acceptodds
Under review as a conference paper at ICLR 2027

PreFix: Robust Inversion-Based Image Watermarking via Generation-Side Pre-Compensation

Abstract

Robustness to distortions is essential for reliable generative image watermarking. Among existing approaches, inversion-based methods offer a promising solution by embedding watermarks into the initial noise latent and recovering them through inversion. However, distortions can perturb the inversion input and compromise watermark recovery. To address this, we take a generation-side perspective on robustness, accounting for distortion effects on watermark extraction before image release. Building on this insight, we propose PreFix, a generation-side pre-compensation framework that optimizes an imperceptible residual per image to improve robustness without modifying the underlying inversion algorithm. Since non-geometric degradations perturb image statistics while geometric transformations introduce spatial misalignment, PreFix combines two complementary mechanisms within a unified residual optimization. Specifically, (1) degradation-pool optimization improves watermark recovery under sampled non-geometric degradations; and (2) alignment-aware optimization uses a lightweight alignment predictor to guide the residual to support transformation estimation and pre-inversion correction. Extensive experiments across watermarking methods and generative models demonstrate substantial robustness gains under both distortion types.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion model, Generative image watermarking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6eE8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PreFix aims to improve the robustness of inversion-based generative image watermarking against geometric and non-geometric distortions. It first learns a shared template and an alignment predictor, then optimizes an additive residual for each image before release. The optimization accounts for sampled non-geometric distortions, while the alignment predictor corrects geometric transformations before inversion. The evaluation covers four watermarking methods on SD-v1.4 and SD-v2.1, with additional GS results on SDXL and SD-v3.

Critical strengths
  1. The reported robustness gains are substantial across several watermarking methods and generative models. For GS on SD-v2.1, geometric bit accuracy increases from 0.502 to 0.944. The experiments across different distortion strengths and the component ablations help clarify the effects of the proposed method.

  2. The study of the optimization budget for each image is useful. In particular, the template and alignment configuration without refinement for each image (K=0K=0) already achieves 0.923 geometric bit accuracy for GS (Table 7). This result helps distinguish the contribution of alignment from the additional improvement obtained through residual optimization. The reported computational costs also make this comparison relevant to practical use.

Critical weaknesses

1.The novelty of PreFix is not sufficiently clear. SynTag [1] has already studied geometric robustness in inversion based image watermarking using synchronization and geometric correction. The paper should explain more clearly how PreFix differs from this work. Table 7 also raises a question about the benefit of the proposed optimization. For GS, using the shared template and alignment predictor without further optimization already improves geometric bit accuracy from 0.502 to 0.923. The full method reaches 0.944, but generation time increases from 7.29 s to 28.14 s. The additional benefit of optimizing the residual for each image therefore seems limited for geometric robustness, given the extra cost.

2.Some important implementation details are missing. Equation (7) defines the optimization objective, but it is unclear how the extraction loss and its gradients are handled for GS and PRC. The paper also does not report the loss weights or the numerical fidelity constraint. These details are important for reproducing the method and understanding how much image distortion is introduced to improve robustness.

Reference [1] "SynTag: Enhancing the Geometric Robustness of Inversion-based Generative Image Watermarking." ICCV, 2025.

Questions for authors
  1. How does PreFix compare with SynTag under comparable watermark payloads, detection operating points, and image fidelity? Could the authors also distinguish the contribution of the shared template and alignment from that of residual optimization for each image?

  2. What differentiable objective is optimized for each watermarking method? Please clarify the losses specific to each method, loss weights, distortion sampling procedure, and gradient treatment of watermark extraction, JPEG quantization, and diffusion inversion.

  3. What numerical fidelity constraints are used in Equations (2) and (7)? Could the authors provide paired image quality measurements, such as PSNR and LPIPS, together with template and residual magnitude statistics for K=0K=0 and the default K=6K=6 configuration? A comparison between robustness and fidelity would help clarify the cost of the additional optimization.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 6eE8 (1/2)

Thank you for your constructive suggestions. To facilitate reproducibility and inspection of implementation details, we provide our code at https://anonymous.4open.science/r/PreFix-640C/. Below, we address your questions.

A1. Comparison with SynTag (W1, Q1)

Both SynTag [1] and PreFix use synchronization for geometric correction. The key distinction is recovery-driven pre-compensation. PreFix uses the watermark recovery loss computed after geometric correction to optimize a per-image residual, whereas SynTag learns synchronization cues independently of the watermark recovery objective. PreFix also optimizes the same residual for watermark recovery under non-geometric distortions. We evaluate SynTag using the code provided by its authors. Tables 1 and 2 show that PreFix achieves higher average robustness under both non-geometric and geometric distortions.

Table 1. Comparison with SynTag on SD v2.1 (TPR/Bit Acc.).

MethodPayloadCleanNon-geo.Geo.FIDCLIP-Score
Tree-Ring11.000/--0.892/--0.578/--25.42 ± 0.120.3651 ± 0.0003
Tree-Ring-SynTag11.000/--0.684/--0.817/--25.22 ± 0.210.3613 ± 0.0004
Tree-Ring-PreFix11.000/--0.980/--0.951/--25.41 ± 0.170.3644 ± 0.0006
GS2561.000/1.0000.958/0.9200.000/0.50225.21 ± 0.200.3637 ± 0.0004
GS-SynTag2561.000/1.0000.995/0.9120.930/0.80825.21 ± 0.170.3615 ± 0.0006
GS-PreFix2561.000/1.0000.998/0.9680.992/0.94425.20 ± 0.220.3637 ± 0.0005

Table 2. Per-distortion results on SD v2.1 (TPR/Bit Acc.).

DistortionTree-RingTree-Ring-SynTagTree-Ring-PreFixGSGS-SynTagGS-PreFix
JPEG0.844/--0.598/--0.992/--0.948/0.9161.000/0.9761.000/0.977
GauBlur1.000/--0.728/--1.000/--0.956/0.8131.000/0.8511.000/0.891
MedFilter0.881/--0.731/--0.988/--0.902/0.9071.000/0.9151.000/0.980
GauNoise0.760/--0.604/--1.000/--0.943/0.8781.000/0.9961.000/0.977
S&P Noise0.790/--0.596/--1.000/--0.928/0.8800.990/0.8451.000/0.956
Brightness0.884/--0.634/--0.883/--1.000/0.9820.980/0.8961.000/0.981
Dropout1.000/--0.850/--0.994/--0.988/0.9900.990/0.8770.986/0.990
Cropout0.975/--0.733/--0.986/--1.000/0.9951.000/0.9411.000/0.995
Avg0.892/--0.684/--0.980/--0.958/0.9200.995/0.9120.998/0.968
Rotation0.547/--0.517/--0.924/--0.000/0.5030.920/0.7910.995/0.985
Scale 0.75x0.097/--0.961/--0.903/--0.000/0.5050.940/0.7830.965/0.945
Scale 1.25x0.356/--0.623/--0.963/--0.000/0.5010.860/0.7951.000/0.938
Translation0.892/--0.990/--0.964/--0.000/0.5010.960/0.8331.000/0.952
Shear1.000/--0.992/--1.000/--0.000/0.5020.970/0.8361.000/0.901
Avg0.578/--0.817/--0.951/--0.000/0.5020.930/0.8080.992/0.944

A2. Contribution of Per-image Optimization (W1, Q1).

Alignment accounts for most of the geometric robustness gain, while per-image optimization primarily improves non-geometric robustness. For GS, K=0K=0 substantially improves geometric bit accuracy (0.502→0.9230.502 \rightarrow 0.923). Increasing KK from 0 to 6 primarily improves bit accuracy under non-geometric distortions (0.919→0.9680.919 \rightarrow 0.968), with a further gain under geometric distortions (0.923→0.9440.923 \rightarrow 0.944). A larger KK suits forensic tasks that prioritize reliable watermark recovery, while a smaller KK is practical for latency-sensitive applications. We will further explore ways to reduce cost, including mixed precision and fewer diffusion inversion steps during optimization.

References

[1] SynTag: Enhancing the Geometric Robustness of Inversion-based Generative Image Watermarking

AuthorsOpenReview ↗

Response to Reviewer 6eE8 (2/2)

A3. Implementation details (W2, Q2).

For GS and PRC, we replace the hard sign decision on the recovered latent z^\hat{z} with σ(αz^)\sigma(\alpha\hat{z}) and minimize

Lwm=BCE⁡(σ(αz^),1[z⋆>0]),\mathcal{L}_{\mathrm{wm}} = \operatorname{BCE}\left(\sigma(\alpha\hat{z}), \mathbf{1}[z^\star > 0]\right),

where α=10\alpha = 10 and z⋆z^\star is the original watermarked latent. For Tree-Ring and RingID, we minimize the ℓ1\ell_1 distance between the recovered and target watermark patterns in the Fourier domain. The recovery, fidelity, and alignment losses have weights (3,10,1)(3, 10, 1) in code, corresponding to λimg=10/3\lambda_{\mathrm{img}} = 10/3 and λali=1/3\lambda_{\mathrm{ali}} = 1/3 in Eq. (7).

At each iteration, we uniformly sample four distortions and average their losses. We use the piecewise differentiable JPEG quantization approximation from MBRS [2]. Gradients pass through the VAE encoder and all inversion steps, updating only the per-image residual. Code is available at https://anonymous.4open.science/r/PreFix-640C/.

A4. Robustness–fidelity trade-off (Q4).

Fidelity metrics are computed relative to the original watermarked images before residual addition. Table 3 shows that increasing KK from 0 to 6 improves robustness (Non-Geo.: 0.919 → 0.968; Geo.: 0.923 → 0.944) with a modest PSNR decrease (38.92 → 37.81 dB), while improving SSIM (0.9561 → 0.9694) and LPIPS (0.0551 → 0.0261). Thus, refinement increases pixel-wise error while improving structural and perceptual similarity, offering a favorable robustness–fidelity trade-off.

Table 3. Robustness–fidelity trade-off (Bit Acc. on SD v2.1).

ConfigPSNR (dB) ↑SSIM ↑LPIPS ↓Residual RMS ↓Non-geo. ↑Geo. ↑
GS----0.9200.502
GS-PreFix (K=0)38.92 ± 1.580.9561 ± 0.01630.0551 ± 0.03160.0115 ± 0.00210.9190.923
GS-PreFix (K=6)37.81 ± 1.120.9694 ± 0.01490.0261 ± 0.01100.0130 ± 0.00170.9680.944

References

[2] MBRS: Enhancing Robustness of DNN-based Watermarking by Mini-Batch of Real and Simulated JPEG Compression.

AuthorsOpenReview ↗

Follow-up on Our Response

As the discussion period is coming to an end, we would like to gently follow up and check whether our response has addressed your concerns. If so, we would appreciate it if you could consider updating your score. We are happy to answer any remaining questions.

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

PreFix performs per-image residual optimization on the generation side before image release. By optimizing a small residual, it combines two complementary mechanisms: Degradation-pool Optimization (DO) for robustness against non-geometric degradations and Alignment-aware Optimization (AO) for robustness against geometric transformations. AO uses a lightweight alignment predictor and a universal anchor template to estimate geometric transformations and correct spatial misalignment before inversion. PreFix aims to improve the robustness of existing inversion-based watermarking methods, including Tree-Ring, Gaussian Shading (GS), RingID, and PRC, without modifying their underlying watermark extraction mechanisms.

Critical strengths

a. Substantial robustness improvements for inversion-based watermarking. By combining alignment-aware optimization with degradation-pool optimization, PreFix substantially improves robustness, particularly against challenging geometric transformations such as rotation, scaling, and translation. Under geometric transformations, it improves the bit accuracy of Gaussian Shading and PRC by 43.95 and 39.95 percentage points, respectively, averaged across Stable Diffusion v1.4 and v2.1. b. Generality and non-invasive, plug-and-play design. PreFix can enhance existing inversion-based watermarking methods without modifying the underlying diffusion model, text-conditioning mechanism, or original watermark extraction mechanism. This makes it a potentially practical enhancement to existing watermarking pipelines.

Critical weaknesses

a. Insufficient evaluation on standardized robustness benchmarks and advanced attacks. Although the paper focuses on robust image watermarking, its robustness evaluation primarily relies on a manually selected set of conventional image-processing operations and geometric transformations. It does not provide a systematic evaluation using established or emerging community benchmarks, such as WaterBench (ICLR 2025), MarkDiffusion (JMLR 2026), WARP (ACM MM 2026), or the classic StirMark 3.0/4.0 benchmark suite. For a more comprehensive evaluation of adaptive attacks, the authors could consider WAVES (ICML 2024). For watermark removal attacks, relevant methods and evaluation settings could include WEvade (ACM CCS 2023), regeneration attacks (NeurIPS 2024), averaging attacks (NeurIPS 2024), NFPA (NeurIPS 2025), and MarkNull (USENIX Security 2026). To be clear, the paper does evaluate reconstruction attacks, diffusion-based purification attacks, and combined distortions. However, the current evaluation does not establish how PreFix performs across a broader range of standardized benchmarks and advanced attack settings. Without such evaluations, it remains difficult to assess its robustness under more diverse and realistic threat models. b. Insufficient explanation of the complementary roles of DO and AO. The ablation study in Table 6 suggests that DO primarily improves robustness against non-geometric degradations, whereas AO primarily improves robustness against geometric transformations. The paper concludes that "PreFix provides more consistent robustness across both distortion families, supporting the complementary roles of DO and AO." I consider this complementarity to be one of the paper's most important contributions. However, the paper primarily supports this claim through empirical results and provides limited explanation of the underlying mechanism. A more detailed analysis of why the two objectives reinforce each other, and under what conditions their combination is beneficial, would strengthen the methodological contribution.

Questions for authors

a. Standardized benchmarks and attack robustnessCould the authors evaluate and report the performance of PreFix-enhanced watermarking methods under the benchmark and the aforementioned removal and adaptive attacks? In particular, reporting metrics such as true positive rate (TPR) and bit accuracy would provide a more comprehensive assessment of robustness under diverse attack settings. b. Further explanation of the complementary mechanisms Could the authors elaborate on why Degradation-pool Optimization (DO) and Alignment-aware Optimization (AO) complement each other? Beyond the empirical results, what underlying principles explain their synergy? c. Reducing per-image optimization latency Given that the default optimization budget of K=6K=6 increases generation time to nearly four times the baseline, it is important for the authors to address this trade-off and discuss potential solutions for the time overhead. d. Generalization to unseen and compound attacks How well do the alignment predictor and degradation pool generalize to complex, compound attacks that are not explicitly covered during training or optimization? Examples include non-rigid deformations and local cropping combined with severe JPEG compression. Evaluating such scenarios would help establish the generalization capabilities and limitations of PreFix.

4 comments
AuthorsOpenReview ↗

Response to Reviewer jbKH

Thank you for your constructive suggestions, and for noting "substantial robustness improvements for inversion-based watermarking". Below, we address each of your questions.

A1. Comprehensive robustness evaluation (W1, Q1).

Following your suggestions, we extend our evaluation to WARP (ACM MM 2026) [1], covering 102 configurations: 26 distortion settings, 52 filters, and 24 attacks. Table 1 reports group-averaged results, showing that PreFix improves robustness to both geometric and non-geometric distortions while maintaining its strong performance across the evaluated attacks.

Table 1. Evaluation on WARP (TPR/Bit Acc. on SD v2.1).

Attack / DistortionGSGS-PreFix
Brightness1.000/1.0001.000/1.000
Contrast1.000/1.0001.000/1.000
FrequencyMasking1.000/0.9041.000/0.981
DoGBlur1.000/0.9791.000/1.000
GaussianBlur1.000/0.9721.000/1.000
JPEG1.000/0.9661.000/1.000
Noise1.000/0.9591.000/0.999
RandomCropout1.000/0.9831.000/0.998
CenterCrop0.000/0.5150.831/0.866
RandomCrop0.000/0.5190.820/0.828
PixelShift0.000/0.4901.000/1.000
Resize0.000/0.5390.932/0.956
Rotation0.000/0.4990.941/0.985
InstagramCSSFilters1.000/1.0001.000/1.000
InstagramAttacks1.000/1.0001.000/1.000
AdversarialEmbedding1.000/1.0001.000/1.000
AdversarialEmbeddingPSNR1.000/1.0001.000/1.000
Averaging1.000/0.9991.000/0.999
BM3DDenoising1.000/0.9991.000/1.000
BlurDeblurFPNInception1.000/0.9981.000/1.000
DIPAttack1.000/0.9991.000/1.000
DIPAttackNoise1.000/0.9971.000/1.000
DISCOAttack1.000/0.9931.000/0.995
DiffPureAttack1.000/0.9991.000/1.000
DiffusionRegeneration1.000/0.9991.000/1.000
DoGBlurDeblurFPNInception1.000/1.0001.000/1.000
ImageEditingInstructPix2Pix1.000/0.9991.000/1.000
ImageWatermark1.000/1.0001.000/1.000
LIIFAttack1.000/1.0001.000/1.000
LatentFrequencyMasking1.000/1.0001.000/1.000
MPRNetAttack1.000/1.0001.000/1.000
NRPLarge1.000/0.9741.000/0.978
NRPSmall1.000/0.9971.000/0.997
RealESRGANAttack1.000/0.9991.000/0.999
StegastampInversion1.000/0.9991.000/0.999
TrustMarkRM1.000/1.0001.000/1.000
VAERegeneration1.000/0.9491.000/0.962
WMForger1.000/1.0001.000/1.000
WPWMAttacker1.000/0.9981.000/0.999

A2. Complementary roles of DO and AO (W2, Q2).

Geometric distortions cause spatial misalignment, while non-geometric distortions degrade pixel values. AO enables spatial realignment, whereas DO pre-compensates for degradation-induced recovery errors. AO alone does not explicitly address non-geometric degradation, and DO alone cannot reliably correct spatial misalignment. Jointly optimizing the same residual therefore promotes watermark recoverability across both non-geometric and geometric distortion families.

A3. Robustness–Efficiency Trade-off (Q3).

PreFix trades extra time for better robustness (Section 4.6, Table 7 in the manuscript). At K=0K=0, it adds only 0.17s to the 7.12s baseline on an NVIDIA H100 GPU, improving geometric bit accuracy (0.502 → 0.923). Increasing KK adds runtime but further improves robustness (0.920 → 0.968 on Non-Geo., 0.923 → 0.944 on Geo.). This additional cost may be acceptable for forensic applications that prioritize attribution accuracy, while small KK (e.g., K=0K=0–22) is a practical choice for latency-sensitive scenarios.

We will discuss ways to reduce this cost, including mixed precision, memory-efficient attention, and fewer inversion steps during optimization.

A4. Generalization to unseen and compound attacks (Q4).

We evaluate generalization under unseen compound attacks and elastic deformation (Table 2). We use JPEG (QF =25=25), Gaussian blur (kernel size =7=7), random crop (retained-area ratio =0.8=0.8), rotation (θ=30∘\theta=30^\circ), and elastic deformation (α=200, σ=5\alpha=200,\ \sigma=5). DO improves robustness to signal degradation, AO provides larger gains under geometric distortions, and PreFix combines both to improve recovery under mixed distortions.

Table 2. Generalization to unseen and compound attacks (TPR/Bit Acc. on SD v2.1).

DistortionBaseDOAOPreFix
JPEG → GauBlur0.920/0.7211.000/0.8930.921/0.7271.000/0.902
JPEG → RandomCrop0.430/0.6480.572/0.7550.930/0.8271.000/0.882
JPEG → Rotation → RandomCrop0.000/0.5330.000/0.5871.000/0.9651.000/0.989
ElasticDeform1.000/0.9321.000/0.9441.000/0.9571.000/0.973
JPEG → ElasticDeform1.000/0.8981.000/0.9371.000/0.9431.000/0.969

References

[1] WARP: A Unified Benchmark for Invisible Image Watermarking-Robustness and Protection Against Attacks.

AuthorsOpenReview ↗

Follow-up on Our Response

As the discussion period is coming to an end, we would like to gently follow up and check whether our response has addressed your concerns. If so, we would appreciate it if you could consider updating your score. We are happy to answer any remaining questions.

AuthorsOpenReview ↗

Follow-up on Our Responses

Thank you again for the helpful suggestions. With the discussion period ending later today, we would appreciate your thoughts on our responses to the questions. We hope the analysis and experiments have helped address them, and we would be happy to clarify anything further before the deadline.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer,

As the discussion period is nearing its end, we would greatly appreciate it if you could acknowledge our response. Please let us know whether we have adequately addressed your concerns and if there is anything that remains unclear or requires further clarification.

We would be happy to provide any additional details that may be helpful. We look forward to hearing from you.

Thank you for your time.

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

The paper proposes PreFix, a generation-side add-on that makes inversion-based diffusion watermarks (Tree-Ring, RingID, Gaussian Shading, PRC) more robust without changing the inversion algorithm. Before release, a small residual is optimized for each watermarked image so that the watermark still decodes after distortions. It has two mechanisms. Degradation-pool optimization backpropagates the watermark loss through DDIM inversion under sampled non-geometric distortions. Alignment-aware optimization uses a universal anchor template and a trained alignment predictor to estimate and undo geometric transformations before inversion.

Critical strengths
  1. The paper addresses a known weakness of inversion-based watermarks with a simple, general add-on. Geometric robustness improves substantially for all four base methods, and the gains are consistent across SD v1.4, v2.1, SDXL and SD3, five samplers, guidance scales, step counts and combined distortions.
  2. The evaluation is broad and the ablation is informative. Table 6 separates the contribution of the degradation pool from the alignment mechanism, and the cost analysis in Table 7 reports time against accuracy for different optimization budgets.
Critical weaknesses
  1. The comparison misses the methods closest to the paper’s contribution. The main claim is geometric robustness for inversion-based watermarks, but the paper does not compare against other methods targeting that problem:
  • Robust inversion-based watermarks such as ROBIN or GaussMarker.
  • Approaches that reduce inversion error, such as exact inversion of diffusion samplers.
  • Synchronization methods for deep watermarks, such as the localization in Watermark Anything, combined with an inversion-based watermark. The anchor template and predictor closely resemble classical synchronization templates, so novelty relative to that line of work needs discussion. The post-hoc baselines are also weak under geometric attacks (TrustMark at 0.09), and stronger ones such as VINE or StegaStamp are missing. In addition, Table 7 shows that K=0 (template and alignment only, no residual optimization) already raises geometric bit accuracy from 0.502 to 0.923. Most of the geometric gain therefore appears to come from the synchronization template rather than the per-image optimization, which weakens the case for the central pre-compensation idea.
  1. Several security and evaluation issues are not addressed.
  • Template security. The same template e is added to every image, initializing each residual. An attacker could estimate it by averaging many watermarked images, subtract it to disable alignment, or add it to unwatermarked images to forge. None of this is tested.
  • Base method guarantees. GS is distribution-preserving and PRC is undetectable. A shared template and per-image residual likely break these guarantees, but the paper does not discuss or measure this.
  • False positive rate. TPR is reported at a theoretical FPR of 10⁻⁶. Since the alignment predictor also warps unwatermarked queries, the null distribution may change. Empirical FPR on clean images processed by the full pipeline is not reported.
  • Removal attacks. Reconstruction and purification results are given only for GS-PreFix, with no GS baseline, so the effect of PreFix is unclear. Stronger attacks such as CtrlRegen or adversarial removal are not evaluated, and purification at strength 0.7 removes the watermark completely.
  • Unseen distortions. The optimization pool overlaps the test set: median filtering with w=15 appears in both, and the geometric pool covers the test ranges. An explicit split between seen and unseen distortions is needed to support the generalization claim.
  • Fidelity. Only FID and CLIP are reported. PSNR or SSIM between the image before and after adding the residual would show how large the change actually is.
Questions for authors
  1. How does PreFix compare with ROBIN, GaussMarker, exact-inversion approaches, and a synchronization baseline such as Watermark Anything combined with GS?
  2. Can the universal template be estimated by averaging watermarked images? What happens to detection if it is subtracted, and can it be used to forge?
  3. What is the empirical FPR on clean images processed by the full alignment and inversion pipeline?
  4. How do GS and the other base methods perform under the same reconstruction and purification attacks without PreFix?
  5. Does PreFix preserve the undetectability of PRC and the distribution preservation of GS?
  6. Can you report results separately for distortions seen and unseen during optimization, and give PSNR and SSIM for the residual?
8 comments
AuthorsOpenReview ↗

Response to Reviewer iyrn (1/3)

Thank you for the review, and for noting "addresses a known weakness of inversion-based watermarks". Most of your questions became experiments.

A1. Comparison with baselines (W1, Q1).

Tables 1 and 2 compare PreFix with stronger post-hoc methods (VINE-B/R), robust inversion-based watermarks (ROBIN and GaussMarker), and GS with improved inversion (ExactInv [1]). Since Watermark Anything localizes its own watermarked regions rather than geometrically aligning GS images, we evaluate GS-Synchro, which adds a post-hoc synchronization module to GS. We will include these comparisons in the revision.

VINE-B/R show limited robustness. GaussMarker handles rotation well but struggles with other geometric distortions. GS-ExactInv improves inversion but does not explicitly address distortions. GS-Synchro improves geometric robustness.

Unlike standalone synchronization, PreFix jointly optimizes alignment cues and distortion compensation within a unified residual using a watermark recovery objective. Table 7 in the manuscript shows that synchronization (PreFix, K=0) improves geometric bit accuracy (0.502 → 0.923), while per-image optimization further improves both Non-Geo. (0.919 → 0.968) and Geo. (0.923 → 0.944).

Table 1. TPR/Bit Acc. on SD v2.1.

MethodsCleanNon-geo.Geo.FIDCLIP-Score
VINE-B1.000/0.9990.791/0.8400.000/0.51725.16±0.250.3293±0.0006
VINE-R1.000/1.0000.824/0.8710.000/0.50825.19±0.160.3280±0.0006
ROBIN1.000/--0.904/--0.458/--25.88±0.080.3690±0.0003
GaussMarker1.000/1.0000.996/0.9280.286/0.64225.11±0.200.3269±0.0005
GS1.000/1.0000.958/0.9200.000/0.50225.21±0.200.3637±0.0004
GS-ExactInv1.000/1.0000.994/0.9240.000/0.50325.21±0.200.3637±0.0004
GS-Synchro1.000/1.0000.995/0.9120.930/0.80825.21±0.170.3615±0.0006
GS-PreFix1.000/1.0000.998/0.9680.992/0.94425.20±0.220.3637±0.0005

Table 2. Detailed TPR/Bit Acc. on SD v2.1.

DistortionVINE-BVINE-RROBINGaussMarkerGSGS-ExactInvGS-SynchroGS-PreFix
JPEG0.990/0.9331.000/0.9800.990/--1.000/0.9490.948/0.9161.000/0.9571.000/0.9761.000/0.977
GauBlur0.870/0.7910.970/0.8470.970/--1.000/0.8350.956/0.8130.990/0.8031.000/0.8511.000/0.891
MedFilter1.000/0.8701.000/0.8990.780/--1.000/0.9210.902/0.9071.000/0.9091.000/0.9151.000/0.980
GauNoise0.600/0.7580.680/0.8230.780/--0.990/0.8640.943/0.8780.960/0.8531.000/0.9961.000/0.977
S&P Noise0.980/0.9131.000/0.9710.980/--1.000/0.9000.928/0.8801.000/0.9050.990/0.8451.000/0.956
Brightness0.880/0.8960.930/0.9200.730/--0.980/0.9591.000/0.9821.000/0.9690.980/0.8961.000/0.981
Dropout1.000/0.9991.000/1.0001.000/--1.000/1.0000.988/0.9901.000/1.0000.990/0.8770.986/0.990
Cropout0.010/0.5570.010/0.5261.000/--1.000/0.9981.000/0.9951.000/0.9971.000/0.9411.000/0.995
Avg.0.791/0.8400.824/0.8710.904/--0.996/0.9280.958/0.9200.994/0.9240.995/0.9120.998/0.968
Rotation0.000/0.5100.000/0.5050.490/--1.000/1.0000.000/0.5030.000/0.5050.920/0.7910.995/0.985
Scale 0.75x0.000/0.5570.000/0.5400.350/--0.000/0.4960.000/0.5050.000/0.5050.940/0.7830.965/0.945
Scale 1.25x0.000/0.4970.000/0.4900.550/--0.200/0.6000.000/0.5010.000/0.5030.860/0.7951.000/0.938
Translation0.000/0.5070.000/0.5020.900/--0.210/0.5980.000/0.5010.000/0.5010.960/0.8331.000/0.952
Shear0.000/0.5140.000/0.5010.000/--0.020/0.5180.000/0.5020.000/0.5000.970/0.8361.000/0.901
Avg.0.000/0.5170.000/0.5080.458/--0.286/0.6420.000/0.5020.000/0.5030.930/0.8080.992/0.944

A2. Template security (W2, Q2).

Although initialized from a shared template, residuals become image-specific through optimization. We estimate a shared pattern from 1,000 watermarked images by averaging Gaussian high-pass residuals (σ\sigma=2) and removing each channel’s spatial mean. We subtract this pattern from watermarked images for removal or add it to unwatermarked images for forgery. Table 3 shows that watermark detection remains at 100% after removal, while no successful forgeries are observed, supporting robustness against these averaging-based attacks.

Table 3. Averaging-based attacks on SD v2.1.

AttackBit Acc.Detection ratePSNR (dB) ↑SSIM ↑LPIPS ↓
Removal1.000100%44.840.98700.0131
Forgery0.5110%44.850.98680.0143

References

[1] On Exact Inversion of DPM-Solvers.

AuthorsOpenReview ↗

Response to Reviewer iyrn (2/3)

A3. Empirical FPR (W2, Q3).

We evaluated 1,000 unwatermarked images through the full alignment, inversion, and detection pipeline using the same threshold as in our TPR evaluation. No false positives were observed (0/1,000).

A4. Robustness to removal attacks (W2, Q4).

Tables 4 and 5 show that GS-PreFix retains robustness comparable to GS under reconstruction and purification attacks. Purification at s=0.7s=0.7 defeats both methods, but at the cost of severe visual degradation and substantial changes to image content (Appendix D, Fig. 7), with PSNR dropping to 15.69/15.92 dB for GS/GS-PreFix.

We further evaluate both methods on the WARP benchmark [2], including adversarial removal, and under CtrlRegen at strengths 0.1 and 0.5 (Table 6). Both maintain strong robustness across all evaluated settings.

Table 4. Robustness under reconstruction attacks.

MethodMetricChengBmshjVQ-VAEKL-VAESD-v1.4
GSPSNR (dB)34.2635.1229.6929.4126.18
TPR0.9680.9090.9660.9790.965
Bit Acc.0.9240.8650.9230.9340.907
GS-PreFixPSNR (dB)34.2935.3229.7829.9026.41
TPR0.9700.9100.9700.9800.970
Bit Acc.0.9350.8730.9290.9400.915

Table 5. Robustness under purification attacks.

MethodMetrics=0.05s=0.05s=0.1s=0.1s=0.3s=0.3s=0.5s=0.5s=0.7s=0.7
GSPSNR (dB)25.2823.6620.2319.0515.69
TPR0.9780.9760.9400.5080.000
Bit Acc.0.9410.9130.8650.7420.598
GS-PreFixPSNR (dB)25.9324.0820.7219.1015.92
TPR0.9800.9800.9400.5200.000
Bit Acc.0.9440.9120.8730.7430.648

Table 6. Robustness under unseen attacks (TPR/Bit Acc.).

AttackGSGS-PreFix
AdversarialEmbedding1.000/1.0001.000/1.000
AdversarialEmbeddingPSNR1.000/1.0001.000/1.000
Averaging1.000/0.9991.000/0.999
BM3DDenoising1.000/0.9991.000/1.000
BlurDeblurFPNInception1.000/0.9981.000/1.000
DIPAttack1.000/0.9991.000/1.000
DIPAttackNoise1.000/0.9971.000/1.000
DISCOAttack1.000/0.9931.000/0.995
DiffPureAttack1.000/0.9991.000/1.000
DiffusionRegeneration1.000/0.9991.000/1.000
DoGBlurDeblurFPNInception1.000/1.0001.000/1.000
ImageEditingInstructPix2Pix1.000/0.9991.000/1.000
ImageWatermark1.000/1.0001.000/1.000
LIIFAttack1.000/1.0001.000/1.000
LatentFrequencyMasking1.000/1.0001.000/1.000
MPRNetAttack1.000/1.0001.000/1.000
NRPLarge1.000/0.9741.000/0.978
NRPSmall1.000/0.9971.000/0.997
RealESRGANAttack1.000/0.9991.000/0.999
StegastampInversion1.000/0.9991.000/0.999
TrustMarkRM1.000/1.0001.000/1.000
VAERegeneration1.000/0.9491.000/0.962
WMForger1.000/1.0001.000/1.000
WPWMAttacker1.000/0.9981.000/0.999
CtrlRegen (s=0.1s=0.1)1.000/0.9961.000/0.998
CtrlRegen (s=0.5s=0.5)1.000/0.9401.000/0.946

A5. Base-method guarantees (W2, Q5).

PreFix does not strictly guarantee preservation of GS’s distribution-preserving property or PRC’s undetectability. These methods are particularly vulnerable to distortions during image dissemination. PreFix aims to improve distortion robustness while preserving the base methods’ properties where possible, with empirical results showing comparable visual quality and robustness to several attacks (Section 4.2, Table 1 in the manuscript).

A6. Generalization to unseen distortions (W2, Q6).

We report PreFix’s performance separately for seen and unseen distortion types in Table 7 and will clarify this distinction in the revision. Results on unseen distortion settings from WARP (Table 8) further support its generalization.

Table 7. Results on seen and unseen distortions (TPR/Bit Acc. on SD v2.1).

GroupIncluded settingsGSGS-PreFix
Seen Non-Geo.JPEG, GauNoise, MedFilter0.931/0.9001.000/0.978
Seen Geo.Rotation, Scale 0.75×, Scale 1.25×, Translation, Shear0.000/0.5020.992/0.944
UnseenGauBlur, S&P Noise, Brightness, Dropout, Cropout0.974/0.9320.997/0.963

Table 8. Generalization to unseen distortions (TPR/Bit Acc.).

Unseen DistortionGSGS-PreFix
FrequencyMasking1.000/0.9041.000/0.981
DoGBlur1.000/0.9791.000/1.000
CenterCrop0.000/0.5150.831/0.866
RandomCrop0.000/0.5190.820/0.828
Resize0.000/0.5390.932/0.956
InstagramCSSFilters1.000/1.0001.000/1.000
InstagramAttacks1.000/1.0001.000/1.000

References

[2] WARP: A Unified Benchmark for Invisible Image Watermarking-Robustness and Protection Against Attacks.

AuthorsOpenReview ↗

Response to Reviewer iyrn (3/3)

A7. Fidelity (W2, Q6).

Table 9 reports PSNR, SSIM, and LPIPS between images before and after residual addition. These results show a favorable robustness–fidelity trade-off.

Table 9. Robustness–fidelity trade-off (TPR/Bit Acc. on SD v2.1).

MethodPSNR (dB) ↑SSIM ↑LPIPS ↓Non-geo. ↑Geo. ↑
GS------0.958/0.9200.000/0.502
GS-PreFix37.81 ± 1.120.9694 ± 0.01490.0261 ± 0.01100.998/0.9680.992/0.944

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.