acceptodds
Under review as a conference paper at ICLR 2027

MVTrack4Gen: Multi-View Point Tracking for Persistent 4D Video Generation

Abstract

Synthesizing a novel-view video from a monocular reference video along a target camera trajectory requires geometric consistency and motion fidelity, both with respect to the reference video. Existing methods based on explicit 3D representations are limited by the accuracy of their reconstruction modules, which often produce inaccurate geometry for dynamic objects in monocular videos. In contrast, camera-conditioning-only methods can achieve high visual quality but often struggle to preserve geometric and motion consistency. In this work, we introduce MVTrack4Gen (Multi-View Point Tracking for 4D Video Generation), a novel framework that leverages multi-view point tracking as an additional geometric and motion supervision signal for camera-conditioning-only 4D video diffusion models. Our key finding is that specific attention layers in 4D video diffusion models encode strong correspondence cues, where query features attend to key features at geometrically corresponding locations across views and over time, and that misalignment of these correspondences causes motion inconsistency. Based on this observation, we route these features into an auxiliary multi-view tracking head and jointly train the 4D video diffusion model with a point-tracking objective. By explicitly strengthening these motion-aware correspondences, MVTrack4Gen improves existing models' ability to follow the motion in the reference view and maintain cross-view geometric consistency. Across diverse benchmarks, our method achieves state-of-the-art geometric consistency and competitive camera accuracy.

Primary area:
generative models
Keywords:
Dynamic Scene Novel View Generation, Video Diffusion Model, Video Generation

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_4KMaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MVTrack4Gen improves dynamic novel-view video generation through auxiliary multi-view point-tracking supervision. The paper analyzes temporal and cross-view correspondences in diffusion attention features, then attaches a tracking head to a selected intermediate layer and directly supervises attention correspondences. The framework also incorporates richer camera conditioning. Experiments on ReCamMaster and Redirector show improved geometric consistency across DAVIS, iPhone, and Neural3D, supported by component and layer ablations.

Critical strengths
  1. The method offers a practical way to improve geometric consistency without requiring explicit reconstruction during generation. On DAVIS, MEt3R decreases from 0.337 to 0.274 for ReCamMaster and from 0.318 to 0.267 for Redirector. Improvements across two backbones and additional real-world benchmarks make the contribution more convincing than a single-model demonstration.

  2. The attention analysis leads to a testable design choice. The component ablations in Table 3 support the usefulness of both tracking and correspondence supervision, while Table 4 supports selecting the correspondence-rich intermediate layer. The tracking head can also be omitted during generation, making the auxiliary training approach practically attractive.

Critical weaknesses
  1. The specific benefit of multi-view supervision over temporal tracking supervision remains insufficiently isolated. Track4Gen already establishes the usefulness of tracking supervision for video generation, as acknowledged in the related work. Table 3 separates the tracking and correspondence objectives, but neither comparison isolates within-view supervision from cross-view supervision. A matched temporal-only baseline would better establish the incremental contribution.

  2. The central claim of preserving reference motion is evaluated mostly indirectly. MEt3Rdynamic measures geometric consistency in dynamic regions, while VBench smoothness and consistency do not directly measure whether the generated subject follows the reference motion. Table 6 evaluates tracking capability rather than motion fidelity in generated videos. Consequently, the evidence for geometric improvement is stronger than the evidence for faithful motion transfer.

Questions for authors
  1. Under identical camera conditioning, training data, and optimization, how does temporal-only tracking/correspondence supervision compare with the full multi-view objectives? This would clarify the contribution beyond existing tracking-supervised generation.

  2. Can you report reference-motion fidelity using independently evaluated trajectories or synchronized ground-truth target tracks, including occlusions and large viewpoint changes? This would directly test the paper’s main motion-preservation claim.

4 comments

Thank you for your careful review and constructive suggestions. We appreciate your recognition of the improvements across backbones and datasets, the connection between our attention analysis and design choices, and the practical benefit of omitting the tracking head during generation. Below, we address your concerns through additional ablations and independent trajectory-based motion evaluation. The corresponding revisions are highlighted in blue in the revised manuscript.

Regarding W1 and Q1:

We acknowledge that the original manuscript did not provide sufficiently detailed ablation studies to clearly isolate the contributions of our design choices. To address this, we have added an ablation study on DAVIS comparing reference-view-only, target-view-only, and joint reference–target supervision, as reported in Table 3. All configurations use the same ReCamMaster backbone and Plücker camera encoding and are trained for 5K iterations under identical settings. In the single-view variants, both tracking supervision and the correspondence loss are restricted to temporal correspondences within the respective view.

ConfigurationAesthetic Quality ↑Imaging Quality ↑Temporal Flickering ↑Motion Smoothness ↑MEt3R ↓MEt3R_dynamic ↓Motion ↑Motion_dynamic ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
(i) Plücker camera encoding only0.5030.6870.9490.9830.2910.3670.4360.4333.430.12360.1740
(ii) (i) + Reference-view supervision0.4970.6800.9510.9830.2930.3720.4320.4284.140.14950.2168
(iii) (i) + Target-view supervision0.5020.6810.9500.9840.2760.3480.4460.4402.620.11610.1539
(iv) (i) + Joint reference–target supervision0.5000.6850.9520.9850.2560.3230.4560.4522.120.10470.1340

Target-view-only supervision improves geometric and motion consistency, and camera accuracy over the camera-encoding baseline, confirming the benefits of single-view supervision. However, joint reference–target supervision further improves all metrics across these three evaluation axes, while maintaining comparable visual quality. These results demonstrate the additional benefits of explicitly supervising correspondences across viewpoints. Single-view supervision in target view encourages temporally consistent trajectories within the generated video, but does not explicitly align them with the reference video. Joint supervision connects the same scene points across time and viewpoints, helping preserve the reference scene’s geometry and object motion under the target camera trajectory. We have clarified this distinction in Section 5.4 and included qualitative comparisons in Figures 7 and 20.

Regarding W2 and Q2 — Directly evaluate motion fidelity between the reference and generated videos.

We agree that geometric consistency, perceptual video quality, and tracking capability alone provide incomplete evidence of motion fidelity between the reference and generated videos. Following your suggestion, we have added Motion, the motion-consistency metric proposed in IC-World [1], to directly compare trajectories reconstructed from the two videos. We also report Motion_dynamic, which restricts the evaluation to foreground regions defined by the DAVIS ground-truth masks.

Specifically, we adapt the IC-World evaluator using SpatialTrackerV2, to reconstruct point trajectories from both reference and generated videos. We align the trajectories using the estimated relative camera pose, match tracks by their temporal-mean 3D positions, and compute the motion-consistency score from their distances at corresponding timestamps. This evaluates whether the generated motion agrees with the reference motion using an external trajectory estimator. The full protocol is provided in Appendix E.4.

DAVIS — Table 1

MethodMEt3R ↓MEt3R_dynamic ↓Motion ↑Motion_dynamic ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
GEN3C0.2900.3280.4340.4312.5380.1270.163
TrajectoryCrafter0.2910.3060.4070.40510.1260.1900.355
CogNVS0.3330.3460.4040.39810.4390.2280.400
NeoVerse0.3020.3230.4260.4244.7050.1590.228
ReCamMaster0.3370.3690.4050.4013.6600.1130.169
Redirector0.3180.3950.4380.4311.7140.0860.109
Camera Anything0.2760.3200.4600.4552.5770.1150.152
MVTrack4Gen (ReCamMaster)0.2740.2870.4690.4661.8580.1000.125
MVTrack4Gen (Redirector)0.2670.3490.4810.4761.7180.0730.097

On DAVIS, both variants of our method outperform the compared baselines in overall and dynamic-region motion consistency, including depth-warping and reconstruction-based methods. These results support the effectiveness of our supervision in preserving reference motion under viewpoint changes.

iPhone — Table 2

MethodPSNR ↑SSIM ↑LPIPS ↓MEt3R ↓Motion ↑
GEN3C10.4190.2700.6990.3820.736
TrajectoryCrafter10.2710.2760.7660.3790.718
CogNVS10.1050.2800.7740.4060.681
NeoVerse11.8860.3940.5790.4930.785
Redirector11.4470.3290.7200.5390.780
Camera Anything10.4550.2870.6630.4250.731
MVTrack4Gen (ReCamMaster)11.5210.3400.6400.3250.800
MVTrack4Gen (Redirector)11.9690.3380.6330.3470.788

On iPhone, both variants achieve better motion consistency than the compared baselines, including methods that explicitly reconstruct and warp scene geometry. Our Redirector-based model also improves visual quality and geometric consistency over the original backbone.

Neural3D — Table 2

MethodPSNR ↑SSIM ↑LPIPS ↓MEt3R ↓Motion ↑
GEN3C16.0720.5190.4660.2220.864
TrajectoryCrafter16.3800.5230.4960.2570.865
CogNVS16.4660.5140.4700.2430.817
NeoVerse15.2680.4890.5110.2600.816
Redirector16.1700.5110.4620.2130.870
Camera Anything14.3070.4730.5170.2440.854
MVTrack4Gen (ReCamMaster)16.0780.5000.4610.2100.875
MVTrack4Gen (Redirector)16.6940.5270.4470.1990.878

On Neural3D, our Redirector-based model achieves the best results across the reported visual-quality, geometric-consistency, and motion-consistency metrics. Both variants outperform the compared baselines in motion consistency, providing further evidence of improved reference-motion fidelity across datasets.

MultiCamVideo architecture ablation — Table 4

ConfigurationPSNR ↑SSIM ↑LPIPS ↓MEt3R ↓Motion ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
ReCamMaster13.230.5320.5370.4050.6325.8920.2324.31
(i) Plücker camera encoding13.690.5460.5050.3910.6494.9220.5023.01
(ii) (i) + Correspondence loss14.930.5540.4720.3860.6984.5718.6721.88
(iii) (i) + Tracking module14.490.5710.4560.3780.7133.4717.8619.43
(iv) (i) + Tracking module + Correspondence loss15.050.5710.4650.3630.7193.9017.4819.31

The MultiCamVideo ablation shows that adding either the tracking module or the correspondence loss improves motion consistency over the same camera-encoding baseline. Combining both objectives achieves the strongest geometric and motion consistency, supporting their complementary contributions. Qualitative comparisons in Figure 16 and the supplementary videos further illustrate how our method preserves reference motion across viewpoints.

Evaluation under large camera motions on DAVIS

To assess motion fidelity under large viewpoint changes, we report results on the 125 pairs of DAVIS evaluation samples with the largest camera-motion magnitudes, selected from our existing evaluation set.

MethodMotion ↑Motion_dynamic ↑
GEN3C0.3050.301
TrajectoryCrafter0.2750.273
CogNVS0.2930.285
NeoVerse0.2980.296
ReCamMaster0.2740.272
Redirector0.3050.296
Camera Anything0.3280.323
MVTrack4Gen (ReCamMaster)0.3340.332
MVTrack4Gen (Redirector)0.3400.333

Both variants of our method outperform all compared baselines in overall and dynamic-region motion consistency on this subset, with our Redirector-based variant achieving the best results on both metrics. These results show that our method maintains its advantage in preserving reference motion even under large camera motions.

Evaluation under occlusion

Quantitative evaluation under occlusion is challenging on real-world datasets because accurately identifying occlusion intervals and obtaining reliable corresponding trajectories across viewpoints is difficult at scale. To investigate this setting, we additionally evaluate our method on synthetic scenes with known object motion and occluder geometry. Specifically, we render a falling ball that becomes temporarily occluded in the reference view and examine whether the generated target video maintains consistent object motion through the occluded interval. As shown in Fig. 17, our method maintains a coherent ball trajectory in the target view even when the ball is temporarily occluded in the reference view, whereas several baselines exhibit missing or misplaced objects and inconsistent motion. These results provide qualitative evidence that our method better preserves motion consistency under reference-view occlusion.

[1] Wu, Fan, et al. “IC-World: In-Context Generation for Shared World Modeling.” European Conference on Computer Vision, 2026.

Dear Reviewer 4KMa,

Thank you for your thoughtful feedback. Your suggestions led us to compare single-view and multi-view supervision and to evaluate motion consistency across viewpoints. These additions strengthen the evidence that our method preserves scene geometry and motion across views.

If you have any further concerns, we would welcome the opportunity to discuss them further. Thank you again for your time and consideration.

Best regards, The Authors

Reviewer_oaG7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies camera-controlled novel-view video generation from a monocular reference video. It analyzes temporal and cross-view correspondences in the query–key features of camera-conditioned video diffusion models and identifies intermediate layers with strong correspondence signals. Building on this analysis, MVTrack4Gen jointly trains the generator with an auxiliary multi-view tracking head and a direct correspondence loss, together with richer camera conditioning. The method is evaluated with ReCamMaster and Redirector on DAVIS, iPhone, and Neural3D, and the appendix provides component ablations, layer-selection experiments, correspondence analysis, and inference-cost measurements.

Critical strengths
  1. The method addresses an important limitation and provides improvements across backbones and datasets. Preserving reference-scene dynamics under viewpoint changes is an important challenge beyond producing visually plausible videos. On DAVIS, the method reduces MEt3R from 0.337 to 0.274 for ReCamMaster and from 0.318 to 0.267 for Redirector. The Redirector variant also improves reconstruction and geometric-consistency metrics on iPhone and Neural3D. These results support the practical value of correspondence supervision for dynamic novel-view generation.

  2. The appendix provides useful evidence supporting the proposed training objectives. Table 3 compares variants fine-tuned for the same number of iterations with a shared camera-encoding baseline, showing benefits beyond the camera-conditioning modification alone. The layer ablation in Table 4 and correspondence improvements in Table 5 are also consistent with the method's motivation. Furthermore, the tracking head is not required during video-only generation, making the approach practically attractive.

Critical weaknesses
  1. The contribution beyond existing temporal tracking supervision is not fully isolated. Table 3 evaluates the proposed components but lacks a matched temporal-only baseline, making it difficult to determine how much improvement specifically requires cross-view supervision.

  2. Reference-motion fidelity is evaluated indirectly. MEt3R and video-quality metrics provide useful evidence but do not directly measure whether generated objects follow the reference trajectories. Evaluation scene counts, MultiCamVideo scene splits, and dynamic-mask construction also need clarification.

Questions for authors
  1. How does the full method compare with an otherwise matched variant using only within-view temporal supervision?

  2. Can you provide direct evidence of generated-motion fidelity on held-out scenes with known trajectories?

  3. Please clarify evaluation scene counts, scene-level separation between training/analysis/evaluation, and the construction of MEt3R_dynamic masks.

6 comments

Thank you for your careful review and constructive suggestions. We appreciate your recognition of our improvements across backbones and datasets, supporting ablations, and the practical benefit of omitting the tracking head at inference. Your comments have helped us strengthen the experimental comparisons and clarify the evaluation protocol. Below, we address your concerns through additional ablation studies, motion-consistency evaluations, and clarifications of the dataset and evaluation protocols. The corresponding revisions are highlighted in blue in the revised manuscript.

Regarding W1 and Q1:Isolate the contribution of cross-view point tracking supervision.

We admit that the original manuscript did not provide sufficiently detailed ablation studies to clearly isolate the contribution of cross-view supervision. To address this, we have added an ablation study on DAVIS comparing reference-view-only, target-view-only, and joint reference–target supervision on top of the same Plücker camera encoding baseline, as reported in Table 3. All variants use the same ReCamMaster backbone and are trained for 5K iterations under identical settings. For the single-view variants, both tracking supervision and the correspondence loss are restricted to temporal correspondences within the respective view.

ConfigurationAesthetic Quality ↑Imaging Quality ↑Temporal Flickering ↑Motion Smoothness ↑MEt3R ↓MEt3R_dynamic ↓Motion ↑Motion_dynamic ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
(i) Plücker camera encoding only0.5030.6870.9490.9830.2910.3670.4360.4333.430.12360.1740
(ii) (i) + Reference-view supervision0.4970.6800.9510.9830.2930.3720.4320.4284.140.14950.2168
(iii) (i) + Target-view supervision0.5020.6810.9500.9840.2760.3480.4460.4402.620.11610.1539
(iv) (i) + Joint reference–target supervision0.5000.6850.9520.9850.2560.3230.4560.4522.120.10470.1340

Target-view-only supervision improves geometric and motion consistency, and camera accuracy over the camera-encoding baseline, confirming the benefits of single-view supervision. However, joint reference–target supervision further improves all metrics across these three evaluation axes, while maintaining comparable visual quality. These results demonstrate the additional benefits of explicitly supervising correspondences across viewpoints. Single-view supervision in target view encourages temporally consistent trajectories within the generated video, but does not explicitly align them with the reference video. Joint supervision connects the same scene points across time and viewpoints, helping preserve the reference scene’s geometry and object motion under the target camera trajectory. We have clarified this distinction in Section 5.4 and included qualitative comparisons in Figures 7 and 20.

Regarding W2 and Q2:Directly evaluate motion fidelity between the reference and generated videos.

We acknowledge that our original evaluation relied on geometric and perceptual metrics that did not directly measure reference-motion fidelity. Following your suggestion, we have added metric named Motion, proposed in IC-World[1], to evaluate the motion consistency between reference and generated video. We also report Motion_dynamic, which restricts this evaluation to foreground regions defined by the DAVIS ground-truth masks.

We adapt the IC-World evaluator using SpatialTrackerV2 to reconstruct point trajectories from both reference and generated videos. After aligning them using the estimated relative camera pose, we match tracks by their temporal-mean 3D positions and measure their distances at corresponding timestamps. The full protocol is provided in Appendix E.4.

DAVIS — Table 1

MethodMEt3R ↓MEt3R_dynamic ↓Motion ↑Motion_dynamic ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
GEN3C0.2900.3280.4340.4312.5380.1270.163
TrajectoryCrafter0.2910.3060.4070.40510.1260.1900.355
CogNVS0.3330.3460.4040.39810.4390.2280.400
NeoVerse0.3020.3230.4260.4244.7050.1590.228
ReCamMaster0.3370.3690.4050.4013.6600.1130.169
Redirector0.3180.3950.4380.4311.7140.0860.109
Camera Anything0.2760.3200.4600.4552.5770.1150.152
MVTrack4Gen (ReCamMaster)0.2740.2870.4690.4661.8580.1000.125
MVTrack4Gen (Redirector)0.2670.3490.4810.4761.7180.0730.097

On DAVIS, both variants of our method outperform the compared baselines in overall and dynamic-region motion consistency, including depth-warping and reconstruction-based methods. These results support the effectiveness of our supervision in preserving reference motion under viewpoint changes without requiring explicit reconstruction during generation.

iPhone — Table 2

MethodPSNR ↑SSIM ↑LPIPS ↓MEt3R ↓Motion ↑
GEN3C10.4190.2700.6990.3820.736
TrajectoryCrafter10.2710.2760.7660.3790.718
CogNVS10.1050.2800.7740.4060.681
NeoVerse11.8860.3940.5790.4930.785
Redirector11.4470.3290.7200.5390.780
Camera Anything10.4550.2870.6630.4250.731
MVTrack4Gen (ReCamMaster)11.5210.3400.6400.3250.800
MVTrack4Gen (Redirector)11.9690.3380.6330.3470.788

On iPhone, both variants achieve better motion consistency than the compared baselines, including methods that explicitly reconstruct and warp scene geometry. Our Redirector-based model also improves reconstruction quality and geometric consistency over the original backbone, showing that the motion improvements are accompanied by gains in the generated views.

Neural3D — Table 2

MethodPSNR ↑SSIM ↑LPIPS ↓MEt3R ↓Motion ↑
GEN3C16.0720.5190.4660.2220.864
TrajectoryCrafter16.3800.5230.4960.2570.865
CogNVS16.4660.5140.4700.2430.817
NeoVerse15.2680.4890.5110.2600.816
Redirector16.1700.5110.4620.2130.870
Camera Anything14.3070.4730.5170.2440.854
MVTrack4Gen (ReCamMaster)16.0780.5000.4610.2100.875
MVTrack4Gen (Redirector)16.6940.5270.4470.1990.878

On Neural3D, our Redirector-based model performs best across the reported visual-quality, geometric-consistency, and motion-consistency metrics. Both variants also outperform the compared baselines in motion consistency, providing further evidence that our supervision helps preserve reference dynamics across different datasets.

MultiCamVideo architecture ablation — Table 4

ConfigurationPSNR ↑SSIM ↑LPIPS ↓MEt3R ↓Motion ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
ReCamMaster13.230.5320.5370.4050.6325.8920.2324.31
(i) Plücker camera encoding13.690.5460.5050.3910.6494.9220.5023.01
(ii) (i) + Correspondence loss14.930.5540.4720.3860.6984.5718.6721.88
(iii) (i) + Tracking module14.490.5710.4560.3780.7133.4717.8619.43
(iv) (i) + Tracking module + Correspondence loss15.050.5710.4650.3630.7193.9017.4819.31

The MultiCamVideo ablation shows that adding either the tracking module or the correspondence loss improves motion consistency over the same camera-encoding baseline. Combining them achieves the strongest geometric and motion consistency, supporting the contribution of both objectives beyond camera encoding alone.

Qualitative comparisons in Figure 16 and the supplementary videos further show that our model preserves reference motion more faithfully than existing baselines.

Reviewer_UP9GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework named MVTrack4Gen. It uses multi-view point tracking as additional geometric and motion supervision for camera-conditioned 4D video diffusion models. The authors find that specific intermediate attention layers encode strong temporal and cross-view correspondences. Based on this observation, the proposed method introduces a multi-view tracking head that shares query and key features from the diffusion model. It also introduces a multi-view correspondence loss to directly supervise the attention maps. The method is applied to two existing backbones, ReCamMaster and Redirector. Experiments on DAVIS, iPhone, and Neural3D demonstrate improvements in geometric and motion consistency.

Critical strengths
  1. The paper provides a detailed analysis of correspondence in 3D attention maps. The authors show that specific intermediate layers encode strong intra-view temporal and inter-view geometric correspondences. This observation provides a clear motivation for the proposed multi-view tracking supervision and correspondence loss.

  2. The proposed method achieves favorable experimental results across different benchmarks and backbones. The method improves geometric consistency on ReCamMaster and Redirector. On DAVIS, MEt3R decreases from 0.337 to 0.274 for ReCamMaster and from 0.318 to 0.267 for Redirector. These results demonstrate the effectiveness and generality of the proposed method.

Critical weaknesses
  1. Track4Gen has already demonstrated that jointly training video diffusion models with point tracking can improve generation consistency. The proposed method extends this idea to multi-view point tracking and introduces an additional correspondence loss. However, this paper does not provide a direct comparison with a single-view point-tracking baseline. This makes it difficult to determine the additional benefits of multi-view tracking over existing point tracking supervision.

  2. The proposed method improves camera encoding by incorporating both reference and target camera intrinsics and extrinsics. However, the main experiments compare MVTrack4Gen with the original backbones, which use different camera conditioning. I have some concerns that how much improvement comes from multi-view tracking supervision rather than improved camera encoding.

Questions for authors
  1. Can the authors compare the proposed method with a single-view point-tracking supervision baseline adapted to the same video diffusion backbone? This would help clarify the effectiveness of multi-view tracking supervision.

  2. How much improvement comes from multi-view tracking supervision rather than improved camera encoding? Is there any way to evaluate it?

4 comments

We sincerely thank the reviewer for the careful and constructive evaluation of our work. We appreciate your recognition of our detailed correspondence analysis in 3D attention maps as motivation for multi-view supervision, as well as the effectiveness and generality of our method across models. Your thoughtful comments and insightful suggestions have helped us examine our experimental comparisons and clarify the contributions of our design choices. Below, we address both concerns through controlled ablations reported in Table 3 and Table 4 of the revised manuscript. The corresponding revisions are highlighted in blue.

Regarding W1 and Q1: Comparison with single-view point-tracking supervision.

We acknowledge that the original manuscript did not provide sufficiently detailed ablation studies to clearly isolate the contributions of our design choices. To better assess the additional benefits of multi-view tracking as a source of geometric and motion supervision, we have added an ablation study on DAVIS, comparing reference-view-only, target-view-only, and joint reference–target point tracking supervision on top of the same Plücker camera encoding baseline, as reported in Table 3. All configurations use the same ReCamMaster backbone and are trained for 5K iterations under identical settings. In the single-view variants, both tracking supervision and the correspondence loss are restricted to the respective view, allowing us to directly assess the benefits of extending supervision across views.

ConfigurationAesthetic Quality ↑Imaging Quality ↑Temporal Flickering ↑Motion Smoothness ↑MEt3R ↓MEt3R_dynamic ↓Motion ↑Motion_dynamic ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
(i) Plücker camera encoding only0.5030.6870.9490.9830.2910.3670.4360.4333.430.12360.1740
(ii) (i) + Reference-view supervision0.4970.6800.9510.9830.2930.3720.4320.4284.140.14950.2168
(iii) (i) + Target-view supervision0.5020.6810.9500.9840.2760.3480.4460.4402.620.11610.1539
(iv) (i) + Joint reference–target supervision0.5000.6850.9520.9850.2560.3230.4560.4522.120.10470.1340

Target-view-only supervision improves geometric and motion consistency, and camera accuracy over the camera-encoding baseline, confirming the benefits of single-view supervision. However, joint reference–target supervision further improves all metrics across three evaluation axes: geometric consistency, motion consistency, and camera accuracy, while maintaining comparable visual quality. These results demonstrate the additional benefits of explicitly supervising correspondences across viewpoints. Single-view supervision in target view encourages temporally consistent trajectories within the generated video, but does not explicitly align them with the reference video. Joint supervision connects the same scene points across time and viewpoints, helping preserve the reference scene’s geometry and object motion under the target camera trajectory. We have clarified this distinction in Section 5.4 and included qualitative comparisons in Figures 7 and 20.

Regarding W2 and Q2: Separating the contributions of camera encoding and multi-view supervision.

We apologize for the unclear explanation in the original manuscript. Table 4 directly isolates the contribution of our proposed supervision from that of improved camera encoding. Configuration (i) adds only Plücker camera encoding to ReCamMaster, while configuration (iv) retains the same camera encoding and additionally incorporates the tracking module and correspondence loss. All variants (i)–(iv) are fine-tuned from ReCamMaster for 10K iterations.

ConfigurationPSNR ↑SSIM ↑LPIPS ↓MEt3R ↓Motion ↑mRotErr (°) ↓mTransErr ↓mCamMC ↓
ReCamMaster13.230.5320.5370.4050.6325.8920.2324.31
(i) Plücker camera encoding13.690.5460.5050.3910.6494.9220.5023.01
(ii) (i) + Correspondence loss14.930.5540.4720.3860.6984.5718.6721.88
(iii) (i) + Tracking module14.490.5710.4560.3780.7133.4717.8619.43
(iv) (i) + Tracking module + Correspondence loss15.050.5710.4650.3630.7193.9017.4819.31

With camera encoding held fixed, our supervision improves over configuration (i). (ii) and (iii) show that the correspondence loss and tracking module each improve the metrics over the same camera-encoding baseline. Their combination achieves the best geometric and motion consistency, supporting the complementary benefits of the two supervision components. These controlled comparisons show that our proposed supervision provides substantial improvements beyond camera encoding. And we have expanded the discussion in Appendix D.1 to make these contributions clearer.

I think the controlled experiments address my concerns about single-view supervision and camera encoding. I will update my rating to Weak Accept.

Dear Reviewer UP9G,

Thank you for taking the time to review the additional experiments. Your suggestion helped us show more clearly how multi-view supervision improves consistency across viewpoints. We appreciate how your feedback helped us strengthen the paper.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.