acceptodds
Withdrawn from ICLR 2027

MPP-GS: Multi-Projection Packing for High-Fidelity 3DGS Watermarking

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Watermarking 3D Gaussian Splatting (3DGS) assets requires modifying selected Gaussian primitives. When watermark carriers are concentrated in a few spatial regions, those regions must carry more of the embedding burden, potentially requiring stronger local perturbations that degrade rendering fidelity. However, existing methods often lack explicit mechanisms for spatially balanced carrier allocation. In this paper, we introduce MPP-GS, a high-fidelity 3DGS watermarking framework with two stages. In carrier construction, Multi-Projection Packing (MPP) organizes Gaussians on a UV grid and assigns each Gaussian multiple candidate destinations, reducing congestion and overflow while producing a more balanced carrier. In watermark embedding, Complementary Pair Modulation (CPM) exploits this carrier to distribute parent–auxiliary pairs and coordinate their updates, reducing rendering interference while preserving decodability. Across three benchmarks (Blender, LLFF, and Mip-NeRF 360), MPP reduces average carrier overflow from 17.75% to 0.08%, retaining substantially more Gaussian primitives. Together, balanced carrier construction and low-interference watermark embedding enable MPP-GS to achieve higher rendering fidelity with comparable or better decoding accuracy across multiple payloads. At 64 bits, MPP-GS achieves 35.15 dB PSNR and 95.95% bit accuracy, compared with 32.79 dB and 91.17% for W2M. Code will be made publicly available on GitHub.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
3D Gaussian Splatting, Digital Watermarking, Privacy

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_5NQCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MPP-GS, a watermarking method for 3D Gaussian Splatting. It uses Multi-Projection Packing (MPP) to distribute Gaussians more evenly across a UV grid and reduce overflow. Complementary Pair Modulation (CPM) then embeds watermarks using parent–auxiliary Gaussian pairs while trying to minimize rendering distortion. Experiments on three datasets show improved rendering quality and bit accuracy compared to existing methods.

Critical strengths
  1. Clear evaluation of carrier construction. The paper compares SPP and MPP under the same carrier budget, showing a large reduction in overflow (32.0% to 0.14% on LLFF) and better coverage across all 23 scenes (Fig. 4b). Table 4 also provides a fair comparison of allocation strategies under the same carrier and pair budget.

  2. Simple design with consistent results. The pair parameterization (Eq. 9) preserves the original opacity under alpha composition, and the resulting model remains compatible with standard 3DGS renderers. The method also shows consistent improvements in bit accuracy and rendering fidelity across three datasets and multiple payload sizes.

Critical weaknesses
  1. The main improvement may come from fixing a problem introduced by UV packing. SPP already loses many Gaussians before watermarking (32% overflow on LLFF), resulting in poor reconstruction quality. Thus, much of the PSNR gain over SPP+CPM may simply come from retaining more Gaussians rather than better watermark allocation. I think an important baseline is missing here: applying balanced spatial quotas directly to native 3DGS using voxel or kNN partitions, without UV packing. Without this comparison, it is hard to tell whether MPP is really necessary.

  2. The pair structure may be vulnerable to adaptive attacks. Parent and auxiliary Gaussians share the same center and rotation, making them potentially easy to identify. An attacker could merge these pairs or selectively remove duplicated Gaussians. Since the opacity and DC updates are designed to cancel, and the differential SH component also cancels under weighted averaging (Eq. 13), such attacks may significantly weaken the watermark. The current robustness experiments mainly consider random model modifications and 2D distortions, but do not test attacks targeting this specific structure.

  3. Fairness and computational overhead need more discussion. It is unclear whether all baselines were evaluated under the same protocol. For example, W2M uses Scaffold-GS, and the methods use different watermark decoders. The paper also reports results from three seeds without standard deviations. Moreover, the method adds around 45–58% more Gaussians, increasing rendering time by roughly 1.5–1.8×. It is unclear how much of the fidelity improvement comes from the additional primitives. Nullspace modulation also relies on a limited set of representative directions rather than actual rendering views.

  4. It would be useful to briefly discuss whether MPP and CPM could extend to dynamic 4D Gaussians. Since MPP relies on Gaussian centers for UV assignment, it is unclear whether packing should be done in canonical space and whether the balanced allocation would still hold after deformation. A short discussion of this limitation would be sufficient.

Questions for authors
  1. How robust is the watermark to attacks that merge parent–auxiliary pairs or selectively remove Gaussians with duplicated centers?

  2. How would MPP compare with balanced watermark allocation directly on native 3DGS, using voxel or kNN-based spatial partitions without UV packing?

  3. In Table 3(b), why does nullspace modulation improve BA from 94.95% to 96.19% despite restricting SH updates? Also, how well do the selected projection directions match the actual training and test views, especially on LLFF? Were the Table 1 baselines rerun under the same evaluation protocol?

Reviewer_4THQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces substantial changes to the 3DGS representation without sufficiently demonstrating why this added complexity is necessary for imperceptible watermarking.

Critical strengths

The experiments across three benchmarks provide the authors with practical research experience in 3DGS watermarking.

Critical weaknesses

Well, I think the key to a watermark is ensuring the embedded information remains invisible. However, to make the plan feasible, the author seems to directly change the structure of the 3DGS model. I think this cannot be accepted.

However, this newly created work also has many problems and limitations. The authors then make small revisions to improve performance.

I am not sure those people in 3DGS reconstruction can accept such a solution. They may not consider this plan.

Besides, the experiments do not clearly establish what drives the improvement. Better fidelity may partly result from retaining more original Gaussians and adding embedding capacity, leaving the independent contribution of the proposed allocation mechanism unclear.

Then, the method improves visual quality at the cost of larger models and slower rendering. Without comparisons under matched resource budgets, its practical advantage over simpler approaches remains uncertain.

At last, the robustness experiments mainly test generic distortions rather than deliberate watermark removal. They therefore provide limited evidence that the method can reliably protect ownership when an attacker knows the embedding mechanism.

Questions for authors

My central concern is that much of the claimed benefit may come from resolving a bottleneck introduced by the chosen representation. The paper does not sufficiently establish why this representation, together with its additional complexity, is necessary for effective 3DGS watermarking.

Reviewer_hh39OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MPP-GS, a 3D Gaussian Splatting watermarking framework that improves the fidelity–decodability trade-off by controlling the spatial distribution of watermark carriers. Multi-Projection Packing assigns Gaussians to multiple candidate UV locations and reduces overflow through occupancy- and importance-aware allocation. Complementary Pair Modulation introduces parent–auxiliary Gaussian pairs whose parameters partially compensate for each other's rendering effects. The method is optimized with a rendering-fidelity loss and a frozen image watermark decoder. Experiments on 23 scenes evaluate multiple payload sizes, baselines, distortions, and ablations, and report strong fidelity and bit accuracy, especially at larger payloads.

Critical strengths
  1. The paper identifies carrier congestion and allocation imbalance as important sources of watermark distortion. MPP greatly reduces Gaussian overflow, while the allocation ablation shows that balanced placement improves rendering quality at similar bit accuracy.
  2. The method achieves a strong fidelity–decoding trade-off across a broad evaluation. The experiments cover 23 scenes, three datasets, three payload sizes, several baselines, multiple distortions, and component ablations. The gains are particularly clear at 48- and 64-bit payloads.
Critical weaknesses
  1. The MPP-vs.-SPP comparison does not clearly separate the benefit of spatial balancing from the benefit of retaining more original Gaussians. SPP discards many primitives because of UV capacity limits, while MPP nearly eliminates this overflow and already achieves much higher clean-carrier fidelity. Therefore, part of the final PSNR gain may come from better carrier reconstruction rather than better watermark allocation.
  2. The robustness evaluation does not include attacks adapted to the CPM structure. Parent and auxiliary Gaussians share the same center and rotation, which may make the pair structure directly detectable from the released model. Random removal, cloning, and Gaussian noise do not test whether an attacker can selectively remove, merge, or refit these pairs while preserving rendering quality.
Questions for authors
  1. Can the authors separate the fidelity gain from primitive retention and spatial balancing, for example with matched retained primitive counts or separate carrier and watermark distortion?
  2. How does the method perform against structure-aware attacks that explicitly detect and modify parent–auxiliary Gaussian pairs?
  3. Does the fidelity advantage remain under matched primitive count, model size, or rendering cost, given the large number of auxiliary Gaussians?
  4. What ownership-verification model is assumed beyond bit recovery, including secret keys, wrong messages, and re-embedding or overwrite attacks?

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

Related papers

Discussion (0)

Sign in to comment.