acceptodds
Under review as a conference paper at ICLR 2027

Observation-Aware QIM for Gaussian Splats

Abstract

When one 3D Gaussian-splatting asset is released to several recipients, a leaked copy should reveal which release it came from. Releases share geometry exactly, so the identifier can live only in the spherical-harmonic (SH) colour coefficients, where it must stay invisible yet survive re-export. Existing marks need trained decoders, and a direct write budgets its change in coefficient distance although camera sensitivity spans a median 7.2 orders of magnitude, leaving margins too thin for 8-bit export. We instead spend the budget in the renderer's own error metric, placing the mark where cameras barely look. Observation-aware quantization index modulation (QIM) solves each keyed quantization write in closed form in a metric built from compositing weights, and a compositing inequality bounds the image error of the complete write on the declared cameras, certifying each write before release. On fourteen scenes at matched training error, exact recovery after 8-bit export rises from 10/70 to 70/70 over Euclidean and diagonal controls, at higher held-out error; at a matched 48-bit payload, every identifier survives half-pruning at a marking cost of dB, against dB for GaussianMarker. Float16 coordinates, SH truncation, and a keyless recipient holding the declared cameras remain failure modes. These results show that solving the write in the renderer's metric lets a training-free mark survive export, and they give native 3D assets a certified release identifier with measured limits.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian splatting watermarking, quantization index modulation, spherical harmonic coefficients, certified distortion bound, robustness to pruning and quantization, asset provenance

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_N3ZXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies release fingerprinting for Gaussian-splatting assets. The authors solve the quantization index modulation write under the renderer’s own metric, so that the mark falls mainly in directions that cameras barely observe. Experiments demonstrate the properties of the method.

Critical strengths

The authors propose a new watermarking scenario in the context of GS assets and provide a solution approach for this scenario via QIM.

Critical weaknesses
  1. This paper does not report bit acc. The baseline methods report accuracy without applying error-correcting codes, whereas this paper uses error-correcting codes, which is an unfair comparison. Additional comparison are needed.

  2. The threat model design is unreasonable. Perturbing geometric parameters is a common evasion technique, especially when the perturbation magnitude does not significantly affect rendering quality. We suggest adding curves of coordinate perturbation versus rendering quality and coordinate perturbation versus bit acc.

Questions for authors

Answer the questions raised by the critical weakness.

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

The paper developed a forensic watermarking (fingerprinting) for 3D Gaussian-splatting assets. The aim is that a leaked copy can be traced back to the release it came from, and the main idea is to put a keyed release identifier into the spherical-harmonic (SH) colour coefficients of a 3D Gaussian-splatting asset.

The paper uses classic QIM, a signal-processing watermark from 2001, instead of a trained neural encoder/decoder in previous work. Methods like GuardSplat and GaussianMarker train neural networks and read the mark from rendered images. This method trains nothing, reads the mark straight from the parameters, and runs in under a second.

Critical strengths
  1. The method needs no training and is cheap. The cost is under a second per write on scenes with millions of Gaussian ellipsoids. It also gives a checkable bound estimating how much the watermark can change the rendered images. None of the learned baseline methods can provide that, which is a unique point of this algorithm.

  2. The core idea is simple and well motivated. The authors show that camera coverage of the SH directions is very uneven, especially in forward-facing scenes. Spending the watermark budget in image space instead of coefficient space makes sense.

Critical weaknesses
  1. The comparison may be designed to specifically favour the method. At equal training-view error, the proposed method makes much larger changes in the unseen directions, so it’s not surprising that it survives rounding better. What happens if the Euclidean baseline is given a larger budget that matches error on held-out views? A clear table for this would help a lot.

  2. The theory contribution is modest. Theorem 1 is standard, as the authors say. Theorem 2 is neat but short, and it’s quite loose in practice. Certified sounds stronger than what is actually guaranteed: low error on the training views only.

  3. I'm worried about the robustness of the marker. The mark is generated where the training cameras don’t look. This has two limitations: 1. It is noticeably more visible from new viewpoints, and new viewpoints are the whole point of a 3D asset. 2. Anyone who has the training camera poses can remove it cheaply. Can authors propose any solution to it?

Questions for authors
  1. If the Euclidean baseline matches the method’s error on held-out views, does it also survive 8-bit export?

  2. In what realistic scenario does the recipient not have the camera poses?

  3. What happens when two marked copies are averaged?

Reviewer_w3S1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies recipient-specific watermarking of 3D Gaussian Splatting assets. It proposes an observation-aware QIM method that replaces Euclidean coefficient distortion with a renderer-aware quadratic metric over spherical-harmonic coefficients. The method computes a closed-form watermark update that prefers SH directions with low rendered impact under a declared camera set. The paper also derives a finite-camera distortion bound and develops a reference-assisted verification protocol based on registered Gaussian centers. Experiments evaluate robustness to export, pruning, SH truncation, noise, cropping, similarity transforms, coordinate perturbations, removal attacks, and transfer to additional captures.

Critical strengths
  1. The renderer-aware metric and finite-camera distortion certificate are technically meaningful contributions. The metric better reflects rendered distortion than Euclidean coefficient distance, and the resulting QIM update has a simple closed-form solution. The certificate also provides a stronger fidelity guarantee than relying only on empirical image-quality metrics.
  2. The controlled evaluation clearly isolates the benefit of the proposed metric. Observation-aware QIM substantially improves identifier recovery over Euclidean and diagonal variants under matched distortion, while the angular control helps identify the contribution of renderer weighting. The paper also evaluates multiple perturbations, keys, fragments, and transfer settings, and reports unsuccessful configurations and failure cases.
Critical weaknesses
  1. The verification pipeline is highly sensitive to coordinate processing. The method relies on registering Gaussian centers to a reference, and small center perturbations or FP16 coordinates already cause substantial verification failures. This limits robustness to practical re-export or compression pipelines that modify geometry representation.
  2. The threat model becomes weaker when the recipient has access to the declared cameras. The proposed metric places watermark energy in directions that are inexpensive under those cameras, but this structure can also be exploited for removal. The reported camera-informed attacks remove most or all identifiers at small visual cost, so the security claim depends strongly on the assumed availability of the camera set.
Questions for authors
  1. What deployment scenarios justify assuming that a malicious recipient does not have access to the declared cameras?
  2. How much of the coordinate-precision failure comes from the current registration method rather than the watermark itself?
  3. How should invisibility be interpreted outside the declared camera set, where the distortion certificate does not apply?
  4. Should the main empirical claim focus on advantages over matched QIM controls rather than broad superiority over existing 3DGS watermarking methods?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.