acceptodds
Under review as a conference paper at ICLR 2027

Demiurge: Forging Navigable 4D Worlds via Autoregressive Video Rerendering

Abstract

Recent advancements in video diffusion models have substantially empowered generative video rerendering, where novel-view videos can be synthesized from a single input video, facilitating a broad spectrum of applications such as world modeling, embodied intelligence, and robotic simulation. However, despite their impressive photorealism, existing video rerendering methods typically treat each camera trajectory in isolation, lacking the cross-trajectory consistency required to form a coherent and traversable 4D world. To address this limitation, we introduce Demiurge, a novel autoregressive video rerendering pipeline for navigable 4D world generation. Built upon a dynamic 4D point-cloud memory cache, Demiurge promotes cross-trajectory consistency through four key components: 1) an appearance video diffusion model that rerenders photorealistic novel-view videos along target camera trajectories conditioned on the 4D cache; 2) a cache-consistent video depth completion model that aligns known-region depths with cached geometry while completing missing regions, mitigating scale ambiguity and discontinuities during cache updates; 3) a memory-aware dual-stream warping strategy that provides high-fidelity visual guidance and occlusion cues for plausible novel-view video synthesis; 4) a geometry-aware video refinement model that further enhances the perceptual quality of synthesized videos. Experiments show that Demiurge achieves state-of-the-art performance with three key capabilities: 1) coherent spatiotemporal lattice construction over dense multi-view trajectories; 2) large-parallax video rerendering under challenging camera motions, such as rotations with no input-view overlap; 3) geometry-faithful outputs that can be directly optimized into high-quality 3D Gaussian Splatting representations with minimal artifacts.

Primary area:
generative models
Keywords:
4D World Generation, Autoregressive Video Rerendering

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UL3aOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Demiurge rerenders a monocular video along multiple camera trajectories while maintaining a shared, time-indexed point-cloud memory. Each iteration generates an RGB video, recovers its depth, and incorporates the resulting observations into memory. The system combines memory-conditioned appearance generation, cache-consistent depth completion, dual-stream warping for appearance preservation and occlusion reasoning, and optional geometry-aware refinement. Evaluation covers 128 videos with 14 target trajectories each, supplemented by large-parallax examples and per-timestamp 3D Gaussian reconstructions.

Critical strengths

1)The dual-stream design addresses a clear trade-off between preserving appearance and maintaining sufficient geometry for occlusion handling. Its contribution is supported by ablations: removing it reduces View Sync from 64.893 to 51.498 and increases P-MSE from 0.0010 to 0.0049.

2)The method shows consistent improvements across perceptual quality and reconstruction compatibility on 128 videos with 14 trajectories each. Together with the component ablations, these results support its practical value for cross-trajectory video rerendering.

Critical weaknesses

1)The consistency metrics provide useful but indirect evidence. P-MSE covers only input-visible regions, View Sync evaluates only the final frame without geometric verification, and 3DGS reconstruction is evaluated on its input views. A held-out or revisit test would better support consistency in newly generated regions.

2)The depth-conversion description needs clarification. The model predicts normalized disparity, but Eqs. (55)–(58) describe affine fits to depth without an explicit reciprocal conversion. This may be a notation omission, but the exact procedure is important for reproducibility.

Questions for authors
  1. Could you clarify how normalized disparity is converted to depth in Eqs. (55)–(58)? Is a reciprocal operation omitted from the description, or does the implementation directly fit disparity to depth?

  2. Do you have held-out or revisit results that assess consistency in newly generated regions, particularly beyond the final frame? Such evidence would help establish how well the reported gains extend to the central cross-trajectory consistency claim.

3 comments

Dear Reviewer UL3a, we are grateful for your careful review of our manuscript and for the constructive feedback you provided. We appreciate your recognition of how our dual-stream design balances appearance preservation and occlusion handling, as well as the consistent improvements in perceptual quality and reconstruction compatibility. We hope our responses below address the points you raised.

Q1. Disparity-to-depth conversion in Eqs. (55)–(58)

We sincerely thank you for pointing out this ambiguity. In fact, we perform global affine alignment in disparity space and apply the reciprocal conversion afterward. The alignment calibrates the normalized disparity prediction to the scale of the cached disparity ξ‾i=1/(D‾i+1)\overline{\boldsymbol{\xi}}^{i}=1/(\overline{\mathbf{D}}^{i}+1), where D‾i\overline{\mathbf{D}}^{i} denotes the partial cached geometric depth. Subsequently, we convert the aligned disparity ξ^i\widehat{\boldsymbol{\xi}}^{i} to geometric depth using Gi=1/(ξ^i+ϵinv)−1\mathbf{G}^{i}=1/(\widehat{\boldsymbol{\xi}}^{i}+\epsilon_{\mathrm{inv}})-1. Here, ϵinv\epsilon_{\mathrm{inv}} is a small constant (e.g., 1e-5) added to the denominator for numerical stability.

We have updated Eqs. (55)–(58) in the revised manuscript (Line 1034 ~ Line 1068). We sincerely thank you again for highlighting this point, which has helped us make the depth recovery procedure clearer and easier to reproduce.

Q2. Held-out reconstruction and consistency in newly revealed regions.

We sincerely thank you for raising this point. Following your suggestion, we extend our 3DGS evaluation to held-out views. We uniformly sample seven of the fourteen generated views to reconstruct the Gaussians. We render the reconstruction at the remaining seven viewpoints and compare the renderings with the corresponding generated frames. We apply the same view split to all methods.

MethodPSNR ↑SSIM ↑LPIPS ↓
TrajCrafter17.2840.5710.452
InSpatioWorld18.9070.6240.388
ReCamMaster19.8360.6490.406
GEN3C19.4240.6380.369
Voyager13.9720.4580.591
Spatia19.0250.6070.415
Demiurge22.7330.6830.317

As demonstrated above, Demiurge achieves the best results across all three metrics. Compared with the strongest baseline for each metric, it improves PSNR by 2.897 dB and SSIM by 0.034, while reducing LPIPS by 0.052. These results provide further evidence that our generated videos are more consistent across different viewpoints. We have included the held-out-view evaluation in the revised manuscript (Appendix M, Table 5, Lines 1662–1688).

To complement these quantitative results, we have added a qualitative comparison of newly revealed content across two trajectories in the revised manuscript(Appendix G, Line 1515 ~ Line 1548, Figure 7). In particular, at the intermediate timestamp T3T_3, both trajectories show the same region RR, which is absent from the corresponding input frame. The doorway and nearby artwork in this region have a consistent appearance and layout across the two generated views.

We sincerely thank you again for this thoughtful suggestion, which has helped us examine the consistency of newly revealed content before the final frame.

Thank you for the detailed rebuttal and for updating the manuscript.

The clarification of the disparity-to-depth procedure in Eqs. (55)–(58) addresses my reproducibility concern. Performing the affine alignment in disparity space and then applying the reciprocal conversion is now explicit, and the revised equations should make the depth recovery procedure easier to follow.

The new held-out 3DGS evaluation is also a useful addition. Reconstructing from seven generated views and comparing renderings at the remaining seven viewpoints is a more direct test of cross-view consistency than the original input-view protocol, and Demiurge is the strongest method under this split. The qualitative comparison of newly revealed content at an intermediate timestamp further helps support consistency beyond the final frame.

Taken together, these responses address the two issues raised in my review. I will keep my original rating of 3 (Weak acceptance).

Dear Reviewer UL3a,

We sincerely thank you for your reply and for carefully considering our responses and revised manuscript! We are pleased that our clarifications and additional results have addressed your concerns. We greatly appreciate your constructive feedback and continued positive assessment of our work.

Best,

Authors

Reviewer_8GZrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript introduces Demiurge, a video generation framework capable of creating navigable 4D world from a single input video. Its core idea is to condition video generation along different trajectories on a continuously updated 4D point-cloud memory, thereby achieving re-enacted consistently across trajectories. Specifically, to mitigate background leakage and appearance degradation, Demiurge maintains two logical memory states: the appearance memory aggregates points only from newly synthesized regions, while the geometry memory incorporates all valid points. These point clouds are back-projected from depth completion on generated videos. During video generation, the depth and appearance condition are constructed by rendering both memories and selecting between them according to their depth discrepancy.

Critical strengths
  1. The idea of maintaining both appearance and geometric memory and dual-stream warping sounds overall interesting and reasonable.
Critical weaknesses
  1. In the proposed method, the point cloud memory associated with each frame is updated independently, without considering cross-frame relationships. This could lead to some potential issues. For example, suppose trajectory A observes a new region at frame 3, whilst trajectory B observes the same region at frame 20. The current design appears to provide no mechanism for ensuring their consistency.
  2. The framework can only generate videos along different trajectories sequentially and relies on several heuristic point-cloud update strategies, which may limit its application to some tasks requiring parallel exploration.
Questions for authors
  1. How is the depth of the input video used to compute P-MSE obtained?
  2. Since the camera trajectories are already known in 3DGS reconstruction, why use a pose-free model? Besides, it is suggested to report reconstruction metrics on held-out views.
  3. Could the noise monitor mask based on depth variation cause boundary regions to be systematically lost during rendering? It would be helpful to add some visualizations of the mask.
4 comments

Dear Reviewer 8GZr, we sincerely thank you for your thoughtful assessment of our work and for sharing your constructive comments and suggestions. We appreciate that you found our paper clear and easy to follow and considered our method interesting and reasonable. We hope the following responses address your concerns.

Q1: Cross-frame consistency of the time-indexed memory.

We sincerely thank you for raising this important point. Our memory stores observations separately at each timestamp to preserve the corresponding dynamic scene state. While the memory updates are indexed by timestamp, both appearance generation and depth completion jointly process the video, allowing information to be shared across frames.

For the example you described, content observed at frame 3 of trajectory A can become part of B’s visual conditioning sequence after warping into frame 3 of trajectory B. The video model’s temporal attention can then use this observation to guide generation at other timestamps, including frame 20. Cross-frame consistency is therefore encouraged through joint video modeling, rather than through direct fusion of point clouds from different timestamps.

To maintain these visual anchors, our autoregressive generation follows progressively changing trajectories with overlapping visible regions. This helps retain historical observations in subsequent conditioning sequences, which the video model can then use across frames. We have included a qualitative comparison in the revised manuscript (Appendix E, Figure 5, Line 1439 ~ Line 1457) showing the same static background region R at (T=5) in trajectory A and (T=70) in trajectory B. The enlarged crops illustrate consistent appearance despite differences in viewpoint and observation timestamps.

We sincerely thank you again for this thoughtful comment, which has helped us explain more clearly how temporal modeling complements the time-indexed memory.

Q2: Sequential generation and applicability to parallel exploration.

We sincerely thank you for this thoughtful suggestion. In this paper, our primary goal is to construct a coherent 4D world from a single input video, since existing video rerendering methods typically synthesize each trajectory independently and struggle to maintain cross-trajectory consistency. To achieve this, we propose Demiurge, which autoregressively generates videos along target trajectories and updates a shared 4D memory after each generation, allowing subsequent trajectories to reuse previously generated appearance and geometry.

On this basis, as you suggested, parallel exploration is an important direction for future work. One possible approach is to use the consistent multi-trajectory videos produced by our framework to construct a training dataset for joint multi-trajectory generation. The appearance generator could then be extended to generate videos along several target trajectories together, while the depth-completion network recovers their corresponding depth videos. A corresponding memory-update mechanism could integrate the resulting RGB-D observations into the shared 4D memory before the next generation step. This would follow our generate–recover–update paradigm while extending each iteration from a single trajectory to multiple trajectories generated in parallel.

We have added this possible extension to the limitations and future-work discussion in the revised manuscript (Appendix P, Line 1739 ~ Line 1747). We sincerely thank you again for this constructive suggestion, which has helped us identify an important direction for extending the capabilities of our framework.

Q3: Input depth estimation for P-MSE evaluation.

We sincerely thank you for this question. The depth of the input video used to compute P-MSE is estimated using DepthCrafter. We project the input RGB-D video onto each target trajectory to obtain a reference RGB rendering and its visibility mask. For each input–trajectory pair, all methods are evaluated using the same input video, estimated depth, projected reference, and visibility mask, providing a consistent basis for comparison.

We have explicitly stated the depth source in the revised manuscript (Appendix D.1, Line 1328 ~ Line 1331). We sincerely thank you again for this helpful question, which has helped us clarify the evaluation protocol.

Q4: Pose-free 3DGS reconstruction and held-out-view evaluation.

We sincerely thank you for this question. Although the target camera trajectories are known, videos generated by several baseline methods may still deviate from the prescribed camera motion. Using the target poses directly for these outputs can introduce additional misalignment and degrade their reconstruction quality. We therefore use AnySplat for pose-free 3DGS reconstruction, allowing the poses to be inferred from the generated views rather than fixed to the prescribed trajectories. The same reconstruction procedure is applied to all methods.

Following your suggestion, we additionally evaluate reconstruction quality on held-out views. Specifically, at each of the three timestamps used in the main experiments, we uniformly select seven of the fourteen views as reconstruction inputs and reserve the remaining seven for evaluation. All methods use the same view split.

MethodPSNR ↑SSIM ↑LPIPS ↓
TrajCrafter17.2840.5710.452
InSpatioWorld18.9070.6240.388
ReCamMaster19.8360.6490.406
GEN3C19.4240.6380.369
Voyager13.9720.4580.591
Spatia19.0250.6070.415
Demiurge22.7330.6830.317

As shown above, Demiurge achieves the best results across all three metrics. Compared with the strongest baseline for each metric, it improves PSNR by 2.897 dB and SSIM by 0.034, while reducing LPIPS by 0.052. These results further support the cross-view consistency of our generated videos.

We have included the held-out-view evaluation in the revised manuscript (Appendix M, Lines 1662–1688). We sincerely thank you again for this thoughtful suggestion, which has helped us strengthen our reconstruction evaluation and present the experimental settings more clearly.

Q5:Boundary preservation and noise-mask visualization.

We sincerely thank you for raising this important point. Although the noise monitor can slightly erode the rendered conditioning near depth discontinuities, the generation model completes these locally masked regions, preserving the overall appearance and object contours in the final output.

To examine this effect directly, we conduct a controlled experiment with the target camera poses fixed to those of the input video. This setting introduces no viewpoint change relative to the input, allowing us to examine masking around existing depth discontinuities and compare the generated result directly with the original reference. We have included this experiment and the corresponding noise-mask visualizations in the revised manuscript (Appendix F, Figure 6, Line 1476 ~ Line 1513). In this example, the noise monitor masks narrow bands around the head and hand boundaries, producing small gaps in the rendered RGB condition. The generation model fills these gaps and produces continuous object contours, while retaining the overall appearance of the reference frame.

These observations support that slight boundary erosion in the conditioning can be accommodated by subsequent generation without noticeably compromising the overall visual quality.

We sincerely thank you again for this thoughtful suggestion, which has helped us clarify the effect of the noise monitor on boundary preservation.

I appreciate the authors' clarifications and discussions on specific design choices and experimental settings, as well as the newly provided results on held-out view evaluation. They can address most of concerns raised in my initial review. However, the response regarding cross-frame consistency assumes that “content observed at frame 3 of trajectory A can become part of B’s visual conditioning sequence after warping into frame 3 of trajectory B.”. But it obviously does not always hold. We cannot expect that trajectories A and B can always observe the same regions at the same frame. However, I consider the work is overall technically solid and therefore decide to keep my original positive rating.

Dear Reviewer 8GZr,

We sincerely thank you for your reply and for your thoughtful engagement in this discussion! We are pleased that our clarifications and additional results have largely addressed your concerns. We also appreciate your recognition of the technical soundness of our work and your decision to maintain your positive rating.

Regarding your further comment on cross-frame consistency, we apologize for the lack of clarity in our earlier explanation. As you pointed out, trajectories A and B may observe different regions at the same timestamp. Hence, during the autoregressive world construction process, we assume that the relevant observation from A remains visible after warping into B’s conditioning frame at that timestamp.

To help satisfy this assumption, we progressively select target trajectories with overlapping visible regions. This overlap helps retain historical observations in subsequent conditioning sequences, allowing the video model to use them across frames through temporal attention. As generation proceeds, these observations progressively expand the shared 4D memory into a coherent spatiotemporal lattice.

Once the scene content has been consistently accumulated across viewpoints and timestamps, the shared 4D memory supports rendering along user-specified trajectories. At this stage, two rendering trajectories can observe the same region at different timestamps, with their views obtained from the corresponding states of the same constructed 4D world.

We will make this more explicit in the revised manuscript. We sincerely thank you again for this thoughtful comment, which has helped us explain this aspect of our framework more clearly. If you have any further questions, we would be happy to provide additional clarifications.

Best,

Authors

Reviewer_iM1HOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to generate videos along multiple camera trajectories from a single monocular video while keeping the scene and its dynamics consistent across views. The method stores RGB-D observations from the trajectories in a time-indexed 4D memory, uses depth completion to align newly generated geometry with that memory, and applies dual-stream warping to preserve appearance and handle occlusions. The authors show quantitative and qualitative improvements over existing methods on a benchmark of in-the-wild monocular videos.

Critical strengths
  1. Making videos generated along different camera trajectories agree on a single dynamic scene is a practically useful problem. The shared 4D memory lets later generations reuse earlier content, while dual-stream warping addresses appearance drift and occlusion errors. The design choices follow directly from the problems the paper identifies.
  2. The evaluation covers video quality, camera control, cross-view matching, and 3DGS reconstruction at three time points. In Table 3, removing dual-stream warping reduces View Sync from 64.893 to 51.498, and replacing depth completion reduces it to 54.226. These results support the contribution of the proposed components to consistency.
Critical weaknesses
  1. It is unclear whether the evaluated baselines reuse previously generated trajectories. This makes it harder to assess the gains over existing approaches.
  2. The depth-completion ablation needs clarification. The full method uses global alignment and Poisson refinement in addition to the completion network. It is unclear whether the monocular-depth variant in Table 3 retains these steps, so the gains from the completion network and the subsequent alignment are not clearly separated.
Questions for authors

1. Comparison with a baseline that reuses results across trajectories

A comparison with a baseline that also reuses generated results across trajectories would be helpful. For example, could GEN3C be adapted to update its cache after each generated trajectory and then compared with the proposed method? This would help show how much the method adds when the baseline can also use its earlier generated results.

2. Does the monocular-depth ablation retain the same alignment steps?

Does the monocular-depth variant in Table 3 use the same global alignment and Poisson refinement? If not, could you add this comparison? I would like to understand how much the trained depth-completion network adds when both variants use the existing cache to constrain depth.

3. Could you report quantitative results for both single-stream variants?

Figure 4 already gives a clear visual comparison of the two single-stream variants, but Table 3 has only one “w/o Dual-stream Warping” entry. Which variant does this row represent? Adding the metrics for the other variant would make the ablation more complete.

4 comments

Dear Reviewer iM1H, we are grateful for your careful review and the valuable feedback you provided on our paper. We appreciate that you found our paper clear and easy to follow, and recognized the practical value of our task and the effectiveness of our key components. We hope the following responses address your concerns.

Q1: Comparison with a baseline that reuses results across trajectories.

We sincerely thank you for this thoughtful suggestion. In our main evaluation, all baselines are reproduced using their official implementations, following the inference settings described in the original papers, with each target trajectory synthesized independently from the same input video.

As you suggested, adapting GEN3C to reuse previously generated trajectories would provide a valuable controlled comparison for assessing the contribution of our proposed pipeline. To enable this comparison, we extend GEN3C’s 3D-cache update strategy to the dynamic multi-trajectory setting. Specifically, we maintain a separate cache for each timestamp, initialized from the corresponding source RGB-D frame. After generating a trajectory, we estimate depth for each generated frame using MoGe (the depth estimator used by GEN3C) and align it to the cached depth rendered at the same timestamp and target camera pose, retaining GEN3C’s original alignment procedure. The aligned RGB-D observation is then incorporated using its original buffer-update policy. These updated caches subsequently provide guidance for generating the next trajectory at the corresponding timestamps.

We denote this baseline as GEN3C + adapted cache. For comparison, we also integrate the same GEN3C appearance generator with the remaining Demiurge pipeline, as demonstrated in Appendix K. We refer to this configuration as GEN3C + our pipeline. Both configurations reuse previously generated observations, allowing us to assess the contribution of our geometry-recovery and memory-processing pipeline with the appearance generator held fixed. For completeness, we include the original GEN3C and full Demiurge results alongside these two configurations. All metrics follow the evaluation protocol used in the main experiments.

MethodFVD ↓Q-Align-V ↑P-MSE ↓View Sync ↑
GEN3C363.4004.2770.00846.562
GEN3C + adapted cache394.2644.2320.00752.683
GEN3C + our pipeline347.8624.4070.00259.528
Demiurge334.5184.5490.00164.893

As shown above, adapted cache updates yield a modest improvement in View Sync, accompanied by higher FVD. We attribute this trade-off to the quality of the reused RGB-D observations. Although GEN3C’s original depth alignment is retained, remaining depth errors can still warp RGB pixels to incorrect locations. Depth alignment also does not directly correct color or texture changes in the generated frames. Reusing these observations can therefore improve agreement across trajectories while also carrying visual artifacts into subsequent generations.

Our pipeline improves the reliability of these observations in complementary ways. Depth completion uses cached geometry during prediction, helping recover surfaces consistent with previous observations. Dual-stream warping preserves established appearance while using denser geometry to handle occlusions. With the GEN3C appearance generator held fixed, the improvements across all four metrics support the benefit of our integrated pipeline when both configurations reuse generated observations. Full Demiurge achieves further gains with our memory-guided appearance generator. We have included this comparison and the adaptation details in the revised manuscript (Appendix L, Line 1620 ~ Line 1660).

We sincerely thank you again for this valuable suggestion, which has helped us strengthen our experimental evaluation and more clearly demonstrate the contribution of our proposed pipeline.

Q2:Does the monocular-depth ablation retain the same alignment steps?

Thank you for highlighting this point. In fact, the monocular-depth variant in Table 3 uses the same global alignment and Poisson refinement as the full method. Both variants therefore use cached geometry to constrain the recovered depth. Only the depth-prediction module is replaced, with all other pipeline components unchanged. Under this shared post-processing, our depth-completion model improves View Sync from 54.226 to 64.893 and reduces FVD from 388.597 to 334.518. These results support the benefit of the trained completion model.

We have clarified the ablation settings in the revised manuscript (Table 3 caption, Line 433 ~ Line 435). We sincerely thank you again for your thoughtful feedback, which has helped us clarify the ablation settings and improve the presentation of this comparison.

Q3:Quantitative results for both single-stream variants.

We sincerely thank you for pointing this out. The “w/o Dual-stream Warping” entry in Table 3 corresponds to appearance-only warping, i.e., w/o Geometry Stream Warping in Fig. 4(d). Following the suggestion, we additionally report the results for geometry-only warping, i.e., w/o Appearance Stream Warping. All other components and evaluation settings remain unchanged.

VariantFVD ↓Q-Align-V ↑P-MSE ↓View Sync ↑PSNR ↑SSIM ↑LPIPS ↓
Appearance-only: w/o Geometry Stream Warping383.5254.4120.004951.49823.8860.7440.254
Geometry-only: w/o Appearance Stream Warping389.2734.3890.005456.32724.0620.7540.265
Full model: Dual-stream Warping334.5184.5490.001064.89324.3600.7690.237

As shown above, the two single-stream variants exhibit different trade-offs. Geometry-only warping achieves higher View Sync and PSNR/SSIM, while appearance-only warping provides better FVD, Q-Align-V, P-MSE, and LPIPS. This pattern is consistent with the complementary roles illustrated in Fig. 4(d), i.e., the appearance stream preserves stable visual content, while the geometry stream provides denser coverage for occlusion handling. Combining both streams achieves the best results across all metrics.

We have clarified the variant names and included results for both single-stream variants in the revised manuscript (Table 3, Lines 440–441). We sincerely thank you again for this constructive suggestion, which has helped us improve the completeness and clarity of the ablation study.

Thank you for the detailed rebuttal and additional experiments. My concerns have been largely addressed, and I appreciate the authors’ efforts to clarify and strengthen the paper.

Dear Reviewer iM1H,

Thanks for your reply! We truly value your involvement in this discussion. We are happy to know that our response has largely resolved your concerns.

If you have any further questions, please feel free to reach out—we are open to continued discussion.

Best,

Authors

Comments on the paper

1 comment

We sincerely thank all reviewers for their time, careful reading, and constructive feedback. We appreciate your recognition of the practical value of cross-trajectory consistency (iM1H), the complementary roles of the appearance and geometry streams (8GZr and UL3a), and the improvements in perceptual quality and reconstruction compatibility (UL3a). In response to your comments, we have added experiments and visualizations accordingly. We have marked the changes in blue in the revised manuscript and indicated their locations in the point-by-point responses below. We sincerely hope that our revisions and the responses below address your questions and concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.