acceptodds
Under review as a conference paper at ICLR 2027

WaterVIB: Learning Minimal Sufficient Watermark Representations via Variational Information Bottleneck

Abstract

Robust invisible watermarking is essential for copyright protection, yet generative editing severely threatens watermark survival by rewriting host-image details. Our quantitative analyses reveal that existing watermark representations become deeply entangled with host-speciffc textures, causing generative modiffcations to disrupt message decoding. To break this entanglement, we formulate watermark extraction as learning a minimal sufffcient statistic for message recovery, thereby substantially reducing the decoder’s reliance on fragile host textures. Guided by this insight, we propose WaterVIB, a plug-in variational bottleneck module that decouples message representations from host-image details across existing backbones. Systematic evaluation across ffve 2D backbones with different architectures under 32 attacks conffrms widespread robustness gains thanks to our WaterVIB. On VINE-R, WaterVIB reduces the bit error rate under global puriffcation by 71.0% while increasing embedding PSNR. Furthermore, extending WaterVIB to NeRF-Signature enhances 3D rendering ffdelity while preserving clean decoding performance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
deep image watermarking, robust watermarking, variational information bottleneck, generative purification, AIGC 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_SWhnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the timely problem of watermark robustness against generative image editing. From an information bottleneck perspective, the authors propose WaterVIB, a plug-in module that reduces dependence on host-image information while preserving watermark messages. Experiments across multiple watermarking frameworks demonstrate encouraging improvements.

Critical strengths

Clear motivation and insightful analysis. The authors identify an important problem through well-designed empirical analyses and provide a principled information-theoretic formulation.

Practical significance and applicability. Robustness against generative editing is increasingly important. The proposed plug-in module is easy to integrate into existing watermarking frameworks and demonstrates promising improvements.

Critical weaknesses
  1. The paper argues that host-image dependence weakens watermark robustness (Section 3), but does not directly examine whether WaterVIB also reduces host dependence in watermark embedding. A useful experiment would be to embed the same message into two different host images, extract their watermark residuals, and swap these residuals between the two hosts. The authors should compare decoding performance before and after swapping, both with and without generative editing, using the original backbone and WaterVIB. If WaterVIB also reduces embedding-level host dependence, it should exhibit better cross-host transferability. This would help clarify the scope of the proposed mechanism.

  2. Table 1 analyzes only HiDDeN and EditGuard, with HiDDeN being a relatively early watermarking framework. Although Table 3 evaluates five backbones, the proposed mechanism may behave differently on more recent architectures. Additional experiments on CIN [1], Lightweight-Mark [2], and VideoSeal [3] would strengthen the plug-and-play claim.

  3. Diffusion-based editing introduces distortions through both VAE reconstruction and the diffusion noising-denoising process. However, the main global purification experiments use an extremely low strength of 0.002 (Appendix G.1). Table 3 also reports nearly identical BER under SD1.5 purification (48.983%) and VAE reconstruction (48.979%) for EditGuard. This raises concerns that the improvements primarily reflect VAE reconstruction robustness rather than robustness to diffusion denoising. Appendix G.3 further shows that the gains become substantially smaller under stronger purification.

  4. Despite focusing on generative-editing robustness, the paper lacks direct comparisons with specialized diffusion-erasure-robust watermarking methods, such as DERO [4]. Such comparisons are necessary to better establish the competitiveness of WaterVIB.

[1] Ma et al. "Towards Blind Watermarking: Combining Invertible and Non-invertible Mechanisms" (CIN). 2022. https://arxiv.org/abs/2212.12678

[2] Qiu et al. "Lightweight-Mark: Rethinking Deep Learning-Based Watermarking." ICML, 2025. https://proceedings.mlr.press/v267/qiu25b.html

[3] Fernandez et al. "Video Seal: Open and Efficient Video Watermarking." 2024. https://arxiv.org/abs/2412.09492

[4] Fang et al. "DERO: Diffusion-Model-Erasure Robust Watermarking." ACM Multimedia, 2024. https://doi.org/10.1145/3664647.3681220

Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer SWhn

Thank you for recognizing the practical motivation and modular design. Following your suggestion, we swapped watermarks between images to test how much an embedded watermark depends on its original image.

Q1/W1 Swapping watermarks across images

Both models read a watermark more accurately on the image it was written into. After VAE reconstruction, WaterVIB recovers more complete messages on that image; after a swap, neither model consistently recovers more. Holding the watermark fixed isolates the effect of texture: amplifying texture causes a smaller BER increase with WaterVIB.

Table 1. 32 pairs, same 100-bit message within each pair, residuals swapped both ways, editing noise fixed. Each cell: BER (%) / exact messages out of 64.

EditOriginal: ownOriginal: swapWaterVIB: ownWaterVIB: swap
Clean0.016 / 632.766 / 150.031 / 622.781 / 18
VAE1.078 / 326.875 / 40.344 / 526.406 / 3
SD1.5 0.1510.109 / 214.844 / 08.984 / 015.203 / 0
SD1.5 0.3022.234 / 024.656 / 021.516 / 024.797 / 0

Table 1b. Texture intervention with the watermark fixed. A positive difference means a smaller BER penalty for WaterVIB.

Edit / texture amplitudePaired imagesBER penalty reduction (pp ± sample variance)
VAE / ×1.532 pairs1.80 ± 3.92

Q2/W2 CIN and VideoSeal

With the same decoder, adding KL lets CIN recover more complete messages after VAE reconstruction and milder diffusion. VideoSeal also makes fewer bit errors and detects more registered watermarks.

Table 2. Released CIN (30 bits) and VideoSeal (256 bits). Control: same Gaussian head with KL off. BER and exact recovery are three-seed means (%) ± sample variances (pp²). VideoSeal uses the same detection threshold for both arms.

Model / editMetricNo KLWith KL
CIN / VAEBER (%)4.748 ± 0.00544.125 ± 0.0023
Exact recovery (%)47.92 ± 1.852.93 ± 0.11
CIN / SD1.5 0.15BER (%)15.284 ± 0.02414.575 ± 0.019
Exact recovery (%)10.22 ± 0.1711.20 ± 0.089
CIN / SD1.5 0.50BER (%)39.123 ± 0.07839.054 ± 0.012
Exact recovery (%)0.00 ± 00.00 ± 0
VideoSeal / SD1.5 0.15BER (%)34.980 ± 0.1134.602 ± 0.069
Exact recovery (%)0.00 ± 00.00 ± 0
Known-key detection (%)48.3152.08

Q3/W3 VAE reconstruction and diffusion denoising

At the original strength of 0.002, the image change comes mainly from VAE reconstruction. WaterVIB also lowers BER on both backbones at all three multistep SD1.5 strengths and under CtrlRegen+. Under CtrlRegen, BER falls on VINE-R and rises slightly on EditGuard.

Table 3. Frozen original and submitted systems read their own 100-bit messages. 128 images per backbone, embedding MSE matched per image. Submitted strength: 0.002. The three SD1.5 settings below use 8, 16 and 26 denoiser calls.

Backbone / attackOriginal BER (%)VIB BER (%)Original exact (%)VIB exact (%)
VINE-R / VAE2.0230.57039.8471.88
VINE-R / SD1.5 0.1512.0168.6950.000.00
VINE-R / SD1.5 0.3023.32821.6020.000.00
VINE-R / SD1.5 0.5034.39132.9380.000.00
VINE-R / CtrlRegen+ 0.3019.71117.4450.000.00
VINE-R / CtrlRegen48.02346.2730.000.00
EditGuard / VAE46.61729.6330.000.00
EditGuard / SD1.5 0.1550.14141.4840.000.00
EditGuard / SD1.5 0.3050.33645.8590.000.00
EditGuard / SD1.5 0.5050.19547.0550.000.00
EditGuard / CtrlRegen+ 0.3049.10246.6090.000.00
EditGuard / CtrlRegen49.80550.5230.000.00

Q4/W4 Comparison with DERO

We could not find an official release of DERO’s code or model weights, so we did not include a direct comparison.

Reviewer_PrW5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes WaterVIB, a plug-in variational information bottleneck module designed to improve watermark robustness against generative editing. By preserving message information while suppressing host-dependent representations, the method improves watermark recovery across five 2D backbones and is further evaluated on NeRF watermarking. The paper also provides theoretical motivation and extensive representation-level analyses.

Critical strengths
  1. The core method largely adapts established variational information bottleneck techniques to watermark decoding. Although a capacity-matched KL ablation is provided for HiDDeN, comparable controls are missing for the stronger backbones. It remains unclear how much of the improvement comes specifically from information bottleneck regularization rather than the modified decoder architecture or stochastic regularization.

  2. The main global purification experiments use a very low strength of 0.002 for several diffusion models. Appendix G.3 shows that the improvements become substantially smaller under stronger purification. For instance, on VINE-R with SD1.5 at strength 0.3, BER decreases only from 21.598% to 20.941%. Moreover, the JPEG evaluation primarily reports Q70, without standard stronger compression tests such as Q50. These limitations weaken the generality of the robustness claims.

Critical weaknesses

Can the authors compare WaterVIB against parameter-matched deterministic bottlenecks and alternative regularizers on VINE-R and EditGuard to isolate the contribution of the variational information bottleneck?

Can the authors provide more comprehensive results under meaningful purification strengths and JPEG Q50/Q70/Q90, ideally with matched image-quality constraints?

Since inference uses the posterior mean rather than the stochastic representation, how directly do the proposed information-theoretic guarantees apply to the deployed deterministic decoder?

Questions for authors

Can the authors compare WaterVIB against parameter-matched deterministic bottlenecks and alternative regularizers on VINE-R and EditGuard to isolate the contribution of the variational information bottleneck?

Can the authors provide more comprehensive results under meaningful purification strengths and JPEG Q50/Q70/Q90, ideally with matched image-quality constraints?

Since inference uses the posterior mean rather than the stochastic representation, how directly do the proposed information-theoretic guarantees apply to the deployed deterministic decoder?

1 comment
AuthorsOpenReview ↗

Response to Reviewer PrW5

Thank you for noting the clarity of the paper and for suggesting controls that distinguish architecture, noise and KL. We address the KL contribution, recovery under stronger attacks and the guarantee for mean decoding below.

Q1/W1 Contribution of the full KL term

At equal parameter count, full KL gives VINE-R the fewest bit errors and most complete messages in every test below. EditGuard recovers more complete messages on clean images and makes fewer bit errors after editing.

Table 1a. BER (%). Revision ablation: residual decoders start from common official weights and train for 500 updates with a fixed writer. 512 test images, three seeds. Mean ± sample variance (pp²).

Backbone / conditionD: deterministicN: noise onlyL2: noise + mean L2V: noise + full KL
VINE-R Clean0.714 ± 0.000370.603 ± 0.000390.594 ± 0.0000460.379 ± 0.000046
VINE-R Watercolor4.186 ± 0.000804.077 ± 0.000444.072 ± 0.000843.815 ± 0.0029
VINE-R Warm light9.359 ± 0.00609.322 ± 0.0129.294 ± 0.0129.072 ± 0.0043
VINE-R JPEG Q500.648 ± 0.000700.542 ± 0.0000620.518 ± 0.00000510.330 ± 0.000072
EditGuard Clean0.350 ± 0.000330.328 ± 0.000320.331 ± 0.000370.139 ± 0.00086
EditGuard Watercolor44.390 ± 0.006744.146 ± 0.03644.587 ± 0.01543.519 ± 0.052
EditGuard Warm light46.348 ± 0.0008646.230 ± 0.001746.480 ± 0.001845.640 ± 0.010
EditGuard JPEG Q5021.266 ± 0.06820.877 ± 0.05022.107 ± 0.02816.769 ± 0.27

Table 1b. Exact 100-bit recovery (%) in the same experiment, mean ± sample variance.

Backbone / conditionD: deterministicN: noise onlyL2: noise + mean L2V: noise + full KL
VINE-R Clean64.71 ± 1.2770.64 ± 3.1070.96 ± 0.3681.84 ± 0.50
VINE-R Watercolor26.17 ± 0.1129.82 ± 0.05130.34 ± 0.01334.83 ± 0.089
VINE-R Warm light16.80 ± 0.6118.42 ± 0.6618.68 ± 0.2422.72 ± 0.36
VINE-R JPEG Q5065.89 ± 5.5471.09 ± 0.5072.27 ± 1.1883.07 ± 0.051
EditGuard Clean79.88 ± 1.3780.92 ± 2.0083.14 ± 1.5492.51 ± 6.15
EditGuard Watercolor0.00 ± 0.000.00 ± 0.000.00 ± 0.000.00 ± 0.00
EditGuard Warm light0.00 ± 0.000.00 ± 0.000.00 ± 0.000.00 ± 0.00
EditGuard JPEG Q500.00 ± 0.000.00 ± 0.000.00 ± 0.000.00 ± 0.00

Q2/W2 JPEG compression and diffusion denoising

At matched embedding MSE, the submitted writer and reader together (V/V) give lower BER than the official pair (O/O) at every JPEG quality and both diffusion strengths. To compare readers on identical attacked pixels, compare O/O with O/V or V/O with V/V.

Table 2. Frozen official (O) and submitted (V) checkpoints. 128 images per backbone, embedding MSE matched per image. Cells give BER / exact 100-bit recovery (%). Columns identify writer/reader.

Backbone (embedding PSNR)AttackO/OO/VV/OV/V
VINE-R (36.730 dB)JPEG Q500.070/93.750.266/78.120.086/91.410.008/99.22
JPEG Q700.047/95.310.203/81.250.047/95.310.016/98.44
JPEG Q900.039/96.090.195/82.030.047/95.310.008/99.22
SD1.5 0.3023.328/0.0025.203/0.0021.188/0.0021.602/0.00
SD1.5 0.5034.391/0.0035.383/0.0033.234/0.0032.938/0.00
EditGuard (49.687 dB)JPEG Q5015.969/0.0049.148/0.0049.930/0.0010.711/0.00
JPEG Q705.016/3.9148.305/0.0050.062/0.004.117/21.88
JPEG Q900.727/81.2547.688/0.0050.117/0.000.398/88.28
SD1.5 0.3050.336/0.0049.711/0.0050.219/0.0045.859/0.00
SD1.5 0.5050.195/0.0049.508/0.0049.758/0.0047.055/0.00

Q3/W3 Guarantees for mean decoding

With an affine final layer, the score at the mean equals the average score under sampling. By convexity of cross-entropy, decoding the mean has no higher loss than the expected loss under sampling. Table 3 measures both on the submitted VINE-R model. If the mean also determines the noise scale, it retains at least as much message information as a sample. All three residual VINE-R models with full KL satisfy this condition. Appendix O.2 tests how much input information the NeRF features retain beyond the target bit.

Table 3. Measured cross-entropy in the submitted VINE-R model (bits per 100-bit message, lower is better).

ReadoutCleanJPEG Q50VAESD1.5 .15SD1.5 .30
Mean0.0420.0501.93537.09570.079
Sampled (expected)0.0430.0511.94437.25670.650
Reviewer_o1SqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the vulnerability of deep image watermarks to generative editing. It argues that watermark and editing residuals exhibit spectral and spatial alignment and that editing perturbations interfere with the decoding objective. Based on this motivation, the authors formulate watermark extraction as learning a minimal sufficient statistic and derive an information-bottleneck objective that seeks to preserve message information while compressing information about the attacked image. The proposed WaterVIB module inserts a stochastic Gaussian bottleneck between the feature extractor and message decoder. It is integrated into five 2D watermarking backbones and NeRF-Signature and evaluated under 32 attack conditions. The paper reports substantial BER reductions under several reconstruction and purification attacks, including a 71.0% relative reduction for VINE-R under SD1.5 purification. It further presents crossed host-message experiments, variance decomposition, mutual-information estimators, and gradient diagnostics to support the claimed mechanism.

Critical strengths
  1. WaterVIB is a broadly applicable module, evaluated across five diverse 2D watermarking architectures and one NeRF-based method under 32 attack conditions.

  2. The method achieves meaningful BER reductions under several reconstruction and purification attacks, generally without a clear loss of embedding fidelity.

Critical weaknesses
  1. The central minimal-sufficient-statistic interpretation is not established for the implemented model. Theorem 3.3 in Section 3.3, with its proof in Appendix A, applies only to deterministic statistics on finite domains. In contrast, WaterVIB uses a continuous stochastic representation during training and the deterministic posterior mean at inference, as described in Section 4.2. The paper provides no result connecting these settings. It also neither measures the sufficiency tolerance in Eq. (5) nor verifies the conditions required to transform Eq. (5) into the information-bottleneck objective in Eq. (6), whose supporting-multiplier argument appears in Appendix B.2.

  2. The KL regularizer does not exactly match the implemented stochastic channel. The compression term is introduced in Eq. (8) of Section 4.1, whereas the actual sampling rule includes an additional fixed noise scale in Eq. (9) of Section 4.2. Appendix C.1 attempts to connect them through the bound in Eq. (34), but under the reported dimensionality and noise scale, this bound is dominated by a very large dimension-dependent term and is therefore weak. Moreover, it concerns the sampled representation, while inference uses the posterior mean.

  3. The proposed host-texture mechanism is not causally established. Section 3.1 and Table 1 report spectral-energy overlap and correlations with the host image, but these statistics do not establish alignment in phase, spatial location, or perturbation direction. Likewise, the first-order analysis in Section 3.2, particularly Eq. (2) and Table 2, largely confirms that loss-increasing perturbations project positively onto the loss gradient. No controlled intervention varies host texture while fixing the message and attack strength.

  4. The experiments do not isolate the contribution of information compression. WaterVIB replaces the original prediction head, changes its capacity, adds stochastic noise, and retrains the full system. The capacity-matched zero-KL comparison in Figure 5(c) of Section 5.3 is provided only for HiDDeN. Comparable controls are absent for the other backbones, leaving architectural changes, stochastic regularization, and end-to-end retraining as alternative explanations.

  5. The robustness claim relies heavily on weak and incompletely calibrated attacks. The headline 71.0% reduction in Section 5.2 uses a very low diffusion strength. The stronger purification sweep in Appendix G.3 and Table 11 shows that the advantage becomes small at strength 0.3 and nearly disappears at strength 0.5. Table 3 also lacks matched perceptual-severity measurements, making comparisons across different purification pipelines difficult.

  6. The representation analysis does not reliably establish increased message information or reduced host information. Section 5.3 interprets differences between separately trained mutual-information estimators as differences in true mutual information, although Appendix J.1 and J.4 acknowledge that these lower bounds need not preserve the ordering of the underlying quantities. The compared head representations also have different dimensions, while the KSG and bootstrap results in Appendix J.5 are not uniformly consistent with reduced host dependence.

Questions for authors
  1. Can the authors provide a theoretical result that directly covers the continuous Gaussian training representation and the deterministic posterior mean used at inference? What value of the approximate-sufficiency tolerance is achieved by the trained models

  2. Can the authors directly intervene on host texture while holding the embedded message fixed? Direct residual cross-spectra, spatial overlap measurements, and controlled texture perturbations would provide stronger evidence for the proposed causal mechanism.

  3. How do the authors reconcile the claimed increase in message information and decrease in host information with the non-orderability of separately fitted DV bounds, mismatched feature dimensions, increased Texture-KSG estimates under some conditions, and negative finite-sample Ross estimates?

  4. For the NeRF study, can the authors report per-scene results, view counts, independent training seeds, and confidence intervals, and evaluate attacks acting on the NeRF representation or rendering process rather than only post-render 2D distortions?

  5. Why does the training objective regularize the unscaled Gaussian posterior rather than the alpha-scaled distribution actually used for sampling? Can the authors report the actual-channel KL, the dimension-dependent correction term, and the aggregated-posterior gap for each backbone?

2 comments
AuthorsOpenReview ↗

Response to Reviewer o1Sq — Part I

Thank you for recognizing the recovery gains and for your questions about the theory.

Q1/W1 Message information retained

Let S be the sampled feature or its mean. Decoding cross-entropy bounds I(M;X | S), the message information lost from image X. With an affine final layer, convexity bounds the loss at the mean by the expected sampled loss.

Table 1a. Information-loss bound (bits/100-bit message). Submitted VINE-R, 128 images × 2 messages.

ReadoutCleanJPEG Q50VAESD1.5 .15SD1.5 .30
Mean0.0420.0501.93537.09570.079
Sampled (expected)0.0430.0511.94437.25670.650

On clean NeRF scenes, KL keeps information about whether the target bit is 0 or 1, while reducing information about its source message and location. The step from Eq. (5) to (6) requires an interior information target and probabilistic mixing of encoders. Appendix B.2 proves the positive multiplier.

Table 1b. NeRF single-bit test, 16 new messages and 32 spatial blocks per message.

Target-bit loss bound with KLScene/seed pairs with less source information
< 0.00000145 bit24/24 (8 × 3)

Q5/W2 Sampling noise and KL

We use predicted standard deviation σ for KL and ασ for sampling to adjust training noise separately. Table 2 gives their exact conversion. Only the variance correction changes the gradient. Sampled KL minus the last column gives information about which input produced the feature.

Table 2. Sampled KL = training KL + both corrections. Gap: pooled features versus reference Gaussian. Nats per representation. CRMark: Gaussian branch.

Model (dim., α)Training KLFixedVarianceSampled KLPrior gap
VINE-R (1024, 1)3769.890.000.003769.893764.35
EditGuard (16384, 0.00005)38784.12162258.74-8191.41192851.45192845.90
HiDDeN (128, 0.007)99.15635.12-163.94570.33564.78
CRMark (256, 0.05)64.04766.91-114.41716.53710.99
Robust-Wide (64, 0.002)6302.84397.73-6485.89214.69209.14
NeRF (4096, 0.007)9.0320323.72-2047.5118285.2418281.24

Q2/W3 Texture amplification and bit errors

We amplified texture while keeping the watermark, message, low-frequency image and editing noise fixed. Both models made more bit errors, but WaterVIB’s increase was smaller. The watermark and edit often change the same frequency bands, yet their signed directions align weakly. Their spatial overlap exceeds the shifted control.

Table 3a. VINE-R, 128 new images, texture ×1.5. BER increase (pp). Reduction: paired-image mean ± sample variance (pp²).

InputOriginalWaterVIBReduction (mean ± variance)
Clean0.380.090.30 ± 0.34
VAE3.771.871.91 ± 6.87
SD1.5 .155.214.250.96 ± 13.41

Table 3b. SD1.5 0.15. Spatial overlap uses the top 10% of pixels by energy. Shifting preserves Fourier magnitudes.

MeasureOriginalWaterVIB
Spectral power overlap.688.685
Signed cross-spectral cosine.006.030
Spatial overlap (observed / shifted).168 / .094.169 / .095

Q3/W6 Mutual information in decoded bits

After VAE, decoded bits carry more message information and the shared 64-D layer has lower estimated texture MI. At unequal-width heads, texture estimates rise on clean and blurred images. Separate DV bounds cannot rank true MI. Negative Ross estimates on zero-MI controls reflect finite-sample error.

Table 4. HiDDeN, new images and messages for decoded-bit MI. Texture information is estimated by KSG.

MeasurementInputOriginalWaterVIB
Decoded message MI (bit/bit)VAE.019.206
Texture KSG, same 64-D featuresVAE.690.339
Texture KSG, 30/128-D headsClean.134.188
Texture KSG, 30/128-D headsBlur.138.260
AuthorsOpenReview ↗

Response to Reviewer o1Sq — Part II

W4 Parameter-matched KL ablations

At equal capacity and training steps, full KL gives VINE-R the lowest BER and highest exact recovery. EditGuard also makes fewer bit errors after editing. N/L2/V share the writer, head and images. CIN and VideoSeal improve. Robust-Wide and CRMark show small VAE differences with their native heads.

Table 5a. Residual decoders: 500 updates, 512 images, three seeds. BER / exact (%). D deterministic, N noise only, L2 noise + mean L2, V noise + full KL.

Backbone / inputDNL2V
VINE-R / Clean0.714/64.710.603/70.640.594/70.960.379/81.84
VINE-R / Watercolor4.186/26.174.077/29.824.072/30.343.815/34.83
VINE-R / JPEG Q500.648/65.890.542/71.090.518/72.270.330/83.07
EditGuard / Clean0.350/79.880.328/80.920.331/83.140.139/92.51
EditGuard / Watercolor44.390/0.0044.146/0.0044.587/0.0043.519/0.00
EditGuard / JPEG Q5021.266/0.0020.877/0.0022.107/0.0016.769/0.00

Table 5b. Other backbones: matched writer and readout, KL off/on.

Backbone / inputBitsBER N/V (%)Exact N/V (%)
CIN / VAE304.748/4.12547.92/52.93
VideoSeal / SD1.5 .1525634.980/34.6020/0
Robust-Wide / VAE640.24/0.2488.80/88.28
CRMark / VAE640.78/0.8168.49/67.19

W5 Watermark recovery after diffusion editing

At matched editing distortion, WaterVIB lowers BER on both backbones under SDXL, including on new images and messages. It also lowers BER in most SD1.5 tests.

Table 6a. Submitted VINE-R, SD1.5, BER (%). At strength 0.002, image changes mainly come from VAE reconstruction.

StrengthOriginalWaterVIB
0.0025.0791.475
0.3021.59820.941
0.5033.41633.300

Table 6b. Frozen official → submitted WaterVIB. LPIPS measures the change caused by editing. 128 images/set. Embedding matched by MSE (original set) or LPIPS (new set). Detection: ≥74/100 matches. Attack rows give BER (%), last rows exact recovery and paired resampling.

Model / diffusionLPIPSOriginal imagesNew imagesDetections on original set /128
VINE-R / SD1.50.2017.98 → 15.3416.97 → 15.48108 → 117
VINE-R / SDXL0.2021.20 → 15.8619.95 → 16.6698 → 123
VINE-R / SD1.50.3029.24 → 27.6027.92 → 28.5244 → 60
VINE-R / SDXL0.3027.57 → 24.0626.62 → 24.5359 → 83
EditGuard / SD1.50.2049.20 → 44.1249.89 → 49.140 → 2
EditGuard / SDXL0.2048.83 → 41.9050.47 → 47.410 → 2
EditGuard / SD1.50.3050.24 → 47.4650.20 → 49.380 → 0
EditGuard / SDXL0.3049.57 → 44.6450.05 → 48.390 → 1
Exact recovery (all settings)—0% → 0%0% → 0%—
Settings supporting lower BER—8/85/8—

Q4 Novel-view quality gains and quantization errors

Mean novel-view PSNR is higher with WaterVIB in every tested scene. At 8, 6 and 4 bits, full KL gives lower BER than noise alone. At a ray step of 1/128, all three readers have similar BER.

Table 7a. Blender, original watermark strength. Field and ray-step attacks. BER / exact recovery (%), 32 new 32-bit messages per scene/seed. BER reduction: seed mean ± variance (pp, pp²).

Field / rendererOriginalNoise onlyFull KLBER reduction vs noise only
Clean0.76/83.334.03/75.780.00/100.004.03 ± 7.47
8-bit2.32/56.128.15/41.023.49/55.474.66 ± 15.13
6-bit24.70/0.0034.26/0.5228.45/0.135.81 ± 1.97
4-bit49.43/0.0050.17/0.0049.14/0.001.03 ± 0.06
Ray 1/12847.68/0.0048.28/0.0048.51/0.00-0.24 ± 0.36

Table 7b. PSNR gain (mean ± variance, dB and dB²). Blender: three seeds, 16 views/scene. LLFF: variance across views, one archived pair/scene. Appendix O gives 95% intervals.

Blender (16 views)Gain ± varianceLLFFViewsGain (mean ± variance)
Chair0.43 ± 2.03Fern32.511 ± 0.0078
Drums1.55 ± 0.81Flower53.498 ± 0.00086
Ficus1.70 ± 3.29Fortress62.617 ± 0.075
Hotdog1.71 ± 1.16Horns82.457 ± 0.091
Lego1.02 ± 2.49Leaves41.931 ± 0.034
Materials1.10 ± 20.24Orchids41.634 ± 0.14
Mic2.54 ± 1.43Room64.487 ± 0.31
Ship0.63 ± 19.09Trex75.487 ± 0.027

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.