acceptodds
Withdrawn from ICLR 2027

4D-GSW: 4D Gaussian Splatting Watermarking with Spatio-Temporal Consistent Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

While 4D Gaussian Splatting (4DGS) has revolutionized high-fidelity dynamic reconstruction, safeguarding the intellectual property of these assets remains an open challenge. Conventional steganographic techniques often neglect the underlying kinematic manifolds, triggering non-physical artifacts such as severe temporal flickering and "FVD collapse". To address this, we propose 4D-GSW, a kinematic-aware watermarking framework designed to embed robust copyright information while preserving high spatio-temporal consistency. Unlike prior 4D steganography that primarily focuses on opacity-guided invisibility, our approach explicitly addresses the physical coherence of motion trajectories. We introduce a Spatio-Temporal Curvature (STC) metric to identify "Dynamic Instants," adaptively gating watermark gradient injection to shield critical motion manifolds from non-physical perturbations. To ensure global coherence across complex deformations, we formulate a joint HMM-MRF energy minimization model that synchronizes watermark phases within both temporal trajectories and spatial neighborhoods. Furthermore, an anisotropic gradient routing mechanism ensures that watermark embedding reduces interference with photometric reconstruction. Extensive experiments have demonstrated the superior performance of our method in robustly hiding watermarks while resisting various attacks and maintaining high rendering quality and spatiotemporal consistency.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Social Good, Watermarking, Digital Reconstruction

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_QhDHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

4D-GSW proposes a kinematics-aware watermarking framework designed to preserve the physical coherence and spatiotemporal consistency of motion trajectories. The method introduces a spatiotemporal consistency (STC) gating strategy to suppress nonphysical perturbations, an HMM–MRF energy-minimization model to synchronize watermark phases.

Critical strengths
  • Kinematics-aware STC gating: Incorporating trajectory velocity into the STC mechanism is an interesting approach to maintaining physically coherent motion during watermark embedding.
  • Optimal transport for spatiotemporal consistency: The use of optimal transport (OT) to promote consistency across space and time is a promising component of the framework.
Critical weaknesses
  • Insufficient baseline and dataset comparisons: The experimental evaluation does not sufficiently establish the method’s advantages and limitations relative to relevant competing approaches. Additional comparisons with recent methods and evaluations on widely used datasets would strengthen the paper.
  • Limited ablation studies: The existing ablations do not clearly isolate the contributions of STC gating and OT. More targeted experiments are needed to substantiate the effectiveness of the main proposed components.
Questions for authors

Major:

  • Coverage of related work and baselines (line 118): Additional papers on 4D watermarking and steganography have been published or made available online in 2026, including Mark4D (CVPR 2026) and Guard4D (IJCAI 2026). The authors should consider discussing these methods as related work and including them as experimental baselines where their settings and implementations permit a fair comparison.

  • Dataset coverage: The authors evaluate their method on the Consistent4D dataset, whereas other methods use DyNeRF and D-NeRF. Please consider including these datasets to assess the method’s generalization across different scene types and motion characteristics. If these evaluations are not feasible, please explain why.

  • Alignment between the ablations and the proposed objective: The ablations in Table 4 (removing the temporal, spatial, or STC components) are difficult to map to the joint objective in Section 4, Equation (13). Please explicitly define which terms or mechanisms are removed in each variant. In particular, the following ablations would help clarify the main contributions:

    • Removing STC gating: the watermark-embedding confidence weight w from L_conn (keep L_conn, remove w)
    • Removing OT (Remove L_conn).

These experiments would give readers a clearer understanding of how each component affects watermark performance, reconstruction quality, and spatiotemporal consistency.

  • Message-length ablation: Please evaluate different message lengths, such as 16, 32, and 48 bits, to characterize the trade-off between payload capacity, watermark recovery, and reconstruction quality.

Minor:

  • Citation formatting in Section 1: Several citations appear to be rendered incorrectly, making the text difficult to read. Please check the bibliography and citation formatting, e.g. Missing spaces:

    • Page 1, line 037: a space is missing between “scenes” and “Hu & Han.”
    • Page 3, line 132: a space is missing in “dynamicsLuiten.”
    • Page 4, line 188: a space is missing in “decoderZhu et al.”
  • Table 1 presentation: Please distinguish more clearly between 4DGS methods and results with and without watermark or steganographic embedding, for example by using separate rows or groups.

  • Notation in Figure 2: The formulations is unclear or potentially incorrect. Please explain the notation {k_il ,k_iw } for all K and W.

  • Embedded message length: Section 5, lines 336–350, does not clearly specify how many bits are embedded (48bit is memtioned somewhere later). Please state the payload length explicitly and clarify whether it remains fixed across experiments.

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

The paper proposes a watermarking method for 4D Gaussian splatting. The main idea is to hide copyright information inside moving 3D objects without affecting the visual quality of the model. Firstly, the proposed method uses a Spatio Temporal Curvature (STC) mechanism to identify regions with complex motion. This method applies stronger watermark updates in stable regions and reduces them in high curvature regions. It also uses Optimal Transport (OT), Hidden Markov Model (HMM), and Gaussian Markov Random Field (GMRF) to maintain spatial and temporal consistency. The experimental results show that 4D-GSW achieves better watermark recovery. It also preserves motion quality compared to other watermaking baselines.

Critical strengths
  1. The paper introduces a novel watermarking technique in 4DGS model using curvature based strategy. Instead of embedding watermarks uniformly through all the frames, the method focuses on stable regions and reduces watermark in regions with high curvature.

  2. The paper evaluates the proposed method under different distortions, such as noise, rotation, scaling, cropping, and JPEG compression. The method achieves 99.10% watermark bit accuracy without distortion and 98.05% average accuracy with distortions.

  3. The training overhead is moderate and the generated watermarked asset introduces no additional rendering-time cost.

Critical weaknesses
  1. The proposed method avoids putting watermarks into parts of the animation where movement is complicated and there is a strong curvature. But the paper doesn't discuss other motion parameters like speed, acceleration. For example, an object is moving rapidly in a simple straight path having zero curvature. On the other hand, the object is moving slowly in a high curvature path. Relying solely on curvature might not capture other motion metrics.

  2. In Table 1, the original SC4D backbone achieves better results than the proposed 4D-GSW method in terms of all metrics which is expected. But the paper does not investigate how this quality gap changes under different perturbations.

  3. The paper mentions HMM(Hidden Markov Model) but the mathematical formula is a penalty that is consistent between Gaussian states. There is no definition of the hidden states, observations, probabilities.

  4. The experiment is evaluated on only seven animated scenes and one main 4DGS backbone. For more generalizability the method should be evaluated on other real scene dataset, indoor and outdoor scene where many objects are present in the scene.

  5. The introduction contains too many technical terms and concepts without providing sufficient definitions or explanations. For example, terms such as "Spatio-Temporal Curvature", "Dynamic Instants", "HMM-MRF", "anisotropic gradient routing", and "FVD collapse" are introduced without clearly explaining.

Questions for authors
  1. Why is curvature better than speed, acceleration or other motion measurements?
  2. How much additional degradation does watermarking introduce when both original and watermarked scenes have the same perturbations?
  3. Can the paper evaluate watermark extraction accuracy from many unseen camera angles?
Reviewer_rUhrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a watermarking method for dynamic 4D Gaussian Splatting. It aims to reduce temporal flickering by combining spatiotemporal consistency (STC), optimal transport (OT) alignment, and a motion regularization term. The watermark is embedded into Gaussian trajectories without requiring an additional rendering network. Experiments on Consistent4D evaluate watermark recovery, rendering fidelity, and robustness under several attacks.

Critical strengths
  1. Interesting and practical problem setting.

Temporal flickering is an important issue in 4DGS watermarking, and the idea of adjusting watermark embedding based on Gaussian motion is intuitive. The method also works directly on Gaussian trajectories without adding rendering overhead, which is a practical advantage.

  1. Broad robustness evaluation.

The paper considers not only 2D image distortions but also model-level attacks such as pruning, control-point noise, and temporal deformation noise. The per-sample results in Tables 6 and 7 and the trajectory noise analysis in Table 8 are also useful.

Critical weaknesses
  1. Several key components are not well justified.

The STC formulation in Eq. 5 seems problematic. The curvature term becomes zero for changes in speed along a straight trajectory, so it does not fully capture motion changes as claimed. It may also become unstable for nearly stationary Gaussians. The normalization and clamping steps are not clearly explained.

I also question whether OT alignment is necessary, since Gaussian identities are already preserved over time in SC4D. In addition, the Lcon term in Eq. 9 directly penalizes temporal changes in Gaussian positions, which seems inconsistent with the claim that it does not restrict motion.

There are also several unclear mathematical details. The HMM inference appears to be a combination of temporal L2 and kNN smoothness terms rather than MAP inference. The claimed anisotropic formulation uses a scalar weight, and the energy in Eq. 16 does not seem consistent with the earlier equations.


  1. The evaluation does not fully support the claims.

The experiments are limited to seven objects from Consistent4D, using a single backbone. This seems narrow given the paper's broader motivation involving dynamic scenes and telepresence.

More importantly, the paper mainly relies on PSNR, SSIM, and FVD to evaluate temporal quality. There is no direct flicker metric, such as temporal LPIPS or warping error. The fidelity comparison is also mainly against ground truth, while the degradation relative to clean SC4D is important for watermarking. For example, LPIPS increases from 0.080 to 0.133.

The baseline comparisons also need clarification. Table 1 mixes watermarking methods with general 4D generation methods, and the HiM reproduction achieves only 89.5% clean bit accuracy. I also noticed that some ablation results seem inconsistent with the main tables, particularly the FVD and accuracy values in Tables 4 and 6.


  1. The novelty claims seem overstated.

The paper describes Hide-in-Motion as “the only existing 4D steganography approach” and claims to be “the first work to explicitly mitigate ‘FVD collapse’, but it does not discuss Mark4D (CVPR 2026), which also studies temporally consistent 4DGS watermarking. A clearer comparison between Mark4D's motion-adaptive weighting and the proposed STC gating would strengthen the novelty claim.


There are also some presentation issues, including incorrect section and equation references, undefined notation, and minor typos.

Questions for authors

Q1. How does STC handle nearly stationary Gaussians, where the velocity magnitude approaches zero? Could the authors show the distribution of the STC weights and a sensitivity analysis of τ? Would acceleration magnitude be a more appropriate measure than curvature for capturing motion changes?

Q2. Since Gaussian identities are preserved over time in SC4D, why is OT alignment needed? Could the authors isolate its contribution without also removing the HMM prior? Also, how much does Lcon actually restrict Gaussian motion?

Q3. Could the authors clarify the apparent inconsistencies between Table 4 and Tables 2 and 6? For example, the FVD of “w/o spatial” (1688.47) matches the Crocodile FVD in Table 6, and its accuracy (96.84) matches the noise accuracy in Table 2. How exactly are SC4D+HiDDeN and the reproduced HiM baseline implemented and validated?

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

Related papers

Discussion (0)

Sign in to comment.