acceptodds
Under review as a conference paper at ICLR 2027

SymRegFlow: Symmetry-Regularized Flow Matching for Video World Models

13 views

Abstract

Flow-matching-based multi-view world models generate realistic videos, but are commonly restricted to fixed camera rigs. Extending them to continuously varying camera poses requires paired pose–video observations with dense pose coverage, which are costly to acquire. We introduce SymRegFlow, a symmetry-regularized flow-matching framework for multi-view-consistent video generation across continuous viewpoints without ground-truth novel-view RGB supervision. For each target pose, SymRegFlow geometrically warps source views into noisy anchors and combines masked dual-anchor supervision with cross-anchor denoising-output consistency to mitigate anchor-specific errors. Under an affine Gaussian surrogate, we prove that suitable consistency regularization recovers the clean-reference optimum at fixed noise levels, strictly outperforming single- and merged-anchor baselines. Experiments on Cosmos-Drive-Dreams and nuScenes demonstrate high-quality, multi-view-consistent autonomous-driving video generation: on nuScenes, SymRegFlow achieves the lowest FVD and FVMD among the evaluated baselines, reducing FVD by over 31% relative to the best baseline, and source-conditioned inference also attains the best FID and instance preservation.

Primary area:
generative models
Keywords:
Symetry Regularization, Novel-view Video Generation, Video World Models, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KQcaOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes SymRegFlow, which utilizes noisy anchors obtained by projecting two source perspectives. Through effective region supervision and denoising prediction consistency under shared noise, it ADAPTS to pre-trained multi-view video models, achieving the generation of new perspectives without RGB supervision of the target perspective. In addition, the method introduces directional cross-view attention to reduce the interference of noise anchors on the generation of the original view.

Critical strengths

The method addresses the practical problem of lacking target perspective supervision by combining complementary geometric pseudo-supervision with consistent regularization, and the design is relatively clear. Post-training reasoning does not rely on anchor construction or geometric projection and has certain practical value. Ablation also verified the benefits of dual anchors and moderate consistency constraints.

Critical weaknesses
  1. The real driving video experiment lacks the target viewing Angle GT and is mainly evaluated through quality indicators such as FID/FVD and VBench, as well as the MEt3R cross-viewing Angle consistency index. These indicators can reflect authenticity and consistency, but they are not sufficient to directly verify the reconstruction accuracy under the specified camera. Especially for the MEt3R in the joint generation mode, the generated source perspective is used as a reference. Internal consistency does not mean it conforms to the real scene. The paper did not report pairing metrics such as PSNR, SSIM, and LPIPS on datasets with real reserved perspectives, so the verification of the accuracy of the new perspectives is still insufficient.
  2. Quantitative experiments only cover the interpolation of the midpoint of the camera ring on the seen scenes and do not verify the pose changes of the unseen scenes or more general cameras. The existing results are insufficient to support the wide applicability of the method for generating continuous perspectives.
Questions for authors
  1. Could supplementary experiments be conducted on driving datasets with real open views (such as WayveScenes101)? It is recommended not to use the RGB of the target camera during training, but only to use it as GT during testing. Report PSNR, SSIM and LPIPS in source-conditioned mode to distinguish the improvement of visual quality from the enhancement of the accuracy of target perspective reconstruction.
  2. Could the scenes that were not adapted be supplemented to verify whether the method can be generalized beyond the midpoint interpolation of the camera ring in the already seen scenes?
2 comments
AuthorsOpenReview ↗

Response to Reviewer KQca

Thank you very much for the constructive feedback and positive rating. In the following responses, "Q" denotes questions and "W" denotes weaknesses.

Q1/W1 — Target-view reconstruction accuracy.

Can you report PSNR, SSIM, and LPIPS for source-conditioned reconstruction of real held-out camera views (e.g., WayveScenes101), using target-view RGB only for evaluation?

Answer: Thank you for this suggestion. We agree that evaluation against real held-out camera RGB using PSNR, SSIM, and LPIPS would directly assess target-view reconstruction accuracy. For nuScenes/Cosmos datasets, withholding an existing rig camera leaves almost no overlap with the observed views. The generated target view Appearance is therefore underconstrained, and plausible generated content may largely differ from the recorded ground truth, in this case, pixel-wise reconstruction metrics cannot provide a meaningful measurement.

Although ground-truth novel-view videos are unavailable for our nuScenes evaluation, we report ATE and rotational RPE in updated Table 1 to assess camera-trajectory accuracy.

WayveScenes101 offers a better camera configuration, with substantial overlap between the held-out front camera and the two forward side cameras. However, it lacks annotated lane geometry and 3D object boxes required to construct the world-scenario controls used by our current Cosmos-Transfer2.5 implementation. Applying our pipeline therefore requires additional annotation estimation or adaptation to a compatible multiview flow-matching backbone without HD-map conditioning. Due to limited rebuttal time and resources, we are still working on the experiments on the suggested WayveScenes101 dataset.

Q2/W2 — Scene and viewpoint generalization.

Does SymRegFlow generalize to scenes excluded from adaptation and camera poses beyond midpoint interpolation between neighboring rig cameras?

Answer: We add lateral camera shift results at ±1, ±2, and ±4 m in Table 2, where SymRegFlow achieves the best FVD and second-best FID among the compared methods. Revised Sec. 3.2 (Fig. 5; Table 1) also evaluates prompts from scenes disjoint from our 30 adaptation scenes under fixed world-scenario controls, demonstrating diverse appearances with multi-view consistency.

We further investigate unseen-pose generalization in the SE(2) toy model using disjoint training and test rotations (Appendix D.1), yielding two observations:

  1. SymRegFlow improves geometric accuracy at poses excluded from training.
  2. Under a fixed training budget, denser sampling of training poses improves generalization, although the gains gradually diminish.

Due to limited rebuttal time and resources, we are still conducting more comprehensive pose generalization experiments on nuScenes.

Thanks for your detailed reply and your updated manuscript. My concerns have been addressed. Overall, the research topic is interesting, and the paper’s contributions are substantial, so I will maintain my score.

Reviewer_qhU6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SymRegFlow warps two observed views into the target view to create two sets of reference images, called anchors. It uses these references to fine-tune a video generator and encourages consistent outputs from the two anchors. The approach is practical, but its symmetry motivation and handling of unreliable references need clearer explanation.

Critical strengths
  1. The method is practical: it reuses a pretrained model, needs no real target-view images for fine-tuning, and requires no geometric warping during generation.
  2. Continuing to supervise the source views and limiting attention between views help preserve source-view content while learning to generate new views. Comparisons and examples also support reduced dependence on a single anchor.
Critical weaknesses
  1. The link between symmetry and the loss is still unclear. Equation (4) mainly encourages agreement at the same target pose, without directly showing how geometric relationships across poses are preserved. Figure 2 supports the method on a 2D rotation task, but the left panel omits the methods' generated outputs. The metric gains on the right also do not directly explain the role of symmetry (§2.2–2.3; Appendix C.5).
  2. Two references can agree because they share the same error. Both warps may contain the same depth or occlusion errors, while Proposition 1 assumes independent errors with a mean of zero. The paper acknowledges this limit. It should still explain how the anchors complement each other in practice and how source-view information and geometric conditions help correct shared errors (§2.4; §5).
  3. Two-way alignment needs a clearer rationale where only one anchor sees a region. Equation (4) still updates both predictions in these regions, without distinguishing their reliability within the consistency term. Supervision from valid anchors and access to the same source views can help. However, the paper should explain what prevents a less reliable prediction from pulling the other one away from the correct result (Eqs. (4)–(6); Algorithm 1).
Questions for authors
  1. How do the gains in Figure 2 connect to the symmetry motivation? Can generated examples show how the model corrects errors in the two anchors?
  2. When both anchors make the same error, what information helps the model determine the correct content of the new view?
  3. Why pull both predictions toward each other where only one anchor sees a region? What prevents the less reliable prediction from misleading the other one?
4 comments
AuthorsOpenReview ↗

Response to Reviewer qhU6

Thank you very much for the constructive feedback and positive rating. In the following responses, "Q" denotes questions and "W" denotes weaknesses.

Q1/W1/Q3/W3 — Symmetry motivation, geometric consistency, and bidirectional alignment.

How does the joint objective in Eqs. (4)–(6) connect agreement at the same target pose to the symmetry motivation and geometric consistency across poses? When only one anchor supports a region, why should consistency update both predictions, and what limits the less reliable prediction’s influence on the directly supervised one? How do Figure 2’s quantitative gains and generated examples demonstrate improved geometry and correction of anchor-induced errors?

Answer: We provide additional generated examples in Figure 17 (Appendix F) to illustrate removal of visible warping artifacts by SymRegFlow. We also provide generated MNIST examples in Figure 12 (Appendix D). Here we would like to provide explanations to clarify the connections between symmetry and our proposed symreg term.

Geometric relationships enter through the source-to-target warps and target-pose conditioning. Equation (4) encourages agreement between two constructions of the same target view across sampled poses, providing a symmetry-motivated consistency constraint. The connection to symmetry becomes clearer when the anchor and consistency losses are considered jointly. Consider a local analysis using binary visibility masks and treating the two predictions as free variables. Let p1p_1 and p2p_2 denote the two predictions at a target pixel supported only by anchor 1, whose value is a1a_1. Their joint contribution is

+ \lambda_c\frac{(p_1-p_2)^2}{N_u},$$ where $N_1$ counts valid pixels in anchor 1 and $N_u$ counts pixels valid in either anchor. The anchor loss constrains $p_1$ through the available supervision, while consistency acts as a bridge that transfers this supervision to the otherwise unconstrained $p_2$. For $\lambda_c>0$, this local objective is minimized at $p_1=p_2=a_1$. The same reasoning applies with the anchors exchanged. Thus, agreement is grounded in available anchor supervision: the joint objective encourages the two source-to-target constructions to produce a consistent target view, regardless of which anchor supplies supervision. This explains the symmetry motivation and why consistency extends over the union of their valid regions. Please see Eqs. (27)–(28) in Appendix F for the derivation and relative coefficient bound. Proposition 1 (proved in Appendix B) addresses error correction: under its assumptions, consistency penalizes sensitivity to anchor-specific corruption and can improve clean-reference prediction. As reported in Table 5 (Appendix D.1), adding consistency to otherwise identical dual-anchor training reduces pose MAE from $20.28^\circ$ to $12.35^\circ$ on the withheld rotation band, and from $17.38^\circ$ to $7.83^\circ$ on the disjoint $K=17$ evaluation grid. Pixel FD also improves in both protocols. **Q2/W2 — Shared anchor errors and complementary information.** When both anchors make the same error, what information helps the model determine the correct content of the new view? How does this relate to Proposition 1’s assumption of independent, zero-mean errors? **Answer:** In practical video warping, errors introduced by depth estimation and projection depend on scene appearance and viewpoint, so different source views produce different warping artifacts. This motivates modeling complementary corruptions, which Proposition 1 analyzes under the idealized assumption of independent, zero-mean errors. However, as acknowledged in the paper, the consistency term alone cannot identify or correct an error shared by both predictions. In this case, we rely on the learned priors of the pretrained model together with the available scene context. Conditioned on the observed source views, novel-view synthesis can be viewed as an imputation or inverse problem. The pretrained priors encourage globally coherent inpainting and therefore help suppress some artifacts shared by the anchors. If neither the observed views nor other conditioning provides evidence that resolves a shared error, the prior can encourage a plausible and coherent reconstruction.
AuthorsOpenReview ↗

Response to Reviewer qhU6

Q3/W3 — Bidirectional consistency under unequal anchor reliability.

Why pull both predictions toward each other where only one anchor sees a region? What prevents the less reliable prediction from misleading the other one?

Answer: As discussed in our response to Q1 (see also Eq. (27) in Appendix F), aligning both predictions over the union region under the joint objective helps promote multi-view consistency: the consistency term transfers supervision from one anchor to the corresponding region unsupported by the other. The directly supervised prediction remains constrained by its masked anchor loss. The consistency term alone, however, cannot prevent an unreliable prediction from influencing the other.

As shown in Eq. (28) in Appendix F, the relative weight of reverse consistency is bounded by λc\lambda_c. In addition, viewing novel-view synthesis as an imputation or inverse problem highlights how scene context and pretrained model priors can help mitigate the influence of unreliable anchors by encouraging coherent reconstructions. Figure 17 in Appendix F provides a qualitative illustration of artifact correction by the complete model.

Dear Reviewer,

Thank you again for your time and constructive feedback. We wanted to gently follow up on our latest rebuttal and would appreciate any of your thoughts when you have a chance. We would be happy to address any remaining concerns.

Best regards,

The Authors

Reviewer_qhU6OpenReview ↗

Response to the rebuttal.

Thank you for the detailed responses and additional results.

The clarifications help explain the motivation behind the method, although some aspects of the theoretical justification remain only partially resolved, particularly the connection to symmetry and the handling of shared anchor errors. Nevertheless, the empirical results are encouraging and support the method’s practical value.

I will therefore maintain my rating.

Reviewer_NGX3OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

SymRegFlow aims to make multi-camera driving video generation models usable beyond their fixed camera rig by generating a virtual camera view between two real rig cameras. To achieve this, SymRegFlow first warps each of the two neighboring camera videos into the target camera using estimated depth, producing two imperfect pseudo-images. The model is trained on these two pseudo-images in two passes that share the same noise, with a loss that makes the two novel-view predictions agree. An attention mask named DCVA prevents the noisy novel view from affecting the real camera views. On 30 nuScenes validation scenes, it reports the best FVD among the compared methods.

Critical strengths
  • The approach is simple and reasonable: the predictions obtained from two imperfect warps are trained to agree with each other.
  • It is backbone-agnostic and needs no warping at inference.
Critical weaknesses
  • The proposed model is trained on the test scenes (Sec. 3.2), yet it is compared with DiST-4D, which never saw these scenes. A fair comparison with a proper train/test split is required.
  • There are no results on the standard benchmark against SOTA methods.
    • The standard ±1/±2/±4 m lateral-shift protocol (used by DiST-4D [R10] on the same 30 nuScenes scenes) is not reported.
    • Table 1 only covers the virtual camera halfway between two neighboring rig cameras (α = 0.5).
  • Comparisons with the strongest and most closely related methods are missing. WorldSplat [R11] is absent, and it is unclear whether the DiST-4D row in Table 1 is the stronger DiST-S variant.
  • There is no evidence that the generated output depicts the requested camera pose; no metric measures whether the generated video actually follows the target camera.
  • Generation "across continuous viewpoints" (Abstract) is overclaimed. The method is quantitatively evaluated only at α = 0.5, the camera halfway between two rig cameras, and the camera embedding moves only along the ring between two neighboring rig cameras.
  • The paper does not compare with depth-projection-based methods such as TrajectoryCrafter [R7], ViewCrafter [R8] and EX-4D [R9].
  • The ablation study does not include a naive baseline that generates the novel view with the frozen Cosmos-Transfer2.5 and a fused camera embedding.

[R1] Noise2Noise: Learning Image Restoration without Clean Data. Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, Timo Aila. ICML 2018. arXiv:1803.04189. https://arxiv.org/abs/1803.04189 [R2] Ambient Diffusion: Learning Clean Distributions from Corrupted Data. Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alexandros Dimakis, Adam Klivans. NeurIPS 2023. arXiv:2305.19256. https://arxiv.org/abs/2305.19256 [R3] DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation. CVPR 2025. arXiv:2410.13571. https://arxiv.org/abs/2410.13571 [R4] ReconDreamer: Crafting World Models for Driving Scene Reconstruction via Online Restoration. CVPR 2025. arXiv:2411.19548. https://arxiv.org/abs/2411.19548 [R5] StreetCrafter: Street View Synthesis with Controllable Video Diffusion Models. CVPR 2025. arXiv:2412.13188. https://arxiv.org/abs/2412.13188 [R6] FreeSim: Toward Free-viewpoint Camera Simulation in Driving Scenes. CVPR 2025. arXiv:2412.03566. https://arxiv.org/abs/2412.03566 [R10] DiST-4D: Disentangled Spatiotemporal Diffusion with Metric Depth for 4D Driving Scene Generation. ICCV 2025. arXiv:2503.15208. https://arxiv.org/abs/2503.15208 (cited in the submission as Guo et al., 2025) [R11] WorldSplat: Gaussian-Centric Feed-Forward 4D Scene Generation for Autonomous Driving. ICLR 2026. arXiv:2509.23402. https://arxiv.org/abs/2509.23402 [R14] ArbiViewGen: Controllable Arbitrary Viewpoint Camera Data Generation for Autonomous Driving via Stable Diffusion Models. arXiv:2508.05236. https://arxiv.org/abs/2508.05236 [R15] World Simulation with Video Foundation Models for Physical AI (Cosmos-Predict2.5 / Cosmos-Transfer2.5). arXiv:2511.00062. https://arxiv.org/abs/2511.00062 (cited in the submission as Ali et al., 2025) [R16] FreeVS: Generative View Synthesis on Free Driving Trajectory. Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang. ICLR 2025. arXiv:2410.18079. https://arxiv.org/abs/2410.18079 (cited in the submission as Wang et al., 2025)

Questions for authors
  1. How does SymRegFlow perform when it is trained on scenes disjoint from the 30 test scenes?
  2. Can you report results under the standard ±1/±2/±4 m lateral-shift protocol on the same 30 scenes, as in DiST-4D [R10]?
  3. Which DiST-4D variant is reported in Table 1? How does SymRegFlow compare with DiST-S, WorldSplat [R11]?
  4. Could you estimate the camera pose of the generated video and check whether it correctly reflects the intended camera trajectory?
  5. Can you report quantitative results at other camera positions (α ≠ 0.5), and extrapolation experiments that show the method can generalize beyond the two rig cameras?
  6. How does SymRegFlow compare with depth-projection-based methods such as TrajectoryCrafter [R7], ViewCrafter [R8] and EX-4D [R9]?
  7. How does the frozen Cosmos-Transfer2.5 with a fused camera embedding (i.e., without SymRegFlow training) perform?
  8. Why do you evaluate at step 2,048, when the ablation training schedule is 100,000 steps and the demo videos are from step 4,096?
  9. Since there are 6 neighboring camera pairs, could you share the novel-view videos of all 6 pairs for the same scene and segment? (Only 4 of them are provided for scene-0012 seg00; back-back_left and back_left-front_left are missing.)

[R1] Noise2Noise: Learning Image Restoration without Clean Data. Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, Timo Aila. ICML 2018. arXiv:1803.04189. https://arxiv.org/abs/1803.04189 [R2] Ambient Diffusion: Learning Clean Distributions from Corrupted Data. Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alexandros Dimakis, Adam Klivans. NeurIPS 2023. arXiv:2305.19256. https://arxiv.org/abs/2305.19256 [R3] DriveDreamer4D: World Models Are Effective Data Machines for 4D Driving Scene Representation. CVPR 2025. arXiv:2410.13571. https://arxiv.org/abs/2410.13571 [R4] ReconDreamer: Crafting World Models for Driving Scene Reconstruction via Online Restoration. CVPR 2025. arXiv:2411.19548. https://arxiv.org/abs/2411.19548 [R5] StreetCrafter: Street View Synthesis with Controllable Video Diffusion Models. CVPR 2025. arXiv:2412.13188. https://arxiv.org/abs/2412.13188 [R6] FreeSim: Toward Free-viewpoint Camera Simulation in Driving Scenes. CVPR 2025. arXiv:2412.03566. https://arxiv.org/abs/2412.03566 [R10] DiST-4D: Disentangled Spatiotemporal Diffusion with Metric Depth for 4D Driving Scene Generation. ICCV 2025. arXiv:2503.15208. https://arxiv.org/abs/2503.15208 (cited in the submission as Guo et al., 2025) [R11] WorldSplat: Gaussian-Centric Feed-Forward 4D Scene Generation for Autonomous Driving. ICLR 2026. arXiv:2509.23402. https://arxiv.org/abs/2509.23402 [R14] ArbiViewGen: Controllable Arbitrary Viewpoint Camera Data Generation for Autonomous Driving via Stable Diffusion Models. arXiv:2508.05236. https://arxiv.org/abs/2508.05236 [R15] World Simulation with Video Foundation Models for Physical AI (Cosmos-Predict2.5 / Cosmos-Transfer2.5). arXiv:2511.00062. https://arxiv.org/abs/2511.00062 (cited in the submission as Ali et al., 2025) [R16] FreeVS: Generative View Synthesis on Free Driving Trajectory. Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang. ICLR 2025. arXiv:2410.18079. https://arxiv.org/abs/2410.18079 (cited in the submission as Wang et al., 2025)

3 comments
AuthorsOpenReview ↗

Response to Reviewer NGX3

We sincerely thank the reviewer for the detailed and constructive feedback. In the following responses, "Q" denotes questions and "W" denotes weaknesses.

Q1/W1 — Scene-disjoint evaluation and comparison fairness.

How does SymRegFlow perform when it is trained on scenes disjoint from the 30 test scenes? Can you provide a fair comparison with DiST-4D under a proper train/test split?

Answer: We follow the established scene-specific novel-view synthesis setting in this research area: fitting observed scenes and evaluating novel viewpoints within them, as in EmerNeRF [R1], Street Gaussians [R2], PVG [R3], OmniRe [R4], UniSim [R5], NeuRAD [R6], MARS [R7], SUDS [R8], and Neural Scene Graphs [R9].

To assess the generalization of SymRegFlow, we additionally evaluate prompts from nuScenes scenes disjoint from our 30 adaptation scenes. Revised Sec. 3.2, “Generalization to unseen prompts,” reports diverse appearances with multi-view consistency and good visual quality (Fig. 5, Table1) at the novel viewpoint. We also extend the SE(2) toy model analysis to disjoint training and test poses, demonstrating improved generalization to unseen rotations (Appendix D.1). Due to limited rebuttal time and resources, we are still conducting the same experiments as in DiST-4D, as requested.

[R1] Jiawei Yang et al. EmerNeRF: Emergent Spatial-Temporal Scene Decomposition via Self-Supervision. ICLR, 2024. [R2] Yunzhi Yan et al. Street Gaussians: Modeling Dynamic Urban Scenes with Gaussian Splatting. ECCV, 2024. [R3] Yurui Chen et al. Periodic Vibration Gaussian: Dynamic Urban Scene Reconstruction and Real-Time Rendering. IJCV, 2026. [R4] Ziyu Chen et al. OmniRe: Omni Urban Scene Reconstruction. ICLR, 2025. [R5] Ze Yang et al. UniSim: A Neural Closed-Loop Sensor Simulator. CVPR, 2023. [R6] Adam Tonderski et al. NeuRAD: Neural Rendering for Autonomous Driving. CVPR, 2024. [R7] Zirui Wu et al. MARS: An Instance-aware, Modular and Realistic Simulator for Autonomous Driving. CICAI, 2023. [R8] Haithem Turki et al. SUDS: Scalable Urban Dynamic Scenes. CVPR, 2023. [R9] Julian Ost et al. Neural Scene Graphs for Dynamic Scenes. CVPR, 2021.

Q2/W2/W3 — Standard lateral-shift benchmark.

Can you report results under the standard ±1/±2/±4 m lateral-shift protocol on the same 30 scenes, as in DiST-4D [R10]? How does SymRegFlow compare with state-of-the-art methods under this protocol?

Answer: Please see the updated lateral shift results in Table 2. SymRegFlow achieves the best FVD and the second-best FID across different shift distances.

Q3 — DiST-4D variant and strongest baselines.

Which DiST-4D variant is reported in Table 1? How does SymRegFlow compare with DiST-S and WorldSplat [R11]?

Answer: Please note that DiST-4D is a two-stage method, composed of DiST-T (a normal I2V model) and DiST-S (novel-view refinement), so DiST-T cannot perform novel-view generation. All results are taken from DiST-S. To the best of our knowledge, WorldSplat has no open-source checkpoints as of 13 Nov. 2026; thus, we cannot compare with it.

Q4/W4 — Camera-pose accuracy.

Could you estimate the camera pose of the generated video and check whether it correctly reflects the intended camera trajectory?

Answer: Please see the updated Table 1, where we include the camera accuracy metrics. SymRegFlow achieves the best/second-best ATE/RPE.

Q5/W5 — Viewpoint coverage and extrapolation.

Can you report quantitative results at other camera positions (α\alpha ≠ 0.5), and extrapolation experiments that show the method can generalize beyond the two rig cameras? Given that the camera embedding varies only along the ring between neighboring rig cameras, what supports the claim of generation “across continuous viewpoints”?

Answer: The updated lateral-shift experiments in Table 2 provide additional evidence that our model is capable of generating other novel-view videos. We emphasize that we present results for α=0.5\alpha=0.5 in the main paper because it is the most challenging pose according to the SE(2) toy model analysis; generation at other α\alpha values would be simpler because the corresponding poses are closer to the endpoint pose, providing more accurate context information. Besides, we also extend the SE(2) toy model analysis to disjoint training and test poses; please see Appendix D.1 for details. Due to limited rebuttal time and resources, we are still conducting the required experiments on disjoint/continuous viewpoints in nuScenes.

Q6/W6 — Depth-projection-based baselines.

How does SymRegFlow compare with depth-projection-based methods such as TrajectoryCrafter, ViewCrafter and EX-4D?

Answer: Please see the updated Tables 1 and 2. We have better performance than TrajectoryCrafter and EX-4D. ViewCrafter was developed for static-scene NVS and cannot directly perform our time-aligned 4D synthesis task without adapting its conditioning to dynamic point-cloud renders.

Q7/W7 — Frozen-backbone baseline.

How does the frozen Cosmos-Transfer2.5 with a fused camera embedding (i.e., without SymRegFlow training) perform?

Answer: Please see the updated Appendix E. A naive fused camera embedding on frozen Cosmos-Transfer2.5 cannot perform novel-view synthesis.

Q8 — Training schedule and evaluation checkpoint.

Why do you evaluate at step 2,048, when the ablation training schedule is 100,000 steps and the demo videos are from step 4,096?

Answer: The 100,000-step value is the configured scheduling horizon, not the training duration: we did not train the ablation models for 100,000 updates. We use LambdaLR with 25 linear warm-up updates followed by cosine decay to 1% of the peak learning rate. The long horizon keeps decay negligible over the adaptation interval: at the common evaluation point of 2,048 optimizer updates, the LoRA learning rate is approximately 4.995 × 10⁻⁴, or 99.9% of its peak (5 × 10⁻⁴). We did not tune this horizon; investigating its effect is left to future work. We clarify the scheduling horizon in Appendix C.

The demo filenames record training iterations (microbatches). Those runs use gradient accumulation of two, so iter004096 corresponds to 2,048 optimizer updates. Thus, the filename does not indicate twice as many optimizer updates.

Q9 — Complete six-pair qualitative results.

Since there are 6 neighboring camera pairs, could you share the novel-view videos of all 6 pairs for the same scene and segment? (Only 4 of them are provided for scene-0012 seg00; back-back_left and back_left-front_left are missing.)

Answer: Please check the updated supplementary files in nuScenes-Example-Videos/scene-0012_seg00_all_six_camera_pairs for all 6 pairs of the scene.

Dear Reviewer,

Thank you again for your time and constructive feedback. We wanted to gently follow up on our latest rebuttal and would appreciate any of your thoughts when you have a chance. We would be happy to address any remaining concerns.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for their constructive feedback, which have helped us significantly improve the clarity and quality of our manuscript.

We have responded to each review individually and highlighted the manuscript revisions in light blue. The main additions are:

  • Viewpoints beyond midpoint interpolation. We evaluated SymRegFlow on 30 nuScenes scenes under the standard lateral-shift protocol. It achieves the lowest FVD and second-lowest FID at every shift distance (±1, ±2, and ±4 m), including in comparisons with DiST-4D and the newly added EX-4D and TrajectoryCrafter baselines (Table 2).
  • Camera accuracy. We measured how closely the generated videos follow the intended novel camera pose. Source-conditioned SymRegFlow matches the SOTA at 0.155 m absolute trajectory error at the reported precision. Joint generation achieves the second-lowest rotational relative pose error (0.770°, Table 1).
  • Generalization. With scene controls held fixed, prompts from outside the 30 adaptation scenes produce varied appearances while preserving scene layout across views (Figure 5, Table 1). In a controlled rotation study, adding our consistency loss reduces mean pose error from 17.38° to 7.83° across 32 unseen angles (Appendix D, Table 5).
  • Role of the consistency loss. We explain the link between symmetry and our proposed regularization, and clarify the limits of that supervision when both views share an error in Appendix F. Figure 17 illustrates how the full model corrects visible warping artifacts.

We hope that we have adequately addressed reviewers' concerns. Please let us know if you have any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.