acceptodds
Withdrawn from ICLR 2027

Robust Generative Image Watermarking via Texture-Aware Initial Noise Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Robust watermark recovery remains challenging in generative image watermarking under lossy channel distortions. In inversion-based diffusion approaches, we find that initial noise, typically sampled randomly and treated as interchangeable, can substantially affect post-distortion latent recovery, even when the model, prompt, and generation configuration are fixed. More importantly, this variation can be anticipated before transmission: the gray-level co-occurrence matrix (GLCM) homogeneity of the generated image exhibits a consistent positive correlation with latent recovery error under both JPEG compression and additive Gaussian noise. Based on this observation, we propose a texture-aware initial noise selection mechanism that samples candidate initial noises and selects a candidate whose resulting watermarked image exhibits low GLCM homogeneity. The mechanism requires neither modification of the embedding mapping nor retraining of the generative model, and introduces only bounded, controllable sender-side generation overhead. Extensive experiments across multiple existing latent embedding schemes, payload capacities, prompts, and channel distortions demonstrate consistent improvements in payload extraction accuracy, with an average gain of 2.78 percentage points. Our work identifies initial noise selection as an effective design dimension for improving the robustness of inversion-based diffusion watermarking.

Primary area:
generative models
Keywords:
Generative image watermarking, diffusion models, gray-level co-occurrence matrix, initial noise selection, robustness

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

The paper studies watermarks for diffusion models that embed a message in the starting noise and detect it by inverting an image to the noise. The authors observe that, for the same prompt and message, some starting noises survive JPEG compression and added noise better than others. They also find that this can be predicted before any compression or noise is applied, because images with more fine texture tend to be recovered more accurately. The proposed method generates up to ten candidate images per message and keeps the first one that is textured enough. Across four watermarking methods, two prompt sets, and several message lengths, this raises bit accuracy by 2.8 percentage points on average under JPEG and Gaussian noise.

Critical strengths
  1. The method is simple and works with existing watermarking techniques. It needs no training, and it does not change how the message is embedded or detected.
  2. The paper introduces a novel way to improve robustness. Prior work changes how the message is embedded in the noise. This paper keeps that fixed and instead picks the initial noise, using the image's texture to predict in advance which noises will survive edits.
Critical weaknesses
  1. Key watermarks that also hide in the starting noise are missing. The paper doesn't check whether selection helps methods like Tree-Ring [1], WIND [2], or SEAL [3], or how they compare under the same attacks.
  2. Choosing images by texture changes which images the model produces, and the paper does not test whether this shifts the distribution of generated images.
  3. The extra robustness gain is limited and expensive. It appears only under JPEG and Gaussian noise, and costs about 7 generations per image for an average gain of 2.8 points.

References

[1] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-Ring Watermarks: Invisible Fingerprints for Diffusion Images. NeurIPS 2023.

[2] Kasra Arabi, Benjamin Feuer, R. Teal Witter, Chinmay Hegde, and Niv Cohen. Hidden in the Noise: Two-Stage Robust Watermarking for Images. ICLR 2025.

[3] Kasra Arabi, R. Teal Witter, Chinmay Hegde, and Niv Cohen. SEAL: Semantic Aware Image Watermarking. ICCV 2025.

Questions for authors
  1. With the same budget of 10 candidates, how does texture-based selection compare with compressing each candidate right after generating it, decoding it, and keeping the one that decodes best? How does it compare with Qi et al.'s recovery score [4]?
  2. Can a simple classifier separate selected images from unselected ones? How do FID and the distribution of the texture score change after selection?
  3. Table 5 shows the cost of different budgets and thresholds, but not their accuracy. How does accuracy change as the number of candidates goes from 1 to 10?

References

[4] Zipeng Qi, Lichen Bai, Haoyi Xiong, and Zeke Xie. Not All Noises Are Created Equally: Diffusion Noise Selection and Optimization. arXiv:2407.14041, 2024.

Reviewer_VjdbOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the impact of initial noise selection on robustness in inversion-based generative image watermarking. The authors observe that different initial noises can lead to substantial variations in latent recovery error and payload extraction accuracy after image distortions, even when the generative model, prompt, and generation configuration are fixed. They further identify Gray-Level Co-occurrence Matrix (GLCM) homogeneity of the generated image as a useful predictor of latent recovery error under JPEG compression and additive Gaussian noise. Based on this observation, the authors propose a lightweight, texture-aware initial noise selection mechanism that samples candidate initial noises and selects candidates with lower GLCM homogeneity. The method requires neither retraining the generative model nor modifying the underlying payload embedding or extraction procedures, and can be integrated into existing inversion-based watermarking pipelines.

Critical strengths

a. Model-Agnostic and Plug-and-Play Integration: The proposed mechanism operates at the initial noise selection stage and does not require retraining the generative model or modifying the underlying payload embedding and receiver-side extraction procedures. This makes it a potentially practical complement to existing inversion-based watermarking methods. b. Empirical Analysis of Texture and Latent Recovery: The paper provides a systematic empirical study of the relationship between GLCM homogeneity and latent recovery error. The comparison across different texture statistics and GLCM spatial distances helps motivate the choice of GLCM homogeneity as the screening indicator.

Critical weaknesses

a. Insufficient Analysis of Prompt-Dependent Selection Bias and Generation Diversity: Although the paper evaluates two prompt sets, it does not sufficiently characterize how prompt semantics influence the distribution of GLCM homogeneity scores or the effectiveness of initial noise selection. Since the method favors candidates with lower GLCM homogeneity, it may introduce a selection bias in the generated outputs. In particular, it remains unclear whether this preference affects visual diversity or semantic fidelity for prompts that intentionally request smooth regions, minimal backgrounds, or other low-texture aesthetics. The reported CLIP Score and MUSIQ results are encouraging, but aggregate quality metrics alone may not fully capture such prompt-specific effects. b. Limited Robustness Evaluation and Incompletely Characterized Computational Cost: The main robustness improvements are demonstrated under JPEG compression and additive Gaussian noise, while the appendix shows that GLCM homogeneity does not consistently predict latent recovery error under Gaussian blur, resizing, center crop, and brightness adjustment. A broader evaluation using established image-watermarking benchmarks and attack suites—such as benchmarks such as WAVES (ICML'24), WaterBench (ICLR'25), MarkDiffusion (JMLR'26), WARP (ACM MM'26), or Stirmark 3.0/4.0—would help assess the method's robustness under more diverse and standardized attack settings. In particular, it would clarify whether the proposed initial noise selection mechanism generalizes beyond the distortion types for which the texture-based indicator demonstrates a consistent relationship with latent recovery error. In addition, although the paper reports an average of 7.231 generation attempts under the default setting (M=10, T=0.10), the actual wall-clock latency and relative computational overhead across different generation configurations remain insufficiently characterized.

Questions for authors

a. Prompt Influence and Selection Bias: How does the distribution of GLCM homogeneity scores vary across prompts, particularly between smooth or minimalist concepts and highly textured scenes? Can the authors quantify the proportion of candidates rejected by the threshold and evaluate whether selection affects semantic fidelity or visual diversity for different prompt categories? b. Robustness Beyond the Targeted Distortions: Given that the proposed indicator does not exhibit a consistent relationship with latent recovery error under Gaussian blur, resizing, center crop, and brightness adjustment, how well does the selection mechanism generalize to stronger distortions and geometric attacks? Can the authors provide additional evaluations under more comprehensive robustness protocols and clarify the scope of attacks for which the method is expected to be effective? c. Computational Overhead and Practical Efficiency: Under the default setting, the method requires an average of 7.231 generation attempts. What is the corresponding wall-clock latency and computational cost relative to generating a single image without selection? How does this overhead vary across different diffusion backbones, samplers, and screening parameters?

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

Based on the observation that latent recovery error after distortion varies depending on the initial noise, this paper proposes a sender-side noise selection method using GLCM homogeneity (gh4gh4). It selects the first candidate that satisfies a threshold; if no candidate satisfies the condition within a maximum number of attempts, it selects the candidate with the lowest gh4gh4. The existing embedding and extraction procedures remain unchanged.

The paper reports an average bit accuracy improvement of 2.78 percentage points across GS, SW, ADPM, and RGSD. Improvements are observed in 162 out of 168 attack configurations involving JPEG compression and Gaussian noise, with the remaining configurations resulting in ties, while maintaining comparable levels of CLIP and MUSIQ scores. Similar improvements are reported on SD2.1 and DPM-Solver++.

Critical strengths

The approach of enhancing robustness through initial noise selection while preserving the existing embedding method is interesting.

It is practically useful as it can be applied without additional training or modifications to the receiver side. Consistent improvements are demonstrated across the evaluated JPEG and Gaussian noise configurations, and all six tied configurations correspond to cases where the accuracy is already 100% (Tables 6–7). The work is meaningful in that it introduces noise selection as a complementary dimension alongside existing payload encoding and latent mapping strategies.

Critical weaknesses
  1. The evaluation scope of embedding methods and generative models is too limited to judge general applicability.

Although the paper claims independence from specific embedding mappings, experiments on frequency-pattern-based methods are lacking (Section 5.1). The validation of generative models is also confined to SD1.5 and SD2.1 (Section 5.5). Given that the relationship between gh4gh4 and recovery error is supported empirically, it is crucial to verify whether this relationship and the performance gains hold across other families of embedding methods as well as flow matching and DiT models.

  1. The generalizability of the metric, statistical reliability, and accuracy-cost trade-off lack sufficient validation.

Table 1 supports the choice of gh4gh4 by comparing the correlation of alternative statistics across various distances. However, it is unclear whether the data used for metric and parameter selection is strictly separated from the evaluation data. The number of candidates per prompt is limited to 10, and the aggregation method of Pearson correlation coefficients across prompts, along with the uncertainty of key performance metrics, is insufficiently explained.

The average number of generation attempts is 7.231. While Table 5 shows gh4gh4 and the number of attempts with respect to MM and TT, it does not provide the relationship between extraction accuracy and actual runtime. A comparison against a direct baseline that selects candidates based on actual extraction accuracy is needed, while accounting for the additional inversion and extraction costs. Furthermore, the paper needs to explain whether the assumptions required for the distributional and security guarantees of the underlying methods are preserved after selection.

Questions for authors
  1. Embedding methods: Can you evaluate the applicability and effectiveness on representative methods among Tree-Ring/RingID, WIND, GaussMarker, and HSTR/HSQR? Please provide the corresponding detection, identification, or bit accuracy suitable for each method, alongside the relationship between gh4gh4 and recovery error.

  2. Generative models: Can you verify the relationship between gh4gh4 and recovery error, as well as performance improvements on representative configurations, using SD3 or FLUX? Small-scale partial experiments would also be helpful.

  3. Metrics and statistics: Are the 800 prompts used for analysis disjoint from the 5,000 prompts in each of the two evaluation sets? What dataset was used to determine dd and TT, and how were the prompt-level Pearson correlation coefficients aggregated? Please also provide standard deviations or confidence intervals across repeated trials for key performance metrics.

  4. Cost and selection effect: Can you present extraction accuracy alongside measured wall-clock time with respect to MM? Please compare this with a baseline that selects candidates based on actual extraction accuracy after channel simulation, distinguishing between matched candidate count and matched computational budget. Additionally, please clarify the distributional differences before and after selection and whether existing theoretical guarantees hold.

  5. Table 4: What are the payload capacity, prompt configuration, averaging methodology, and regeneration procedures for each model and sampler? The GS values can be reproduced by averaging GS-1024 from Table 6 across both prompt sets, but the SW values cannot be reproduced using the same method, necessitating clarification on the configuration.

The approach of selecting initial noise is interesting and well-suited for ICLR. However, there is insufficient validation regarding the scope of application, metric generalizability, statistical reliability, and accuracy relative to computational cost. I am willing to raise my score if representative results on additional embedding methods and models, reinforced statistical reporting, and cost-aware baseline comparisons are provided in the rebuttal.

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

Related papers

Discussion (0)

Sign in to comment.