acceptodds
Under review as a conference paper at ICLR 2027

Velocity 4D Gaussian Splatting: Velocity is All You Need

Abstract

Novel view synthesis for dynamic scenes remains a significant challenge in computer vision, primarily due to the complexity introduced by the temporal dimension and diverse motion patterns. The additional dimensionality in 4D Gaussians complicates their constraint, resulting in motion inconsistencies and artifacts. Methods that model time independently struggle to decouple temporal and spatial dimensions effectively, while time-slicing techniques fall short in capturing temporal consistency. To address motion inconsistencies, we introduce velocity 4D Gaussian Splatting (V4DGS), a dynamic reconstruction approach that incorporates velocity consistency regularization into 4D Gaussians, leveraging temporal constraints alongside spatial priors. By using velocity to quantify time dimension, we learn transferable temporal features that enable accurate motion prediction at new timestamps. Our velocity consistency regularization framework includes both temporal and spatial components, effectively addressing motion inconsistency by constraining velocity variations within an infinitesimal time interval. V4DGS excels in dynamic scenes with multiple objects and even reconstructing reflective regions. Through extensive evaluations on both monocular and multi-view motion datasets, V4DGS achieves real-time rendering at high resolutions, surpassing existing methods in both quantitative and qualitative metrics. V4DGS substantially improves quality (PSNR gain of 3.31 on the Plenoptic dataset), maintaining comparable rendering speeds and training time to previous methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Novel View Synthesis, Dynamic Scene, Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ASHkOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Velocity 4D Gaussian Splatting (V4DGS) for dynamic scene reconstruction and novel view synthesis. The method builds upon a HexPlane-based Gaussian deformation framework and introduces temporal and spatial velocity consistency regularization to improve motion coherence. Experiments on the Plenoptic Video and D-NeRF datasets demonstrate improvements in rendering quality while maintaining real-time performance.

Critical strengths

The paper addresses an important problem in dynamic scene reconstruction, where motion inconsistencies can lead to temporal artifacts and degraded rendering quality. The proposed method is relatively simple and can be integrated into an existing Gaussian deformation framework without substantially increasing inference complexity. The quantitative results are promising, particularly the 3.31 dB PSNR improvement over Deformable4DGS on the Plenoptic Video dataset.

Critical weaknesses
  1. The mathematical formulation of spatial velocity consistency is questionable (Eq. 13). The authors claim that multiplying the velocity difference by a factor of 2 extends spatial consistency constraints from a single timestamp to an entire temporal interval. However, Eq. 13 is mathematically equivalent to standard neighborhood velocity smoothing multiplied by 2. This constant factor can be directly absorbed into the regularization weight without introducing any additional temporal constraint. The authors should clarify this mathematical justification and compare their formulation against standard velocity smoothing with matched regularization weights.

  2. Limited methodological novelty. The overall framework closely follows existing deformation-based Gaussian Splatting methods. The main modifications are temporal velocity smoothness and local spatial velocity alignment, both of which are relatively standard motion regularization strategies.

  3. Comparisons with recent methods are missing. The quantitative comparisons in Tables 1 and 2 primarily involve methods published in or before 2024.

  4. Insufficient ablation studies to isolate the contributions of the proposed regularization terms. Table 3 only reports the baseline, spatial-only, and full configurations, without evaluating temporal regularization independently. Moreover, spatial regularization alone provides only marginal improvements on average and even degrades performance on two of the five reported scenes.

Questions for authors

Please refer to the weakness.

1 comment
AuthorsOpenReview ↗

V4DGS Exhibits Ample, Comprehensive, and Consistent Experiments

1. We would like to clarify that a detailed derivation has already been provided on p. 7. The proof follows a clear and logically consistent sequence of steps and is presented in a mathematically rigorous form.

2. Our contributions include an optimization method for 4D Gaussian properties that encourages neighboring Gaussians to move consistently, aiming to maintain that consistency over continuous time intervals. Furthermore, we present two theorems and give their proofs. Theorem 1 shows that, for a differentiable trajectory, finite-difference velocity converges to instantaneous velocity as the interval shrinks. Theorem 2 formulates the neighboring Gaussians penalty.

3. The Introduction (p. 2) analyzes how representative dynamic rendering methods model motion: Deformable4DGS uses a deformation field, RealTime4DGS uses native 4D Gaussian primitives, and 4D-RotorGS renders time slices of 4D Gaussians. We then present Theorem 1, which motivates our use of a local constant-velocity approximation over short time intervals. Tables 1 and 2 (p. 8) compare V4DGS with these methods on both Plenoptic Video and D-NeRF. The evaluation, therefore, includes a comprehensive comparison.

4. Table 3 gives the most direct experimental evidence supporting the claims. Holding the spatial term fixed, adding the temporal term increases PSNR in all five reported D-NeRF scenes by 0.16–0.52 dB, with a mean gain of 0.29 dB. Tables 1–2 and Figures 5–6 place this ablation alongside benchmark results and rendered examples. The supplement adds qualitative comparisons on both datasets (Figures A and C) and visualizes estimated optical flow (Figure D). These results support the benefit of our contributions.

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

This paper proposes V4DGS, which adds temporal and spatial velocity-consistency regularization to a deformation-based Gaussian splatting framework. It reports improved PSNR score on Plenoptic Video and on D-NeRF over its stated deformation baseline, while maintaining the same reported training times and rendering speeds.

Critical strengths
  1. Improving motion consistency in dynamic Gaussian reconstruction is a relevant problem, and motion regularization is a reasonable approach.

  2. The reported PSNR improvements on real multiview and synthetic monocular datasets are meaningful.

Critical weaknesses
  1. Theorem 1 establishes the standard convergence of a finite-difference approximation to instantaneous velocity. It does not justify dropping delta_t in Eq. (3) or equating velocity with displacement in Proposition 1, Eq. (10). Defining delta_t=t_2-t_1 does not imply delta_t=1. Moreover, the deformation output is previously described as a canonical-space offset; even with a unit time step, velocity generally requires the difference between offsets at two timestamps. These ambiguities affect what the proposed losses actually regularize.

  2. Equation (13) is exactly twice an ordinary neighborhood velocity-consistency penalty after factoring out the constant two. The factor can be absorbed into the regularization weight. It does not introduce another timestamp or enforce consistency at both interval endpoints. Consequently, the claimed temporal extension and special justification for the factor two are unsupported.

  3. The temporal loss and essential implementation details are unclear. Equation (12) averages the predicted velocity change delta_v_i without specifying a norm or another nonnegative scalar reduction. The paper also does not explain how this prediction is constrained to equal the actual velocity change along the rendered trajectory. The paper leaves several implementation details unclear: how the time interval and neighboring Gaussians are selected, how many neighbors are used, and how the regularization losses are weighted and combined with the reconstruction loss. Without these details, it is difficult to evaluate or reproduce the method.

  4. The Lego comparison in Figure 6 appears misaligned with the ground truth and other baseline, particularly in the enlarged bucket region. Please clarify whether all methods use identical camera parameters, timestamps, and crop regions. This discrepancy makes the claimed improvement difficult to assess.

  5. In the introduction, the Theorem 1 paragraph, it ends with " See the proof in Sec.. ". This kind of unfinished sentence or mistake should not appear in the submitted version.

  6. The distinction from closely related work is insufficiently explained. Although the paper cites and discusses 4D-RotorGS, it does not address that method’s neighborhood velocity-consistency loss, whose algebraic form matches Eq. (13) up to a constant factor. The authors should clarify their "first" claim, what is new in their spatial regularizer, how the velocity definitions and neighborhoods differ, and what additional benefit the proposed temporal term provides.

Questions for authors

How are delta_x, v, and delta_t defined in the implementation? Is delta_x a canonical-space offset or an inter-timestamp displacement, and what justifies the equalities in Eq. (3) and (10)?

How does Eq. (13) enforce consistency at both interval endpoints? Since the factor two can be absorbed into the loss weight, is an additional temporal operation missing from the formulation?

What is the exact implemented temporal loss in Eq. (12), and how is delta_v_i tied to the rendered trajectory? Please also specify the time-interval selection, neighborhood construction, K, and full training objective with loss weights.

Do all Lego results in Figure 6 use identical camera parameters, timestamps, and crop regions? What accounts for the apparent misalignment in the enlarged bucket region?

What distinguishes the proposed regularization from 4D-RotorGS’s neighborhood velocity-consistency loss, and what supports the “first” claim? Can you provide a temporal-only ablation to clarify the spatial term’s additional benefit?

1 comment
AuthorsOpenReview ↗

V4DGS Boosts Rigorous Mathematical Foundations

1. Proof of Theorem 1 and clarification of Eq. (10)

Page 6 provides the step-by-step proof of Theorem 1 in Eqs. (4)–(9). For a smooth trajectory xi(t)x_i(t), we expand xi(t+Δt)x_i(t+\Delta t) about tt, subtract xi(t)x_i(t), and divide by Δt\Delta t. Taking Δt→0\Delta t\to 0 proves convergence to the instantaneous velocity. We stand by this derivation. It retains Δt\Delta t throughout and does not require setting Δt=1\Delta t=1.

The final equality in Eq. (10) is intended to indicate that, in the limit as Δt→0\Delta t\to 0, velocity is identified with displacement under the formulation adopted in this derivation. We have explicitly added this clarification to the main text.

2. Clarification of the detailed derivation on p. 7

We would like to clarify that a detailed derivation has already been provided on p. 7. The proof follows a clear and logically consistent sequence of steps and is presented in a mathematically rigorous form.

3. Domain-level temporal consistency and hyperparameter selection

Equation (12) measures the average rate of velocity change over the domain and characterizes domain-level temporal consistency, rather than consistency along individual rendering trajectories. The selection of the principal hyperparameters is described in detail in the supplementary material. The remaining hyperparameters were determined through grid search under the corresponding experimental settings.

4. Novel-view generalization and reconstruction quality on the LEGO dataset

Unlike the other datasets, the training and test sets of the LEGO dataset are captured from different viewpoints. This setting is specifically designed to evaluate the model’s generalization capability to novel views. In the LEGO reconstruction task, our method achieves stronger reconstruction accuracy than the compared baselines, especially in recovering the bucket’s geometric structure, surface texture, and pose. Although some fine-grained discrepancies remain, our method produces the highest perceptual quality when rendering from novel viewpoints.

5. Typos

Due to formatting changes introduced for ICLR 2026, minor display errors occurred. We have carefully reviewed the manuscript and corrected all issues.

Reviewer_9xkdOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies motion inconsistency in dynamic 4D Gaussian Splatting, aiming to improve temporal coherence and rendering quality while maintaining efficiency. The authors propose V4DGS, which incorporates velocity consistency regularization into a deformation-based framework to constrain motion variations across both temporal intervals and neighboring Gaussians. By leveraging velocity as an explicit motion attribute, the method promotes smoother and more consistent dynamic scene representations.

Critical strengths
  1. The paper clearly identifies temporal inconsistency as a key challenge in dynamic 4D Gaussian Splatting and provides a reasonable motivation for velocity consistency regularization.

  2. This paper is generally well-written, with a coherent presentation of the proposed method and experimental results.

Critical weaknesses
  1. Although the paper evaluates V4DGS on the Plenoptic Video and D-NeRF datasets, additional experiments on real-world monocular datasets, such as NeRF-DS or HyperNeRF, would help validate its generalizability to more challenging motion patterns and appearance variations.

  2. The evaluation could be strengthened by including more 4DGS methods that explicitly decouple spatial and temporal representations, rather than focusing primarily on deformation-based approaches. Such comparisons would help clarify the advantages of the proposed velocity consistency regularization over alternative strategies for modeling dynamic Gaussian motion.

  3. The proposed method introduces velocity consistency constraints on top of a deformation field, but its relationship to existing motion-sharing and anchor-based representations could be discussed more thoroughly. A clearer comparison would help establish the novelty of the velocity-based formulation and distinguish its contribution from existing spatial and temporal motion constraints.

Questions for authors
  1. How well does V4DGS generalize to real-world monocular scenes? Evaluation on datasets such as NeRF-DS or HyperNeRF would provide further evidence.

  2. A comparison with additional 4DGS methods that explicitly decouple spatial and temporal representations would help clarify the advantages of the proposed approach.

  3. To what extent does V4DGS improve upon existing anchor-based or motion-sharing representations? A direct experimental comparison would help establish its novelty.

Convincing answers would strengthen my assessment and potentially increase my rating.

2 comments
AuthorsOpenReview ↗

What Is Temporal Consistency in Dynamic Scene Rendering?

1. Thank you for the suggestion. We would like to clarify that our evaluation already includes challenging real-world scenes through the Plenoptic Video dataset. As described on p. 8 and in supplementary Appendix B, this benchmark contains six real-world scenes with complex phenomena, including the sudden appearance of flames, moving shadows, and translucent and reflective materials. These introduce substantial challenges in modeling both dynamic geometry and time-varying appearance.

Our choice of Plenoptic Video therefore provides a demanding evaluation of precisely the motion and appearance variations highlighted in the review. We evaluate these scenes at 1352 × 1014 resolution, testing reconstruction quality under realistic visual complexity. D-NeRF complements this evaluation with synthetic monocular sequences, where each timestamp is observed from a single viewpoint. Our datasets assess performance under real-world scene complexity and monocular observation constraints in complementary settings.

2. The Introduction (p. 2) analyzes how representative dynamic rendering methods model motion: Deformable4DGS uses a deformation field, RealTime4DGS uses native 4D Gaussian primitives, and 4D-RotorGS renders time slices of 4D Gaussians. We then present Theorem 1, which motivates our use of a local constant-velocity approximation over short time intervals. Tables 1 and 2 (p. 8) compare V4DGS with these methods on both Plenoptic Video and D-NeRF. The evaluation therefore includes a comparison beyond deformation-based methods.

3. Our contributions include an optimization method for 4D Gaussian properties that encourages neighboring Gaussians to move consistently, aiming to maintain that consistency over continuous time intervals. Furthermore, we present two theorems and give their proofs. Theorem 1 shows that, for a differentiable trajectory, finite-difference velocity converges to instantaneous velocity as the interval shrinks. Theorem 2 formulates the neighboring Gaussians penalty.

The most direct experimental evidence is Table 3. Holding the spatial term fixed, adding the temporal term increases PSNR in all five reported D-NeRF scenes by 0.16–0.52 dB, with a mean gain of 0.29 dB. Tables 1–2 and Figures 5–6 place this ablation alongside benchmark results and rendered examples. The supplement adds qualitative comparisons on both datasets (Figures A and C) and visualizes estimated optical flow (Figure D). These results support the benefit of our contributions.

The authors have adequately addressed my concerns, and I will raise my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.