acceptodds
Under review as a conference paper at ICLR 2027

UniCaMo: Unified Camera and Motion Control for Video Diffusion Models

Abstract

We present UniCaMo, an unified framework for joint control of camera and object motion in image-and-text-to-video diffusion models. Given a reference image, text prompt, sparse 3D object trajectories, and a 3D camera trajectory, UniCaMo constructs motion-consistent input noise that directly encodes the desired motion. Our 3D-grounded noise construction combines reference-derived noise propagation with a spherical Gaussian noise representation for newly revealed regions, providing a simple and effective motion prior without requiring reconstruction of unseen scene geometry. UniCaMo requires no architectural modifications or dedicated control modules and can be adapted using lightweight LoRA fine-tuning with low training data volume. Experiments on standard controllable video generation benchmarks demonstrate state-of-the-art motion controllability and high visual quality with minimal inference overhead. Qualitative videos are presented on the HTML webpage provided in the submitted supplementary zip files.

Primary area:
generative models
Keywords:
Video generation, video diffusion model, controllable video generation, computer vision

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_MQpVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents UniCaMo, a framework for jointly controlling camera and object motion in image-to-video diffusion models. Given a reference image, text prompt, sparse 3D object trajectories, and camera poses, the method constructs motion-consistent initial noise to guide video generation. Specifically, it propagates foreground noise along object trajectories, warps background noise according to camera motion, and introduces spherical Gaussian noise to handle newly revealed regions. These noise components are combined using depth-aware compositing and used to initialize a LoRA-finetuned video diffusion model.

Critical strengths

1.Simple and effective framework. The method provides a unified way to control camera and object motion through structured noise initialization, without introducing additional control modules or modifying the architecture of the pretrained video diffusion model.

2.Strong experimental performance. The method achieves competitive or superior results across multiple benchmarks, particularly in joint camera and object motion control.

3.Reasonable geometric design. The combination of foreground and background noise propagation, spherical noise, and depth-aware composition is intuitive. The ablation studies demonstrate the effectiveness of the proposed components.

Critical weaknesses

The paper lacks discussion and experimental comparisons with closely related recent methods. In particular, CameraNoise (ICML 2026), which explores geometry-guided noise warping for camera control, is not cited or discussed. Moreover, although VerseCrafter (CVPR 2026) has demonstrated joint control of camera and object motion, it is not included in the experimental comparisons.

Questions for authors

1.Non-rigid motion: Can UniCaMo explicitly control articulated or non-rigid object movements through independently specified 3D point trajectories? If so, could the authors provide examples or quantitative evaluations beyond object-level 6-DoF transformations?

  1. Robustness to random initialization: Since motion control is achieved through stochastic initial noise, how sensitive is the generated motion accuracy to different random seeds? Reporting the mean and variance of control errors across multiple seeds would help demonstrate the reliability of the method.

  2. The proposed spherical representation provides temporally consistent noise for unseen regions, but it does not model their actual depth or geometry. Although the ablation study demonstrates its effectiveness, further evaluation under large camera translations and complex occlusions would help clarify the limitations of this approximation.

  3. Trade-off between motion controllability and generation quality. The noise mixing ablation reveals a trade-off between motion accuracy and visual quality. Stronger structured noise improves trajectory fidelity but degrades FVD. This raises questions about how well the proposed noise initialization aligns with the pretrained diffusion prior and whether the trade-off becomes more severe for complex motions.

7 comments

Thank you for the review, and for pointing us to the two missing baselines. Both are now cited, discussed and compared, and we ran the experiments you asked for on seeds and on large motion. All additions to the paper are in red. We have also made an anonymous website to show the requested results explicitly.

Weakness: CameraNoise and VerseCrafter

  • CameraNoise is the closest work in spirit: it also writes the camera into the noise. It warps the noise with a reprojection flow computed from the camera poses and a constant pseudo-depth, so it uses no scene depth and does not control object motion. We now discuss it in Section 2 and compare with it on RealEstate10K. Overall, Unicamo excell on all 5 metrics using for camera-control. We have updated Table 2 with this new result. | Method | FID↓ | FVD↓ | RotErr↓ | TransErr↓ | CamMC↓ | |---|---:|---:|---:|---:|---:| | CameraNoise | 10.89 | 109.54 | 2.22 | 10.58 | 11.49 | | UniCaMo | 5.99 | 39.93 | 0.26 | 1.89 | 1.98 |

  • VerseCrafter controls camera and objects jointly through an additional adapter. We compare on camera control (300 RealEstate10K clips, on our joint benchmark (Table 3) and on a new stress set with much larger motion (Extreme-200, Appendix E.2):

FID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
Joint benchmark: VerseCrafter25.85301.3715.750.633.865.3534.0736.53
Joint benchmark: UniCaMo18.19192.4318.880.722.951.2316.2716.74
Extreme-200: VerseCrafter26.38411.5512.730.4710.4218.0633.5945.59
Extreme-200: UniCaMo18.43222.2014.840.586.683.5817.4619.27

UniCaMo is ahead on all eight metrics in both settings, without an additional control module.

Q1: Can UniCaMo manipulates articulated and non-rigid motion ?

Yes. We can explicitly control articulated or non-rigid object movements through independently specified 3D point trajectories, please refer anonymous website and Figure 16 for qualitative results. Given single first frame images, with no reference video, we scripted the 3D tracks by hand via our GUI for non-rigid and articulated manipulation: bending a rope, folding a T-shirt, turning a page, scrubbing a plate, sawing bread, closing a drawer and opening a refrigerator, plus a cyclist taking off into the sky.

Q2: Sensitivity to the random seed

We re-scored the original outputs and evaluated four additional seed runs on the same clips (those with a foreground mask). The second row is the mean and standard deviation of the benchmark averages over these 8 runs; the first row repeats the published result of Table 3:

Joint benchmarkFID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
Paper, Table 3 (single run)18.19192.4318.880.722.951.2316.2716.74
8 runs, mean ± std17.90 ± 0.32188.75 ± 5.4618.85 ± 0.040.72 ± 0.012.96 ± 0.061.27 ± 0.0816.18 ± 0.3416.75 ± 0.33

The results are stable across seeds: benchmark averages vary by about 3% or less on all eight metrics, and in each of the eight runs UniCaMo stays ahead of every baseline of Table 3 on all eight metrics.

Q3: The sphere under large camera motion and complex occlusion

To further stress-test our model, we built Extreme-200 for these cases (Appendix E.2). A vision-language model ranked the real clips in RealCOD-25K by the extremity of their motion. Among the top 600 clips, two authors independently reviewed them and selected the 200 with the most extreme object and camera motion.

Extreme-200 (UniCaMo / VerseCrafter)ClipsEPE↓RotErr↓CamMC↓
All2006.68 / 10.423.58 / 18.0619.27 / 45.59
Camera rotation < 45°1265.12 / 7.762.22 / 10.2015.25 / 36.68
Camera rotation ≥ 45°749.37 / 14.935.95 / 31.4525.48 / 60.76
Occlusion, then reappearance157.52 / 10.823.93 / 16.4730.60 / 44.81

Errors increase with camera rotation for both methods, but UniCaMo remains clearly ahead in all groups. For rotations beyond 45°, UniCaMo achieves a RotErr of 5.95, compared with 31.45 for VerseCrafter. The limitations are also evident, and we state them in the paper: the sphere cannot provide the content of unseen regions, which must be synthesized by the model; the noise sampled from the sphere only conveys the camera trajectory. Because Extreme-200 clips often contain both camera and object motion, we further isolated UniCaMo's camera-control capability by evaluating the 60 RealEstate10K test clips with the most extreme camera motion: the 30 clips with the largest translation and another 30 clips with the largest rotation (Appendix E.3, Table 11).

RealEstate10K, 60 most extreme clipsFID↓FVD↓RotErr↓TransErr↓CamMC↓
Gen3C24.60187.920.572.342.59
CameraNoise32.93294.668.5920.1025.52
VerseCrafter27.33252.752.169.089.88
SymphoMotion27.65233.242.007.218.34
UniCaMo19.27157.400.541.812.11

UniCaMo outperforms all previous baselines across all five metrics.

Q4: How well the proposed noise initialization aligns with the pretrained diffusion prior and does the control-quality trade-off worsen for complex motion?

According to our reply to reviewer i3S2 Q2, our structured noise is close to i.i.d Gaussian noise spatially. However, our noise is not random Gaussian noise temporally, our structured noise does contain correlation across the time channel to supply diffusion prior with motion cues. Thus, our method does not work with pretrained diffusion model that trained on i.i.d random Gaussian noise in training-free manner but required a light LoRa finetuning phase.

To quantify the severity of the trade-off under complex motion, we repeated the θ study from Table 5(a) on Extreme-200, with θ fixed for each run (Appendix E.2):

θ on Extreme-200FID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
0.018.81229.4314.890.586.433.5217.2519.08
0.218.43222.2014.840.586.683.5817.4619.27
0.518.46220.3014.420.567.854.4518.6421.17

The control-quality trade-off remains evident in this more challenging setting. Nevertheless, all settings maintain reasonable performance, with no indication of abrupt degradation or complete failure. Overall, the trade-off does not become more severe under complex motion. This experiment further confirms that if we want better motion-control then we can choose θ=0. In most of our experiment, we choose θ=0.2 to balance visual-quality and control-fidelity.

We would like to thank the reviewer again for your thoughtful review and valuable feedback. We would appreciate it if you could let us know if our responses have addressed your concerns and whether you still have any other questions about our rebuttal. We would be happy to do any follow-up discussion or address any additional comments. Thanks!

Reviewer_Y5JeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UniCaMo encodes the 3D motion of objects and the camera into the initial noise of a video diffusion model. It constructs noise with cross-frame geometric correspondences through foreground propagation, background projection, spherical noise, and depth synthesis, and utilizes a lightweight LoRA to achieve joint motion control. Experiments demonstrate that this method exhibits strong control capabilities and high video quality in object, camera, and joint control tasks.

Critical strengths
  1. The paper proposes a clear and relatively intuitive approach to joint control: representing both camera motion and object motion uniformly as 3D trajectories and further encoding them into the initial noise of a video diffusion model. This design addresses the core challenge of distinguishing between object motion and camera motion in 2D trajectories, while requiring no additional inference control branches or complex model architecture modifications.

  2. The experiments covered object motion control, camera motion control, and joint motion control. In particular, the joint control experiments align closely with the core objective of this paper, demonstrating that the method can simultaneously improve the control performance of both object trajectories and camera trajectories. Ablation experiments also show that the combined effects of foreground noise propagation, background noise propagation, spherical noise, and depth-aware synthesis collectively contribute to the performance improvement.

Critical weaknesses
  1. The paper has provided a relatively comprehensive explanation of the mechanism through 2D noise warping comparisons, layer-by-layer noise visualization, and ablation experiments; however, a more rigorous quantification of the relative contributions of key design elements is still lacking.

2.The existing results demonstrate the effectiveness of the complete 3D-grounded noise construction and provide an intuitive understanding of the roles played by foreground, background, spherical, and synthetic noise. However, the main ablation experiments still rely on a step-by-step module addition approach, making it difficult to distinguish the individual contributions of 3D geometric correspondence, cross-frame noise consistency, and spherical noise completion. Adding a set of matched noise baselines that maintain the same backbone, training data, and training strategy would clarify the paper’s conclusions regarding the underlying mechanisms.

Questions for authors
  1. The supplementary materials already show a comparison between UniCaMo and Go-with-the-Flow using 2D noise warping, as well as visualizations of foreground, background, spherical, and composite noise. Could you further report a strictly matched quantitative comparison: under the same Wan2.1 backbone, training data, LoRA configuration, and inference settings, comparing 2D noise warping, 3D propagation without spherical noise, 3D propagation with spherical noise, and full depth-aware composition?

  2. The supplementary materials illustrate the process by which users generate videos after editing 3D point tracks via the GUI. The trajectories used during the training phase are derived from automatically extracted real-world video signals, while those used during the inference phase result from user edits. Did the paper test different levels of trajectory sparsity, significant camera motion, or user-generated trajectories that differ from the motion distribution in the training videos? If so, can you report on changes in control accuracy and video quality?

3 comments

Thank you for the careful review and the concrete suggestions on the ablation. We have added the two new ablation settings to Table 4 and a 2D noise-warping baseline in the joint setting to Table 3 (the new content is highlighted in red). We have also made an anonymous website to show the requested results explicitly.

W1, W2, Q1: contribution of each component, and 2D noise warping in the joint setting

Step-by-step module addition. Instead of performing a step-by-step module addition, we now follow a "leave-one-out" approach. Settings C and D of Table 4 already remove exactly one component from the full method. We added the two missing ones as rows F and G, with foreground noise only removed and background noise only removed, respectively:

SettingFGBGSph.DAFID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
F✗✓✓✓22.09236.6316.960.674.191.2516.3116.82
G✓✗✓✓18.20193.4718.670.713.111.5535.4935.98
C✓✓✗✓18.32191.5118.150.733.161.5118.4219.11
D✓✓✓✗18.51196.2318.850.713.351.7418.8519.56
E✓✓✓✓18.19192.4318.880.722.951.2316.2716.74
  • Without Background Noise (Setting G), TransErr and CamMC metrics increase more than double (16.27 → 35.49, 16.74 → 35.98), while video quality and EPE change little, indicating that the major contribution of background noise is to camera control.
  • Without Foreground Noise (Setting F), PSNR drops from 0.72 to 0.67 and EPE rises from 2.95 to 4.19. This confirms that all of the conditioning signals for object motion go through foreground noise.

2D noise warping in the same joint setting. We added Go-with-the-Flow (Wan2.1-I2V-14B) to Table 3. To ensure fairness in the evaluation, we synthesize its conditioning signal from the first frame and the provided motion cues only. Its flow is computed on an animation of the first frame: the image is rendered along the target camera trajectories using the same depth as ours, and each object mask is moved along its annotated point tracks in the same manner as in our method.

Joint-control benchmarkFID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
Go-with-the-Flow (Wan2.1-14B)24.32266.7217.760.703.673.0820.4922.14
UniCaMo (Wan2.1-14B)18.19192.4318.880.722.951.2316.2716.74

We outperform Go-with-the-Flow on all metrics under the strictly matched setting with the same backbone, training data, and training strategy, leaving the noise-warping strategy as the only difference.

Q2: sparsity, large camera motion, user-edited trajectories

  • Sparsity. Table 5(b) reduces the number of 3D tracks from 1024 to 256 on MoveBench: FID and FVD change little and EPE rises from 1.92 to 2.38, so control degrades gradually rather than collapsing outright.
  • Large motion. We built a stress set, Extreme-200 (Appendix E.2): a vision-language model ranks the real clips of RealCOD-25K [1] dataset by how extreme their object and camera trajectories are, and among the top 600, two authors independently reviewed the clips and selected the 200 with the most extreme motion and camera trajectories.
Extreme-200FID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
SymphoMotion21.67274.4714.170.567.947.5221.3125.99
VerseCrafter26.38411.5512.730.4710.4218.0633.5945.59
UniCaMo18.43222.2014.840.586.683.5817.4619.27

UniCaMo remains the best on all eight metrics. For all methods, accuracy does drop as the camera trajectory being harder. Qualitative results are shown in Figure 15 and our newly provided webpage.

  • User-edited trajectories. The supplementary page ("In-the-Wild Control" and "UI Demo") shows hand-authored cases based on a single image, e.g., a person walking behind or in front of another. These trajectories are drawn via our GUI, which is quite different from tracks extracted from real videos, and the model follows them. We also added qualitative examples of non-rigid, articulated, and OOD motion (a man on a bicycle who then suddenly flies up into the sky). Such cases cannot exist in our real training dataset, yet UniCaMo still follows the user's intentions faithfully. Please refer to our anonymous webpage for qualitative results.

[1] Guiyu Zhang, Yabo Chen, Xunzhi Xiang, Junchao Huang, Zhongyu Wang, and Li Jiang. Symphomotion: Joint control of camera motion and object dynamics for coherent video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11127–11137, 2026

We would like to thank the reviewer again for your thoughtful review and valuable feedback. We would appreciate it if you could let us know if our responses have addressed your concerns and whether you still have any other questions about our rebuttal. We would be happy to do any follow-up discussion or address any additional comments. Thanks!

Reviewer_Y5JeOpenReview ↗

Official Review of Submission28367 by Reviewer Y5Je

Thank you for the detailed rebuttal and the additional experiments. The leave-one-out ablation (rows F and G) clearly isolates the contribution of foreground and background noise. The strictly matched Go-with-the-Flow comparison in the joint setting addresses my main concern about whether the gains come from the noise construction rather than the backbone or training. The sparsity and large-motion results, the eight seeds variance study, and the θ study on the harder set are also helpful. I have read the updated paper and the other reviewers’ discussions, and I think the revisions have strengthened the work. I will keep my score.

Comments on the paper

5 comments

Thank you for the careful review. We compared our method against Go-with-the-flow and VerseCrafter in joint-control setting and conducted a stress test on extreme cases (object rotation, substantial disocclusion, reappearance and crossing order, motion along depth). We also clarify below how the noise handles newly exposed regions and made an anonymous website to show the requested results explicitly. All additions to the paper are in red.

W1, W2: The sphere is fixed in the world; what do newly exposed regions receive? How does UniCaMo perform under extreme cases?

We thank the reviewer for acknowledging that our sphere noise is an useful prior. We further clarify our contribution as follows :

  • Background uncovered by an object or by the camera has no point in the reference cloud, so it is filled from the sphere. The sphere gives it noise that remains consistent as the camera moves. While it does not recover the true geometry of the unseen region, it indicates the camera's direction of motion. The geometry of the novel-view regions is handled by the diffusion prior and the text prompt.
  • New surfaces of a moving object. While newly exposed surfaces of moving or rotating objects do not receive object-attached noise when reference-derived propagation is unavailable, these regions are filled by spherical noise. We want to clarify that the tracks that start on the visible surface of the object still stay visible in our structured noise and they have provided the model motion cues. Therefore, our structured noise is imperfect by design and we performed LoRa finetuning on existing video diffusion model to produce plausible outputs.
  • Occlusion and reappearance. We employed Depth-Aware composition to tackle this issue, For each latent location receiving multiple contributions from the foreground trajectories or background point cloud, we apply a z-buffer rule and retain the contribution with the smallest camera-space depth, i.e., the point closest to the camera. However, suppressed tracks are not discarded, once they passed the depth test, the same noise associate with that track will be used. Thus, our method can preserve the object's fidelity in such cases.

On our newly provided anonymous website, we provide various examples of extreme camera motion, 3D rotation, occlusion, and reappearance to demonstrate UniCaMo's capabilities in such cases.

Stress test. We built Extreme-200 for these cases (Appendix E.2): a vision-language model ranks the real clips of RealCOD-25K dataset by how extreme their object and camera trajectories are, and among the top 600, two authors independently reviewed the clips and selected the 200 with the most extreme motion and camera trajectories.

Extreme-200FID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
SymphoMotion21.67274.4714.170.567.947.5221.3125.99
VerseCrafter26.38411.5512.730.4710.4218.0633.5945.59
UniCaMo18.43222.2014.840.586.683.5817.4619.27

UniCaMo performs best on all eight metrics. We further explicitly compare our method with VerseCrafter across all requested events (Table 9) (object rotation, crossing order, motion along depth, etc.). The advantage holds for every event you listed. In the following table, each result is reported as UniCaMo / VerseCrafter. Please also refer to the updated Tables 1, 2, and 3 for comparisons with VerseCrafter on standard datasets.

EventClipsPSNR↑EPE↓RotErr↓CamMC↓
Object rotation revealing new surfaces10015.34 / 12.875.65 / 9.332.74 / 16.7316.71 / 43.50
Crossing with a depth order6714.55 / 12.586.54 / 10.223.31 / 18.1818.89 / 44.30
Motion along depth9814.38 / 12.527.83 / 12.004.68 / 21.9221.55 / 50.12
Background disocclusion3014.97 / 12.777.45 / 10.636.42 / 26.0716.69 / 50.29
Occlusion, then reappearance1513.76 / 12.187.52 / 10.823.93 / 16.4730.60 / 44.81

Large camera motion. The Extreme-200 dataset we created contains not only camera motion but also object motion. To isolate UniCaMo's camera-control performance, we also evaluated the 60 RealEstate10K test clips with the most extreme camera motion, where the sphere must fill large newly revealed regions (Appendix E.3, Table 11).

MethodFID↓FVD↓RotErr↓TransErr↓CamMC↓
Gen3C (Ren et al., 2025)24.60187.920.572.342.59
CameraNoise (Zhao et al., 2026)32.93294.668.5920.1025.52
VerseCrafter (Zheng et al., 2026)27.33252.752.169.089.88
SymphoMotion (Zhang et al., 2026)27.65233.242.007.218.34
UniCaMo (Ours)19.27157.400.501.912.20

Q1: VerseCrafter and Go-with-the-Flow in the joint setting

Both are now included in Table 3 (red rows). For Go-with-the-Flow, we followed your suggestion and derived its guidance from the same reference depth and camera poses: we render the first frame along the target camera trajectories, move each object mask along its annotated 2D track, and compute the flow from this animation. No subsequent frames of the target video are used, either as appearance information or as flow input.

Joint-control benchmarkFID↓FVD↓PSNR↑SSIM↑EPE↓RotErr↓TransErr↓CamMC↓
VerseCrafter25.85301.3715.750.633.865.3534.0736.53
Go-with-the-Flow (Wan2.1-14B)24.32266.7217.760.703.673.0820.4922.14
UniCaMo18.19192.4318.880.722.951.2316.2716.74

UniCaMo outperforms both methods on all eight metrics.

Q2: Within-frame statistics, blending rule, and temporal mapping

Statistics (Appendix E.1, Table 7). To validate the preservation of spatial i.i.d. Gaussianity, we follow the evaluation protocol outlined by Go-with-the-Flow. We use Moran’s I to measure the spatial correlation of the warped noise and the Kolmogorov-Smirnov (K-S) test to assess its normality. This experiment is conducted on 200 videos, yielding 67200 latent maps in total.

NoiseMoran's IMean p-valueK-SMean p-value
Random Gaussian0.00010.500.01100.50
Bilinear warp0.15760.030.03130.09
Bicubic warp0.12510.100.06040.08
Nearest warp0.15250.090.07240.05
Go-with-the-Flow0.00020.510.01070.52
Ours0.00340.480.01160.48

We make the following observations:

  1. A Moran’s I value close to 0 indicates little to no spatial autocorrelation, as expected for i.i.d. noise. Our method achieves a Moran’s I value of 0.0034 with an average p-value of 0.48, providing strong evidence that the warped noise has no significant spatial autocorrelation.
  2. The K-S test measures the difference between the empirical distribution of the warped noise and a standard normal distribution. A small K-S statistic combined with a high p-value suggests that the two distributions are similar. Our method achieves a K-S statistic of 0.0116 with an average p-value of 0.48, indicating that the warped noise remains consistent with a normal distribution.

Blending. We use the variance-preserving rule from Go-with-the-Flow, z = ((1 − θ)·z_motion + θ·ε) / √((1 − θ)² + θ²), where ε is freshly sampled Gaussian noise, z is the blended noise, z_motion is the clean structured noise. One value of θ is drawn per video: from U[0, 0.5] during training and around the default value of θ = 0.2 during inference.

Temporal mapping. The Wan 2.1 VAE encodes an 81-frame video into 21 latent frames. Given camera poses and point tracks for all 81 input frames (indexed 0–80), we align latent frame k with the camera pose and track positions at pixel-space frame 4k, where k=0,…,20

AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful and constructive feedback. We greatly appreciate their recognition of UniCaMo's ability to jointly control camera and object motion without requiring external control adapters, as well as the extensive experimental validation, comprehensive ablation studies, and generally clear presentation. Several reviewers also highlighted the strong empirical performance of our method and the simplicity and effectiveness of our 3D-grounded noise-warping framework.

In response to the reviewers' comments and questions, we provide detailed clarifications, additional evaluations, and new ablation studies. Furthermore, we manually curated and evaluated UniCaMo on a newly introduced benchmark, Extreme-200 (more details can be found in our updated pdf, highlighted in red), which consists exclusively of videos with highly challenging object and camera trajectories. We also provide additional qualitative results and visualizations on our anonymous website at https://unicamo-x.github.io/unicamo/.

We hope these additional analyses, experiments, and clarifications adequately address the reviewers' concerns. We sincerely welcome any further questions or suggestions and thank the reviewers again for their time, careful assessment, and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.