours215.270.210.110.0817.670.4400.092
Single-View w/ iter.230.740.240.210.1115.730.3990.087
Neoverse w/ iter.260.970.280.201.7413.210.2850.151

[1] Plenoptic Video Generation. CVPR 2026

[2] NeoVerse: Enhancing 4D World Model with in-the-wild Monocular Videos. CVPR 2026

2. Analysis of the View4D multi-view video dataset (in response to Weakness 3 and Question 2)

We acknowledge that MultiCamVideo [3] and Syn4D [4] both contain synchronized dynamic scenes captured by moving cameras. Their main limitation is that camera motion is defined relative to a fixed point in 3D space, while object motion mainly occurs in place. In contrast, camera and object motion in in-the-wild videos are not subject to these restrictions. Below, we compare our dataset with these datasets in these two respects.

  • Camera motion. Existing datasets mainly define camera trajectories under object-centric spherical coordinates. MultiCamVideo [3] samples start positions within a character-centered hemisphere and sets camera movement relative to the distance to the center. Syn4D [4] defines the starting and ending camera positions in spherical coordinates around a reference center. In contrast, our cameras move through the scene without being anchored to a fixed reference point.

  • Object motion. Existing datasets mainly feature in-place object motion with limited global displacement. In contrast, our dataset also includes many dynamic objects moving through the scene, such as characters walking and running.

We further provide visual examples for better comparisons at https://anonymous-results-1122.github.io/view4d/1/.

To demonstrate the added value of our dataset, we conduct a controlled ablation by replacing it with MultiCamVideo while keeping the model and training conditions unchanged. As shown below, our dataset improves both visual quality and cross-view consistency.

MethodFVD ↓KVD ↓RotErr-V ↓TransErr-V ↓PSNR-m ↑SSIM-m ↑MET3R ↓
Ours (View4D data)215.270.210.110.0817.670.4400.092
Ours (MultiCamVideo)317.390.480.230.3314.100.3340.130

Moreover, we are very willing to release our dataset and to help advance progress in this area. To ensure the reproducibility of our work, we will provide a detailed document describing how to collect and construct the training data. We are actively working through the open-sourcing process. However, since the dataset is relatively large, the full release may require some additional time. In the meantime, we can first provide a subset of the data so that the community can quickly iterate on experiments and verify the correctness of the dataset.

[3] ReCamMaster: Camera-Controlled Generative Rendering from A Single Video. ICCV 2025

[4] Syn4D: A Multiview Synthetic 4D Dataset. ECCV 2026

3. Controlled ablation studies (in response to Weakness 2)

Our method combines relative-projection multi-view attention with an optional depth-reprojection module and a hybrid training strategy to efficiently generate synchronized multi-view videos with moving cameras. It achieves better visual quality and cross-view consistency than previous methods.

We conduct ablations under controlled training conditions to demonstrate the value of each component.

  • Inter-view synchronization modules: We implement single-view variants (Single-View and Single-View w/ iter. in Table 3). The results demonstrate that our method achieves better performance through joint prediction.
  • Relative projective camera encoding: We implement a learnable camera encoder variant (Ours w/ Cam Enc. in Table 3). The results demonstrate that our camera encoding improves cross-view consistency.
  • Point-based geometric guidance: We report results with explicit geometric guidance (Ours w/ guid. in Table 1). The results demonstrate that this guidance improves overall camera accuracy while having little impact on visual quality.
  • Hybrid training strategy: We implement two variants by removing each real-world dataset individually (Ours w/o MV-V and Ours w/o MV-I in Table 3). The results demonstrate that both real-world data sources improve our model’s robustness on real-world test cases.

4. Evaluation on a dataset with ground-truth geometry (in response to Question 3)

We evaluate our method and the baselines on 14 multi-view video clips from our dataset with ground-truth RGB images and depth maps. These clips are excluded from model training. We compare the generated target-view videos directly with the ground-truth RGB videos. As shown below, our method achieves the best performance across all three image quality metrics.

MethodOursNeoverse w/ iter.NeoverseSynCamMasterReCamMasterTrajectoryCrafter
PSNR ↑20.3617.6317.9415.9116.0717.03
SSIM ↑0.6580.5720.5940.5240.5570.592
LPIPS ↓0.4190.5270.5090.5520.5860.492

5. Details of the pixel-level fidelity metric (in response to Question 3)

For PSNR-m/SSIM-m, we use the input camera poses Tj,iT_{j,i} and intrinsic matrix KK described in Section 3 for reprojection. We will clarify this in the revised paper. We also report the valid mask coverage for each method below. Our method achieves the best PSNR-m/SSIM-m scores while maintaining the highest valid mask coverage.

MethodOursNeoverse w/ iter.NeoverseSynCamMasterReCamMasterTrajectoryCrafter
Coverage36.9%31.3%33.1%34.4%29.4%34.8%

6. Analysis of the number of target views (in response to Question 4)

Our method can scale to more synchronized views thanks to the design of camera conditioning with relative positional encoding. We extend the evaluation in Table 5 to eight target views, matching the computational cost analysis in Table 7. As shown below, the per-view rotation and translation errors remain stable as the number of target views increases. These results demonstrate that our framework can handle varying numbers of target views while maintaining geometric accuracy.

Number of target views2345678
RotErr-V / view ↓0.0270.0260.0250.0260.0260.0260.026
TransErr-V / view ↓0.0300.0200.0160.0160.0190.0190.018

We further randomly select 10 video clips and simultaneously increase the number of target views and spatial coverage. We evaluate PSNR-m and SSIM-m between adjacent views. As shown below, our method maintains cross-view consistency as both the number of views and spatial coverage increase. We provide visual results of eight target views at https://anonymous-results-1122.github.io/view4d/2/.

Number of target views2345678
Spatial coverage (deg)3045607590105120
PSNR-m ↑20.9021.4421.3421.4021.3721.2821.13
SSIM-m ↑0.5360.5600.5620.5740.5710.5700.563

Dear Reviewer WjbV,

Thank you for your time and thoughtful comments. We greatly appreciate your recognition of the usefulness of the problem, the clarity of our framework, and the performance improvements of our method over previous works.

As the discussion period will close on November 18 (AoE), we would be grateful to know whether our responses have addressed your concerns. Please let us know if any points require further clarification—we would be happy to provide it.

Thank you again for your valuable feedback.

Thank you for the additional experiments. I have several follow-up questions.

  1. Why would joint modeling yield better results than an autoregressive approach? How much of the reported improvement can be attributed to differences in the foundation models? My understanding is that the attention cost of joint generation grows quadratically with the number of views. Does this limit how many views the model can generate at once and make it difficult to generate videos from many novel viewpoints?

  2. The performance improvement demonstrates the value of your dataset. When you replaced it with MultiCamVideo, did both runs use the same amount of training data? Could you describe the setup of this comparison in more detail? I would like to understand how much of the improvement comes from the dataset itself and how much of that gain is due to the differences in motion.

  3. I still have questions about the technical novelty of the method. The core modules were introduced in prior work, and your approach primarily combines them and trains the resulting model on a larger dataset that you collected. Would existing methods achieve similar improvements if trained on your dataset? If so, where does the novelty of your architecture and algorithm lie?

  4. Thank you for extending the evaluation to eight target views. How does the actual inference time change as the number of target views increases? What is the maximum number of target views your method can generate jointly?

Reviewer_kMcaOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

View4D synthesizes synchronized videos of the same dynamic scene from multiple moving viewpoints, conditioned on a single input video. Its core component is relative-projection multi-view attention, which exchanges information across views at each time step while incorporating relative camera geometry. The model also supports optional depth-based geometric guidance. Although the reported results are competitive, the experiments do not fully isolate the contribution of the proposed projection-based attention design, leaving questions about its specific effectiveness and technical novelty.

Critical strengths
  1. The paper is well written and easy to follow, with a clear presentation of the overall framework.
  2. The proposed model achieves strong quantitative results.
Critical weaknesses

Generally, my main concerns are all about the effectiveness and novelty of the proposed method.

  1. There is no controlled ablation that isolates the contribution of the projection transformation in relative-projection multi-view attention. It remains unclear how the block-diagonal lifting matrix contributes to performance. An ablation in which this transformation is omitted would help clarify its effectiveness. A clearer intuitive explanation, supported by theoretical analysis or relevant references, would also be helpful.

  2. According to Table 1, the proposed geometric guidance provides mixed benefits rather than consistent improvements across metrics. Its contribution therefore appears incremental, and the paper should clarify when it is beneficial.

  3. Compared with single-view baselines such as ReCamMaster, the paper aims to generate multiple views of the same scene simultaneously. To this end, it introduces an attention module that exchanges spatial information across views. However, this is a straightforward approach to encouraging cross-view consistency, and the paper does not sufficiently establish the novelty of its specific attention design or its advantage over simpler alternatives.

Questions for authors

No questions for authors. Please refer to weakness items.

3 comments

Thank you for your detailed comments and insight suggestions. We will revise our paper based on your feedback. Here are our responses to your comments:

1. Ablation study on projection transformation (in response to Weakness 1)

We conduct an ablation in which the projection transformation is omitted by replacing it with an identity matrix. As shown below, removing this transformation substantially degrades cross-view consistency, indicating that the model fails to model the given multi-view camera configuration {T0→j}j=1n\lbrace T_{0 \to j} \rbrace_{j=1}^{n}.

MethodRotErr-V ↓TransErr-V ↓PSNR-m ↑SSIM-m ↑
Ours w/o trans.1.585.6811.990.290
Ours0.110.0817.670.440

2. Intuitive explanation of projection transformation (in response to Weakness 1)

As stated in the third paragraph of Section 3.2, our relative-projection multi-view attention is inspired by ProPE [1] and incorporates camera poses into attention. This design makes the network learn how each camera is situated relative to every other camera. To be specific, the transformation makes attention scores explicitly depend on the relative camera projection matrices, allowing cross-view interactions to account for camera geometry. Further discussion of the design can be found in the fourth paragraph of Section 3.2. This design [1] was originally developed for tasks such as sparse-view synthesis [2], while we adapt the underlying idea to synchronized multi-view video generation.

[1] Cameras as Relative Positional Encoding. NeurIPS 2025

[2] LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias. ICLR 2025

3. Effectiveness of geometric guidance (in response to Weakness 2)

As shown in Table 1, geometric guidance substantially improves overall cross-view consistency, with a 54.5% reduction in RotErr-V, a 25.0% reduction in TransErr-V, a 24.7% reduction in MSE (PSNR-m: 17.67 dB → 18.90 dB), and a 10.7% increase in SSIM-m, while causing only a slight increase in MET3R (3.3%). Meanwhile, it has little impact on overall visual quality, as reflected by the nearly unchanged FVD and KVD scores.

Geometric guidance is beneficial for regions that are visible in the input video, have high-confidence depth estimates, and remain unoccluded after warping to the target views. We provide detailed information on how geometric guidance is constructed in Section A.2 of the supplementary material.

4. Novelty of multi-view attention design (in response to Weakness 3)

We agree with the reviewer that multi-view attention is a common approach to encouraging cross-view consistency. Our contribution lies in how we incorporate the multi-view camera configuration into the attention module. To be specific, we effectively incorporate camera pose information as relative positional encoding, achieving better cross-view camera accuracy and generalization than the learnable encoders used in previous methods [3, 4]. This camera pose conditioning scheme, together with the optional geometric guidance module and hybrid data training strategy, enables our model to generate synchronized multi-view videos with moving cameras in the open world. This task remains underexplored, and we are encouraged that Reviewers WjbV and ksWr also recognize its usefulness and importance.

[3] ReCamMaster: Camera-Controlled Generative Rendering from A Single Video. ICCV 2025

[4] SynCamMaster: Synchronizing Multi-Camera Video Generation from Diverse Viewpoints. ICLR 2025

5. Advantages over simpler alternatives (in response to Weakness 3)

We evaluate the effectiveness of our multi-view attention design against simpler alternatives in Section 4.3. Specifically, we consider a variant that injects camera parameters through a learnable encoder, as in previous works [3, 4] (Ours w/ Cam Enc.). As shown in Table 3, our method achieves better visual quality and cross-view consistency.

We provide further analysis showing that our design generalizes better across diverse multi-view configurations in the paragraph titled “Analysis on camera pose conditioning” in Section D of the supplementary material and we have moved this part to Section 4.3 of the main paper in the update version. We also provide more visual results of multi-view videos in Figure 5 and the supplementary videos.

Dear Reviewer kMca,

Thank you for your time and thoughtful comments. We greatly appreciate your recognition of the clarity of our presentation, and the strong performance of our model.

As the discussion period will close on November 18 (AoE), we would be grateful to know whether our responses have addressed your concerns. Please let us know if any points require further clarification—we would be happy to provide it.

Thank you again for your valuable feedback.

Thank you for the detailed responses and additional experiments. I believe that most of my previous concerns, including those regarding the effectiveness of the projection transformation and geometric guidance, have been sufficiently addressed. My main remaining concern is the technical novelty of the proposed method.

My understanding is that the camera-conditioning mechanism in Section 3.2 largely adapts the projective encoding and attention transformation introduced in PRoPE [1] to a multi-view video diffusion setting. The new experiments provide convincing evidence that this adaptation is effective. However, I am still unclear about what constitutes the methodological advance beyond [1].

This distinction seems particularly important considering related developments in camera-controlled video generation, e.g., [2], together with existing multi-view video generation methods. I therefore encourage the authors to clearly separate the components inherited from [1] from those newly proposed in this work and, more importantly, explain what methodological modification or new formulation is necessary for the multi-view video setting.

In other words, beyond demonstrating that PRoPE-style conditioning works well when incorporated into the proposed framework, what is the specific algorithmic novelty of Section 3.2 compared with [1]? Clarifying this point would help resolve my main remaining concern regarding the paper.

[1] R. Li, B. Yi, J. Liu, H. Gao, Y. Ma, and A. Kanazawa, “Cameras as Relative Positional Encoding,” NeurIPS, 2025.

[2] M.-J. Kim, J. Kim, H. Jin, J. Hyung, and J. Choo, “Infinite-Homography as Robust Conditioning for Camera-Controlled Video Generation,” ECCV, 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.