acceptodds
Under review as a conference paper at ICLR 2027

SphericalMark: Third-Moment Preserving Diffusion Watermarks with Adaptive Augmentation Detection

Abstract

Diffusion watermarking embeds a learnable signature into the initial noise so that generated images remain traceable without degrading generation quality. Existing in-processing methods normalize this signature to match only the first two moments of the target Gaussian—mean and variance—leaving its skewness unconstrained; the resulting distributional mismatch is a structural limitation that current normalization schemes do not address. To address this gap, we propose SphericalMark, which introduces spherical third-moment normalization to explicitly constrain the third moment of the injected signature through a differentiable regularizer that combines bounded odd-symmetric transformation (tanh), re-centering, and norm restoration. This yields a watermark distribution more faithful to the target Gaussian while remaining fully end-to-end optimizable. To strengthen detection robustness, we further introduce an adaptive augmentation sampler (OAS) that continually reallocates training probability toward the perturbations on which the detector currently performs worst. Evaluated on Stable Diffusion 2.1, SphericalMark achieves 78% lower quality degradation than state-of-the-art SERUM (FID: 0.81 0.18) while maintaining competitive robustness (average TPR: 98.55% @ 1% FPR, within 1.5% of SERUM for 7/8 perturbations). We validate the distributional benefits through direct statistical measurement and demonstrate that SphericalMark offers a practical, drop-in refinement that better respects the target noise distribution without sacrificing efficiency.

Primary area:
generative models
Keywords:
Diffusion watermarking, Third-moment normalization, Adaptive augmentation, Distribution preservation, Watermark robustness

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_97poOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SphericalMark, a tuning-free diffusion watermarking method that addresses two key limitations of existing initial-noise watermarking approaches such as SERUM: unconstrained third-moment (skewness) of the injected signature, which causes distributional mismatch and quality degradation, and uniform augmentation sampling during detector training, which wastes compute on easy perturbations and leaves hard ones under-trained.

Key contributions:

  1. Spherical third-moment normalization (SMP): a differentiable regularizer that combines an odd-symmetric tanh transformation, re-centering, and norm restoration to explicitly constrain the skewness of the watermark signal. This produces a watermark distribution far more faithful to the target Gaussian while remaining fully end-to-end trainable.
  2. Adaptive augmentation sampler (OAS): dynamically reallocates training probability toward the perturbations on which the detector currently performs worst, concentrating compute on hard cases instead of sampling all augmentations uniformly.
  3. A two-stage progressive training pipeline that decouples quality optimization (via SMP) and robustness optimization (via OAS), avoiding the conflicting-objective degradation observed in joint training.

Validation:

Evaluated on Stable Diffusion 2.1. Compared with state-of-the-art SERUM, SphericalMark reduces quality degradation by 78% (ΔFID: 0.81 → 0.18) while maintaining competitive detection robustness: an average TPR of 98.55% at 1% FPR, within 1.5% of SERUM on 7 out of 8 standard perturbations. Ablation studies confirm the complementary roles of SMP and OAS and validate the advantage of progressive training over joint training.

Critical strengths
  1. Principled third-moment normalization with formal guarantees and full differentiability

The paper introduces spherical third-moment normalization built on the odd-symmetry property of the tanh transformation, which mathematically enforces zero skewness of the watermark signature while preserving zero mean, unit variance, and fixed norm. The regularizer is fully differentiable and end-to-end trainable within the detection loop, instead of being a post-hoc correction, and recovers standard two-moment normalization in the limiting case. This is consequential because it addresses a structural distribution-mismatch flaw in existing initial-noise watermarking with a theoretically grounded solution, not an engineering trick, giving the contribution depth and generalizability. 2. Rigorous ablative evaluation with clear characterization of the quality-robustness tradeoff

The evaluation benchmarks against the state-of-the-art SERUM method across eight perturbations, including component ablations, regularization weight sweeps, and training-strategy comparisons to isolate the contributions of SMP and OAS, and demonstrates that progressive two-stage training avoids the optimization conflict observed in joint training. Importantly, the paper does not overclaim — it explicitly acknowledges the rotation robustness tradeoff and delineates applicable scenarios.

Critical weaknesses
  1. Substantial and structural drop in rotation robustness with no mitigation pathway

Rotation robustness drops markedly from 99.34% (SERUM) to 92.22% (SphericalMark) — a 7.12 percentage point gap — and the paper attributes this to an inherent tradeoff of the spherical manifold projection, which preserves spectral characteristics at the cost of geometric transformation invariance. For a watermarking method, robustness to common geometric transformations is a core functional requirement, and rotation is a prevalent real-world attack. The paper frames this as a deliberate design choice but provides no mitigation strategy, configurable tradeoff knob, or alternative variant to close this gap. This materially limits practical deployability and weakens the claim that the method offers a broadly superior balance. 2. Limited benchmarking scope and insufficient validation of perceptual imperceptibility

The evaluation only benchmarks against SERUM, a single tuning-free initial-noise baseline, and does not compare against other established watermarking paradigms such as post-processing methods, tuning-based methods, or inversion-based approaches like Tree-Ring. Additionally, quality improvement is measured primarily via FID, without supplementary human perceptual studies or systematic tests of imperceptibility, and there is no evaluation against stronger adversarial removal attacks. This narrow scope makes it difficult to assess the method’s standing relative to the broader watermarking landscape and to verify that the FID improvement corresponds to meaningful perceptual benefits in real use cases.

Questions for authors

The paper reports a 7.12 percentage point drop in rotation TPR compared with SERUM, attributing it to an inherent tradeoff of the spherical manifold projection that prioritizes spectral fidelity over geometric invariance. Could the authors clarify whether this degradation is a necessary consequence of third‑moment normalization itself, or whether there exist architectural variants or configurable tradeoff knobs that can narrow this gap while retaining the quality benefits? Additionally, have any mitigation strategies been explored for rotation robustness beyond the current default configuration?

The evaluation primarily benchmarks against SERUM as the only tuning‑free initial‑noise baseline, without comparisons to other established watermarking paradigms such as post‑processing methods, tuning‑based methods, or inversion‑based approaches like Tree‑Ring. Furthermore, quality gains are measured only via FID, without human perceptual studies or systematic tests of visual imperceptibility, and no evaluation against adversarial removal attacks. Could the authors clarify whether supplementary results exist for these additional benchmarks and validations?

1 comment
AuthorsOpenReview ↗

Rotation-quality tradeoff is tunable, and benchmark scope clarification

Dear Reviewer 97po,

Thank you for recognizing the motivation of SMP and the ablation coverage. We agree rotation and benchmark scope need tightening, and we will revise accordingly.

Q1: Rotation robustness drop and mitigation pathway.

Response. The drop is real: 99.34% to 92.22% at 1% FPR, -7.12pp. We do not claim it is a proven necessary consequence of third-moment normalization itself.

Existing evidence already shows tunability. In Table 2, rotation rises with ww: 91.70% (w=0.001w=0.001) to 92.22% (w=0.005w=0.005) to 93.22% (w=0.01w=0.01). OAS also directly mitigates it: rotation receives ~35% priority vs. 12.5% uniform, improving from 86.06% to 92.22% (+6.16%) in Fig. 3. Our default w=0.005w=0.005 was chosen as the quality optimum (Δ\DeltaFID 0.18), not the rotation optimum.

Revision. We will reframe Sec. 4.2 from “inherent tradeoff” to “observed tradeoff under the quality-optimal setting,” add a ww-vs-rotation-quality curve, document OAS rotation priority as the current mitigation, and list rotation-equivariant detector and rotation-heavy Stage-2 as explicit future variants. We will also narrow the “broadly superior balance” claim to high-fidelity scenarios where JPEG/noise/crop dominate.

We are also tightening the “guarantee” wording raised in discussion: SMP is an empirically effective differentiable regularizer, not a proof of zero skewness for arbitrary centered inputs.

Q2: Limited baselines and perceptual validation.

Response. Scope was intentionally narrow: tuning-free initial-noise methods with frozen generator and inversion-free detection. SERUM is the direct SOTA in that cell. Tree-Ring requires diffusion inversion, Stable Signature requires decoder fine-tuning, and post-processing watermarks have a different threat/cost model, so published numbers are not directly comparable.

We agree FID alone does not prove perceptual imperceptibility, and we have no human study or adversarial-removal evaluation.

Revision. We will add a paradigm-scope table clarifying cost and detection assumptions, keep SERUM as the primary matched baseline, cite other paradigms only qualitatively without implying matched comparison, add CLIP-stability plus LPIPS limitation discussion, and explicitly list missing human evaluation and adaptive/adversarial removal attacks as limitations.

We hope this clarifies the tunable nature of the rotation gap and the intended benchmark scope. If these revisions address your concerns, we would be grateful if you could consider updating your rating.

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

This paper extends SERUM with a proposed third-moment normalization and progressive training to improve generation quality while retaining watermark robustness. The objective is relevant, and the reported detection results are strong under standard perturbations. However, I have substantial concerns about the central mathematical guarantee, the attribution of novelty, and the validity of the headline quality comparison.

Critical strengths
  1. Improving the quality–robustness trade-off while retaining fast, inversion-free detection is a practically relevant objective.

  2. The paper reports TPR at a specified FPR across eight perturbations and includes ablations over regularization weights and training strategies.

Critical weaknesses
  1. The claimed zero-skewness guarantee is incorrect. Section 3.2 argues that the odd symmetry of tanh implies zero expectation for any centered distribution. Zero mean does not imply symmetry. For example, repeating the entries (−1,−1,−1,3)(-1,-1,-1,3) and applying the complete proposed mapping—standardization, tanh, re-centering, and norm restoration—still yields skewness 2/3≈1.1552/\sqrt{3}\approx1.155. The explicit skewness penalty may reduce skewness empirically, but this is different from the claimed guarantee.

  2. The distribution-preservation argument does not follow from the analysis. Under the stated model, fixed signature norm and watermark strength give DKL(q∣p)=−d2log⁡(1−α),D_{\mathrm{KL}}(q|p)=-\frac d2\log(1-\alpha), independently of signature skewness. Thus, reducing the empirical skewness of signature coordinates does not improve this joint latent KL. The paper needs to distinguish these statistical quantities and substantiate the connection between its regularization and image quality.

  3. The novelty and motivation are inaccurately positioned relative to SERUM. SERUM already includes the adaptive augmentation sampler, with the same sampling formula, update rules, and default hyperparameters. It also samples its signature once before detector training, so the motivation based on gradient-induced signature skewness does not describe this baseline. Section 4.1 acknowledges inheriting SERUM’s OAS, contradicting its presentation elsewhere as a newly introduced component.

  4. The claimed 78% quality improvement is not established by a matched comparison. SERUM’s published ΔFID=0.81\Delta\mathrm{FID}=0.81 comes from 18.43→19.2418.43\rightarrow19.24, whereas this paper reports 104.96→105.14104.96\rightarrow105.14, using a different sample count and evaluation setup. Comparing these differences does not isolate an improvement attributable to SMP. No uncertainty estimates are provided. Moreover, a small difference between two FIDs against COCO does not establish that the clean and watermarked image distributions are nearly identical.

  5. The implementation and experimental descriptions are inconsistent. Section 4.1 and the reproducibility statement disagree on the sampler, prompt dataset, optimizer, learning rate, and batch size. The manuscript also describes normalization as both continuously attached to training and a one-time initialization operation. These discrepancies prevent a clear understanding of the implemented method.

Questions for authors
  1. Is low skewness an empirical consequence of the explicit penalty rather than a guarantee of the mapping? Which distributional quantity improves when the stated KL remains unchanged?

  2. What specifically distinguishes the proposed OAS from SERUM’s existing sampler, and how is the claimed baseline signature skewness obtained?

  3. Does the quality advantage persist when both methods are evaluated with identical prompts, seeds, sampling settings, sample counts, and FID implementation, with repeated measurements?

  4. Which parameters are updated during SMP fine-tuning, and how do gradients reach the signature? Please reconcile the conflicting configurations and include a continuation baseline with the same training budget but without SMP.

1 comment
AuthorsOpenReview ↗

Clarifications on skewness claim, KL scope, OAS attribution, and matched FID evaluation

Dear Reviewer tRPy,

Thank you for the careful and precise review. We agree with the core technical points on the skewness guarantee, KL scope, OAS attribution, and matched FID comparison, and we will revise accordingly.

Q1: Zero-skewness property of the tanh mapping.

Response. You are correct; our guarantee statement is wrong.

Odd symmetry alone does not imply zero mean for an arbitrary centered but asymmetric input. Your counterexample stands. What reduces skewness empirically in our training is the explicit penalty w⋅skew(w)2w\cdot \text{skew}(\mathbf{w})^2 together with re-centering, not a mathematical guarantee from tanh⁡\tanh.

Revision. We will delete the claim that tanh⁡(−x)=−tanh⁡(x)\tanh(-x)=-\tanh(x) ensures E[tanh⁡(X)]=0\mathbb{E}[\tanh(X)]=0 for any centered XX in Sec. 3.2, replace “guarantees zero skewness” with “empirically reduces coordinate-wise skewness,” and report measured ∣γ∣|\gamma| before/after mapping plus the penalty contribution in ablations.

Q2: Relation between signature skewness, joint KL, and image quality.

Response. You are correct; under fixed norm ∥A′∥22=d\|\mathbf{A}'\|_2^2=d and fixed α\alpha, the stated DKL(q∥p)=−d2log⁡(1−α)D_{KL}(q\|p)=-\frac{d}{2}\log(1-\alpha) does not depend on skewness.

Our intended claim is narrower: coordinate-wise skewness is a marginal-shape diagnostic whose effect propagates through denoising/decoding and is therefore evaluated by FID, not by that joint latent KL. We conflated the two quantities.

Revision. We will keep the KL expression only as a norm/strength diagnostic, explicitly state it does not improve with lower skewness, and reframe SMP as an empirical quality regularizer validated by matched FID/CLIP and direct γ\gamma measurement.

Q3: Novelty and attribution of OAS.

Response. We agree the presentation is inconsistent.

We reuse SERUM’s OAS formulation, including sampling formula, update rule, and defaults, and Sec. 4.1 already states we build on it. It should not have been framed as a newly introduced component alongside SMP.

Revision. We will move OAS to “reused baseline component” throughout Abstract/Intro/Method, keep only the OAS-pretraining vs. joint-training ablation as our analysis contribution, and clarify that gradient-induced skewness is our motivation for SMP, not a description of SERUM’s sampler.

Q4: Validity of the 78% quality improvement.

Response. We agree the current comparison is not matched.

SERUM Δ\DeltaFID and ours use different clean baselines, sample counts, prompts, sampler settings, and likely FID implementations, so 0.81→0.180.81\to0.18 does not establish a causal improvement, and a single FID difference does not imply near-identical distributions.

Revision. We will either (a) rerun both methods under identical prompts, seeds, 50 DDIM steps, CFG 7.5, sample counts, and pytorch-fid with repeated runs reporting mean$\pm$std, or (b) if space/time does not permit, narrow the claim to “lower Δ\DeltaFID under our protocol” and remove “78% over SERUM” from Abstract/Conclusion. We will also add uncertainty and state FID limitations explicitly.

Q5: Implementation consistency and SMP training mechanics.

Response. The inconsistencies you list are real. Stage 1/Stage 2 optimizer, LR, batch size, prompt set, and sampler description differ between Sec. 4.1 and Reproducibility Statement, and “continuously attached” vs. “one-time projection” is contradictory.

Revision. We will unify all settings to the actual run configs, specify exactly which parameters are updated in SMP fine-tuning and how gradients reach the signature A\mathbf{A}, add a continuation baseline with the same training budget but without SMP, and publish exact configs/seeds. The “one-time projection” sentence will be removed or corrected to describe initialization only.

We hope these concrete revisions address your concerns, and we welcome further guidance on which matched evaluation you consider decisive.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.