acceptodds
Under review as a conference paper at ICLR 2027

ZeroStep: Rethinking Diffusion Inversion for Watermark Verification

Abstract

Embedding watermarks into the initial noise provides a natural mechanism for copyright protection and source attribution in diffusion generation. Yet existing verification methods still invert the generated image back to the initial watermarked noise, making costly diffusion inversion a standard step in watermark recovery. We revisit this requirement by tracing watermark evidence along the generation trajectory and find that, although similarity to the initial watermarked noise drops rapidly, watermark recoverability remains high even near the generation endpoint. This observation motivates ZeroStep, a zero-step watermark verification method that removes diffusion inversion entirely. Our method trains a lightweight LoRA adapter on the VAE encoder with trajectory supervision across timesteps and latent components, enabling direct watermark extraction without diffusion inversion. For robustness during online image distribution, we further leverage paired clean and distorted views to guide distortion-robust training. Experiments across diffusion and flow-matching backbones show that our method consistently outperforms state-of-the-art inversion-based methods. Notably, on the challenging SD v3.5 backbone, it improves watermark bit accuracy from 95.88% to 99.11% while reducing the deployed model size from 2.5B to 35.01M parameters. Moreover, the learned adapter can be reused across compatible model versions and scales, demonstrating strong cross-model adaptability. Code is available at https://anonymous.4open.science/r/ZeroStep-Release.

Primary area:
generative models
Keywords:
Diffusion Watermarking, Zero-Step Extraction, Copyright Protection, Source Attribution, Generative Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WdZSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper rethinks the verification process for initial-noise watermarks in diffusion models, breaking away from the traditional paradigm that relies on costly diffusion inversion. The authors discovered that, although the similarity between intermediate latent representations and the initial noise decays rapidly, the watermark information remains highly preserved throughout the generation trajectory and even within the final latent representation. Based on this insight, the proposed ZeroStep framework, which involves fine-tuning only a lightweight VAE Encoder LoRA and incorporates trajectory distillation and distortion-resilient training. Experimental results demonstrate that this method significantly improves extraction accuracy and robustness against distortions while reducing computational latency and exhibiting strong cross-model transferability.

Critical strengths

The proposed inversion-free watermark extraction scheme reduces verification latency to 19.0 ms and model parameters from 2.5B to 35.01M.

Critical weaknesses
  1. While the trajectory distillation loss in Eq. (7) constrains clean image predictions to match sampled intermediate trajectory states ztwz_t^w, the distortion-robust loss in Eq. (8) rigidly anchors distorted image predictions to the initial watermarked noise zTwz_T^w. This design choice lacking theoretical justification.

  2. The threat model is overly idealized, as it fails to consider critical security threats such as regeneration attacks (e.g., diffusion regeneration, vae regeneration) and adversarial perturbations.

  3. The scope of validation is somewhat limited, as both the main text and appendix evaluate only four initial noise watermarking schemes (Tree-Ring, Gaussian Shading, RingID, and PRC).

  4. The learned adapter is strictly coupled to a specific VAE architecture and latent dimensionality. As admitted in Appendix C.2, direct transfer fails across different VAE families.

Questions for authors
  1. In Section 3, the authors present the paper's core motivation: watermark recoverability remains high near the generation endpoint despite a rapid drop in latent similarity to the initial noise. However, this motivating experiment is restricted solely to Tree-Ring. Can this analysis be generalized to other initial noise watermarking schemes?

  2. While the trajectory distillation loss in Eq. (7) constrains clean image predictions to match sampled intermediate trajectory states ztwz_t^w, the distortion-robust loss in Eq. (8) rigidly anchors distorted image predictions to the initial watermarked noise zTwz_T^w. Could you provide some theoretical explanations?

  3. The authors introduce 1−Ct1 - C_t into the weighting function wt=Rt(1−Ct)w_t = R_t(1 - C_t) to assign higher supervision weights to states farther from the initial noise. There is a lack of theoretical explanation for why states closer to the original watermarked noise zTwz_T^w should be down-weighted during trajectory distillation.

  4. Figure 5 lacks essential experimental details in both the main text and appendix. Specifically, the metric "Correct evidence retained" is left undefined, and the exact setup (model backbone, watermark method, and sample size) used to generate the plot is not specified.

  5. Comparing Frozen VAE with ZeroStep under image corruptions is unfair because ZeroStep benefits from data augmentation (Eq. 8), which is not the paper's core contribution. To evaluate the true impact of VAE adaptation, ZeroStep without data augmentation (the "No Noise" setting) should be compared against Frozen VAE.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We thank the reviewer for the insightful and detailed comments. We address each point below.

Re: Generality of Motivation and Trajectory Weighting (Q1 & Q3)

Generality of motivation. Our Tree-Ring analysis shows that latent similarity to the initial watermarked noise decreases along the generation trajectory, while watermark evidence remains recoverable. We extend this analysis to GS and PRC, measuring latent cosine similarity (CtC_t) and watermark recoverability (RtR_t) at intermediate and final stages using their original verifiers.

Watermark recoverability along the generation trajectory.

WatermarkMetricCT/2C_{T/2} ↓RT/2R_{T/2} ↑C0C_0 ↓R0R_0 ↑
Tree-RingTPR@10−310^{-3}0.870.970.350.92
GSBit Acc.0.920.970.380.89
PRCBit Acc.0.900.980.330.93

GS and PRC exhibit the same trend, confirming that this behavior is not specific to Tree-Ring. We have added the corresponding visualizations to Figure 8 of the revised manuscript.

Motivation for trajectory weighting. This observation also motivates our timestep weighting strategy. Our goal is to learn verifier compatible watermark representations rather than reconstruct the initial noise. For normalized latent vectors, 1−Ct1-C_t reflects the deviation from the initial watermarked noise, as ∥zˉt−zˉTw∥22=2(1−Ct)\|\bar z_t-\bar z_T^w\|_2^2=2(1-C_t). States with high similarity retain watermark evidence but provide supervision largely redundant with initial noise regression. We therefore use 1−Ct1-C_t to favor more distinct trajectory states, while RtR_t ensures that their watermark evidence remains recoverable. Accordingly, we define the timestep weight as wt=Rt(1−Ct)w_t=R_t(1-C_t), balancing supervision distinctiveness and watermark recoverability.

Re: Watermark Trajectory Distillation (W1 & Q2)

Rationale for Distortion Supervision. We thank the reviewer for raising this point. Eqs. (7) and (8) use different supervision targets for complementary objectives. For clean images, intermediate states ztwz_t^w retain sufficient watermark evidence and provide suitable supervision. For distorted images, we use the initial watermarked noise zTwz_T^w as a shared alignment target across different views of the same watermark instance. As a sample-specific reference determined by the watermark embedding process, it is independent of timestep sampling and provides consistent alignment across distorted views. This constraint serves as an auxiliary regularizer rather than requiring exact reconstruction of the initial noise. To validate this design, we compare different distortion supervision targets.

Ablation of distortion supervision targets.

VariantAlignment TargetTR TPR (Clean) ↑TR TPR (Avg.) ↑GS Acc. (Clean) ↑GS Acc. (Avg.) ↑
Trajectory-state anchorztwz_t^w1.0000.9981.00000.9965
Initial-noise anchor (Ours)zTwz_T^w1.0000.9991.00000.9983
AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Re: Robustness and Security Evaluation (W2)

Additional attacks and security robustness. We further evaluate robustness under true crop and resize, as well as compound distortions, to assess generalization beyond the original training distribution. We also include three representative regeneration attacks: Regen-Diff, Rinse-2xDiff, and CtrlRegen. All methods are evaluated on the same attacked images using the original GS verifier.

Robustness under compound distortions and regeneration attacks on SD v2.1 with Gaussian Shading.

AttackDDIM-50DDIM-1FARIZeroStep
R.Crop + Resize0.95220.84790.95830.9997
Resize + JPEG0.98180.96430.99210.9991
G.Blur + M.Blur0.98130.94810.99350.9995
Regen-Diff0.94860.83000.95470.9671
Rinse-2xDiff0.92070.79740.94110.9584
CtrlRegen0.89530.72510.90920.9204

ZeroStep achieves the highest bit accuracy across all evaluated attacks. However, the ability to resist attacks remains largely dependent on the underlying design of the base watermarking methods, as Zerostep serves as a plug-and-play approach to enhance their detection performance.

Re: Generalization to Additional Watermark Schemes (W3)

We thank the reviewer for the suggestion. To further evaluate the applicability of ZeroStep to different watermark mechanisms, we extend our experiments on SD v2.1 to two additional schemes: WIND and Spherical Watermark.

Top-1 identification accuracy of WIND (100,000 candidate keys across 2,048 groups).

MethodNFECleanAvg.JPEGR.CropR.DropResizeG.BlurM.BlurG.NoiseS&PBright
DDIM-50500.99900.96400.99300.91700.92900.99100.99200.99100.95300.94700.9630
DDIM-110.98400.90220.94900.81300.83700.93200.93600.91900.89200.88700.9548
FARI11.00000.98730.99900.97000.97300.99900.99900.99500.98200.97800.9907
ZeroStep01.00000.99271.00000.98200.98501.00001.00000.99800.98700.98500.9973

ZeroStep achieves the highest average identification accuracy of 0.9927, outperforming FARI while requiring zero denoiser NFE.

Bit accuracy of Spherical Watermark (512-bit payload).

MethodNFECleanAvg.JPEGR.CropR.DropResizeG.BlurM.BlurG.NoiseS&PBright
DDIM-50500.99990.96290.98290.94870.93590.99870.99340.99180.93210.92370.9592
DDIM-110.99140.92140.96380.87500.86420.97840.97410.95930.86820.85960.9500
FARI10.99990.97850.99580.95720.94360.99870.99750.99540.96290.95880.9966
ZeroStep00.99990.99030.99920.97910.97481.00000.99970.99820.98710.98180.9928

ZeroStep achieves an average bit accuracy of 0.9903, compared with 0.9785 for FARI, further demonstrating its effectiveness across different watermark encoding mechanisms without online diffusion inversion.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Re: Adapter Generalization and Frozen VAE Comparison (W4 & Q5)

We thank the reviewer for the helpful suggestions.

Cross-backbone adapter reuse. Although direct transfer across different latent dimensions remains challenging, ZeroStep supports bidirectional adapter transfer between SD v1.5 and SD v2.1, as well as direct reuse between SD v3.5 Medium and Large without retraining. We further demonstrate transfer to and from SDXL-Turbo, reducing the need for backbone-specific retraining and lowering deployment costs.

Cross-backbone adapter transfer between SD v1.5, SD v2.1, and SDXL-Turbo.

Source → TargetDirect Transfer TR ↑Direct Transfer GS ↑Adapted Transfer TR ↑Adapted Transfer GS ↑
SD v1.5 → SDXL-Turbo0.9870.96470.9990.9773
SDXL-Turbo → SD v1.50.9910.99420.9990.9997
SD v2.1 → SDXL-Turbo0.9900.96550.9990.9794
SDXL-Turbo → SD v2.10.9890.99450.9990.9995

Controlled Frozen VAE comparison. To isolate encoder adaptation from distortion augmentation, we introduce ZeroStep (No Noise), which uses the adapted VAE encoder without distortion augmentation. All variants are evaluated on the same test images under identical verification protocols. Results on SD v2.1 are shown below.

MethodClean TPR ↑Avg. TPR ↑Clean Bit Acc. ↑Avg. Bit Acc. ↑
Frozen VAE + Original Verifier0.8660.2350.93730.8692
ZeroStep (No Noise)0.9950.8890.99450.9180
ZeroStep (Ours)1.0000.9991.00000.9983

Encoder adaptation alone substantially improves verification over the Frozen VAE, while distortion augmentation provides further gains in robustness.

Re: Figure 5 Metric Definition (Q4)

Correct Evidence Retained (CER). We define CER as the fraction of latent elements whose predicted signs match those of the corresponding original watermarked latents. Specifically, we partition the predicted latent elements into four quartiles based on their absolute magnitudes and calculate the sign agreement rate within each quartile. CER is therefore the proportion of sign matched elements in each quartile. We have clarified this definition and the evaluation setting (SD v2.1, GS, and 1,000 test samples) in the revised manuscript.

We sincerely thank the reviewer for the score increase. Your expertise and insightful comments have substantially improved the quality of our paper.

Reviewer_To6XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper targets verification of initial-noise diffusion watermarks (TreeRing, Gaussian shading, etc.) Instead of reconstructing the entire initial noise, the paper proposes to FT the VAE encoder via LoRA so that a verifier can still verify the watermark exists. The verification through experiments seems to be surprisingly well even on the generated image. The training includes a clean/distorted consistency term plus alignment of distorted predictions to the initial noise.

Experiments use SD1.5, SD2.1, SDXL-Turbo and SD3.5-M, 1,000 test prompts and nine distortions. ZeroStep reports higher distortion-averaged GS bit accuracy and TR TPR than multi-step, few-step and learned one-step inversion baselines;

Critical strengths
  1. Simple, practical, and consistently effective.
    • Removing the denoiser from verification entirely, with a tiny adapter, gives higher reported distortion-averaged accuracy than strong inversion baselines on four backbones (diffusion and flow) and four watermark schemes.
    • Faster and uses less memory than the strongest learned one-step method (FARI) and uses ~10× less memory (Table 5, Fig. 8).
    • Adapter transfer across compatible model variants (App. C.2) suggests the training cost can be amortized.
    • For a provider verifying large volumes of images, this is a meaningful engineering result.
  2. Useful controls and analyses.
    • Frozen VAE zero step baseline (Table 8).
    • Unseen distortions eval training (Table 11).
    • Reconstruction vs verification study a verifier-compatible latent need not reconstruct zTz_T well.
Critical weaknesses

1. Limited novelty relative to learned one-step extraction, and the proposed ingredients add little over plain regression.

At its core ZeroStep is an amortized inverse. A small network (the VAE encoder plus LoRA) is regressed from watermarked images to watermark-bearing latents and trained with augmentation; this is the same recipe as learned one-step extraction (FARI, OSI), with a smaller network.

The paper's own ablation (Table 3) shows:

  • Plain regression to zTz_T ("Initial noise MSE": TR 0.997, GS 0.9883) already beats every baseline in Table 1 (e.g., FARI on SD2.1: 0.993, 0.9824).
  • The proposed trajectory targets, recoverability weighting and magnitude weighting together add +0.002 TPR and +1.0 point of bit accuracy.

The motivating analysis (Sec. 3) is neither new nor framed correctly.

  • "Low similarity" is misleading. The paper calls the endpoint cosine C0≈0.35C_0\approx0.35 "low". But for a 16,384-dimensional latent, the cosine between independent vectors has a standard deviation of ≈1/d≈0.0081/\sqrt d\approx0.008, so 0.350.35 is tens of standard deviations above chance.
  • Already known and exploited. That the clean latent remains strongly correlated with its initial noise has been observed before and used for verification without any inversion or diffusion-model access:
    • NoisePrints (ICLR 2026) checks the cosine between the VAE encoding and the seed's noise. It reports a mean of 0.48 on SD2.0 against a threshold of 0.10 at FPR 2−1282^{-128}, and a 14–213× speedup over inversion-based verification.
    • Staniszewski et al. (ICLR 2026) also analyse the noise–image relation.
    • Neither is cited.
  • Visible in the paper's own baseline. The frozen-VAE result (Table 8: GS bit accuracy 0.94 and TR TPR 0.82–0.87 with no inversion at all) shows the same effect.
  • Encoder fine-tuning has precedents. OSI jointly fine-tunes the VAE encoder with the backbone. TraceMark-LDM (ESWA; arXiv 2503.23332; not cited) fine-tunes the LDM encoder to make an initial-noise watermark robust, though it still inverts.

So the "reveal" contribution and the "removes inversion entirely" framing need to be repositioned against this line of work.

Finally, the closest learned competitor, OSI, is excluded from all comparisons. The stated reason, that it "cannot be integrated with Tree-Ring and RingID" (App. B.1), does not apply to Gaussian Shading or PRC, which are sign-based schemes where sign classification is natural. A comparison on GS/PRC, and against a direct bit/sign classifier trained with the same data and augmentations, is needed to support the novelty and state-of-the-art claims.

2. The baseline comparison is not controlled, and the robustness and security evaluation favours the learned extractor and is incomplete.

(0) Baseline provenance and generation mismatch. I compared Table 1 with FARI's published Table 1 (ICLR 2026).

  • FARI's Table 1 has 30 baseline rows: DDIM-50, DDIM-1, EDICT, BELM, AMED, LCM-LoRA, DMD2, ExactDPM and FARI, on SD1.5 and SD2.1, in both the GS and Tree-Ring blocks. For 29 of them, the clean value and all nine per-distortion values in ZeroStep's Table 1 are identical to four decimals. The one exception is a Tree-Ring row.
  • Only the "Avg." column differs, and here the submission's values are the correct means. FARI's own averages contain arithmetic errors, e.g. SD1.5 DDIM-50: FARI gives 0.9777, but the mean of its nine values is 0.9752.
  • The SD3.5 GS FARI row matches FARI's Table 5 exactly, and the SDXL-Turbo DDIM-1 and FARI rows match FARI's Table 6.
  • Yet the paper states that "official implementations and recommended settings are used" and that baselines are reproduced (Sec. 4; App. B.1).
  • It also lists PNDM / 50 steps / CFG 7.5 as the generation sampler for SD1.5/2.1 (Table 6), whereas FARI generated with "the DDIM scheduler for all generations".
  • If the baseline numbers come from FARI's paper, ZeroStep and the baselines were evaluated on different generated images, and Table 1 is not a controlled comparison.

(a) Train/test distortion overlap. The adapter is trained on the same nine distortions at the same strengths used in the main evaluation (App. B.3, Table 7), whereas the inversion baselines are training-free.

  • The leave-one-out study helps, but each held-out distortion has a close relative in training: Gaussian↔median blur, Gaussian↔salt-and-pepper noise, crop↔drop.
  • Crop keeps spatial alignment (it zero-fills rather than cropping and resizing).
  • There is no evaluation under regeneration or purification attacks, composed distortions, geometric attacks, or adaptive attacks.
  • Adaptive attacks matter especially here: ZeroStep turns the verifier into a small, cheap, differentiable network, which lowers the cost of white-box removal or forgery.
    • FARI itself notes that one-step inversion "dramatically shrinks the computational graph" for such attacks (FARI §5).
    • Lin & Juarez (USENIX Security 2025) remove Tree-Ring using only the public VAE, cutting detector ROC-AUC from 0.993 to 0.153. ZeroStep's extractor is a LoRA on that same public encoder.
    • Forgery and removal attacks on TR/GS from one or a few images are neither evaluated nor cited.

(b) False positives are under-examined for an extractor trained only on watermarked images.

  • TR TPR@10−310^{-3} is extrapolated from a fitted ROC with only 1,000 negatives (App. B.1).
  • RingID reports identification accuracy but not false identification on unwatermarked images.
Questions for authors
  1. OSI and direct-classifier baselines.
    • Please compare with OSI on Gaussian Shading and PRC, including parameters and latency. OSI reports GS on SD2.1, SDXL and SD3.5, and PRC on SD2.1, so the stated reason for excluding it (App. B.1) does not apply to these schemes.
    • Please also compare with a direct bit/sign classifier (or the "Initial noise MSE" variant) trained with identical data and augmentations.
    • If these perform similarly, what remains as the contribution beyond using the VAE encoder as the backbone?
  2. False positives.
    • Are the null distances or thresholds computed with unwatermarked images passed through the adapted encoder?
  3. Robustness beyond the training distribution.
    • Regeneration attacks: VAE re-compression and diffusion purification, as in WAVES.
    • Composed distortions, rotation, and true crop+resize.
    • An adaptive white-box attack against EϕE_\phi (removal and forgery). How does the cost of such an attack compare with attacking an inversion-based verifier?
    • A public-VAE surrogate attack (Lin & Juarez, USENIX Security 2025), and forgery from a single image (Müller et al., CVPR 2025).
  4. Training-free correlation baselines.
    • Prior work verifies initial-noise provenance without inversion by correlating the VAE latent with the (known) initial noise (e.g., NoisePrints, ICLR 2026; on the noise–image relation, see also Staniszewski et al., ICLR 2026). How does ZeroStep relate to this line of work?
    • For schemes where the verifier knows the embedded pattern (Tree-Ring key region; GS/PRC sign pattern for a claimed message), how does a training-free correlation or matched-filter test on E(x)E(x) compare with ZeroStep under the same distortions?
    • How much of the gain over the frozen-VAE baseline (Table 8) remains against such a test?
  5. Keys and deployment cost.
    • Is the adapter trained per key or key set (Tree-Ring has a fixed key; RingID uses 32 keys)? Does it generalize to keys never seen in training?
    • How many adapters are needed per backbone × scheme × key configuration?
    • What is the end-to-end training cost, including trajectory generation?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

Re: Methodological Novelty and Baseline Comparisons (W1, Q1 & Q4)

We thank the reviewer for the thoughtful discussion of our work's novelty. Unlike methods that design new watermarking schemes or verify images by comparing their latents with the initial noise, ZeroStep investigates the shared trajectory behavior of existing initial noise watermarks. We find that watermark evidence remains recoverable at the generation endpoint despite a substantial decrease in latent similarity to the initial noise. Based on this observation, ZeroStep learns to extract verifier compatible latent representations directly from generated images, eliminating online diffusion inversion without modifying existing watermark mechanisms, while enabling adapter reuse across compatible backbones.

Cosine similarity. An endpoint cosine similarity of 0.35 indicates that some correlation with the initial noise remains. Nevertheless, the latent states deviate substantially from their initial state along the generation trajectory, as also illustrated by the latent space PCA visualization. Despite this divergence, watermark evidence remains highly recoverable. We have added trajectory analyses for GS and PRC to the revised manuscript, further demonstrating that this phenomenon is not specific to Tree-Ring.

Relation to prior work. NoisePrints studies seed based verification through initial noise correlation, while Staniszewski et al. investigate noise and inversion representations for image editing. However, these works focus on specific verification tasks or the relationships between noise and unwatermarked images, rather than watermark behavior across generation trajectories. In contrast, ZeroStep exploits these shared trajectory characteristics and develops a general latent extraction framework compatible with original verifiers, without modifying the watermark embedding or verification mechanisms.

Frozen VAE and encoder adaptation. The original Frozen VAE already preserves some watermark evidence, motivating direct verification from the final image's VAE latent without diffusion inversion. In the original manuscript, we directly feed its encoded latents into the original verifiers without additional training or UNet inference, but performance under distortions remains limited. ZeroStep addresses this limitation through trajectory supervised VAE encoder adaptation. We further compare against a VAE direct classifier, a frozen VAE, and the ZeroStep (No Noise) variant on SD v2.1. The results are summarized below.

Comparison on SD v2.1

MethodClean TPR ↑Avg. TPR ↑Clean Bit Acc. ↑Avg. Bit Acc. ↑
Direct ClassifierN/AN/A0.99020.9779
Frozen VAE + Original Verifier0.8660.2350.93730.8692
ZeroStep (Initial Noise MSE)0.9960.9970.99990.9883
ZeroStep (No Noise)0.9950.8890.99450.9180
ZeroStep (Full)1.0000.9991.00000.9983
AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

Re: Evaluation Fairness, False Positives, and Security Robustness (W2, Q2 & Q3)

We thank the reviewer for the careful examination of our experimental settings and evaluation protocols.

Baseline fairness and comparisons. We evaluate baselines using publicly available implementations and released checkpoints. For configurations not covered by prior work, we reproduce the baselines under the same generation settings, test images, and verification protocol as ZeroStep. To further assess its performance against learned extraction methods, we compare ZeroStep with FARI and OSI on Gaussian Shading across SD backbones. FARI and OSI still require a UNet forward pass, whereas ZeroStep uses only a single VAE encoder pass, with zero denoiser NFE. The results are included in the revised manuscript.

Comparison with FARI and OSI on Gaussian Shading across SD backbones.

BackboneMethodNFECleanAvg.JPEGR.CropR.DropResizeG.BlurM.BlurG.NoiseS&PBright
SDXL-TurboFARI10.99980.95110.99030.91150.91610.98710.93700.99220.98010.93540.9102
SDXL-TurboOSI11.00000.95230.96200.93500.92700.98500.96900.98700.96200.97900.8647
SDXL-TurboZeroStep01.00000.97910.99120.98380.98670.99060.93600.99440.99460.97400.9607
SD v3.5FARI10.99950.95150.96710.97130.94690.97560.93840.98520.87650.91910.9833
SD v3.5OSI10.99930.95460.93020.95360.92900.95380.93060.97650.94950.99330.9749
SD v3.5ZeroStep01.00000.99110.99240.99950.99630.99260.96190.99410.99700.99050.9954

False positive evaluation. We follow the original evaluation protocols of each watermark scheme. Tree-Ring is evaluated using TPR at a fixed FPR of 10−310^{-3}, while GS and PRC use bit accuracy. We further increase the negative sample set beyond the original 1,000 samples and evaluate Tree-Ring at a stricter FPR of 10−610^{-6}.

Tree-Ring detection performance at FPR = 10−610^{-6}.

BackboneMethodNFECleanAvg.JPEGR.CropR.DropResizeG.BlurM.BlurG.NoiseS&PBright
SD v1.5DDIM-50500.9990.9360.9820.9960.9940.9970.9900.9970.5900.9250.955
SD v1.5FARI10.9990.9890.9960.9960.9940.9980.9960.9970.9550.9920.978
SD v1.5ZeroStep01.0000.9960.9990.9990.9980.9990.9980.9990.9820.9980.993
SD v2.1DDIM-50500.9990.9490.9850.9960.9950.9970.9910.9970.6800.9600.940
SD v2.1FARI10.9990.9840.9930.9970.9960.9980.9960.9970.9250.9900.967
SD v2.1ZeroStep01.0000.9951.0001.0000.9990.9990.9970.9990.9760.9900.996

Additional attacks and security robustness. We further evaluate robustness under true crop and resize, as well as compound distortions, to assess generalization beyond the original training distribution. We also include three representative regeneration attacks: Regen-Diff, Rinse-2xDiff, and CtrlRegen. All methods are evaluated on the same attacked images using the original GS verifier.

Robustness under compound distortions and regeneration attacks on SD v2.1 with Gaussian Shading.

AttackDDIM-50DDIM-1FARIZeroStep
R.Crop + Resize0.95220.84790.95830.9997
Resize + JPEG0.98180.96430.99210.9991
G.Blur + M.Blur0.98130.94810.99350.9995
Regen-Diff0.94860.83000.95470.9671
Rinse-2xDiff0.92070.79740.94110.9584
CtrlRegen0.89530.72510.90920.9204

ZeroStep achieves the highest bit accuracy across all evaluated attacks. However, the ability to resist attacks remains largely dependent on the underlying design of the base watermarking methods, as Zerostep serves as a plug-and-play approach to enhance their detection performance.

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

Re: Key Generalization and Deployment Cost (Q5)

We thank the reviewer for raising questions about key generalization and deployment costs. ZeroStep does not take keys or messages as input. For GS, we randomly sample messages, keys, and nonces during training. For RingID, the same adapter supports candidate key sets ranging from 32 to 2,048 keys, as demonstrated in the original manuscript. Thus, only one shared adapter is needed per backbone and watermark scheme. We further demonstrate adapter transfer across SD v1.5, SD v2.1, and SDXL-Turbo.

Cross-backbone adapter transfer between SD v1.5, SD v2.1, and SDXL-Turbo.

Source → TargetDirect TR ↑Direct GS ↑Adapted TR ↑Adapted GS ↑
SD v1.5 → SDXL-Turbo0.9870.96470.9990.9773
SDXL-Turbo → SD v1.50.9910.99420.9990.9997
SD v2.1 → SDXL-Turbo0.9900.96550.9990.9794
SDXL-Turbo → SD v2.10.9890.99450.9990.9995

We also report the complete offline adaptation costs, including trajectory generation, caching, and LoRA training, to assess the practical deployment overhead.

Offline adaptation costs across backbones and watermark schemes.

BackboneWatermarkTrajectory Gen. (s)Caching (s)LoRA Training (s)Total Time (s)Cache Size (GiB)
SD v2.1Tree-Ring299.05354.82402.491056.368.057
SD v2.1GS300.76350.76402.661054.188.057
SDXL-TurboTree-Ring49.85311.01402.03762.897.446
SDXL-TurboGS50.09311.84402.25764.187.447
SD v3.5MTree-Ring1204.79349.46402.321956.5711.353
SD v3.5MGS1206.08348.12402.341956.5411.353

These costs are incurred only during offline adaptation and are not required at verification time. The total adaptation time is approximately 17.6 minutes for SD v2.1, 12.7 minutes for SDXL-Turbo, and 32.6 minutes for SD v3.5 Medium.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer To6X,

We appreciate your insightful feedback. We have carefully considered your comments and provided a point-by-point response, and we hope it addresses your concerns.

We are very keen to clarify any remaining points and look forward to your feedback.

Best regards,

Authors

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

This paper proposes ZeroStep, a lightweight framework for verifying diffusion watermarks without iterative diffusion inversion. Motivated by the observation that watermark detectability can persist even when intermediate representations differ substantially from the original noise, the method adapts a VAE encoder through LoRA to directly produce verifier-compatible latents. Training combines recoverability-guided trajectory supervision, distortion consistency, and magnitude-weighted regression. Experiments across multiple generative backbones and watermarking schemes demonstrate competitive verification performance with substantially reduced inference latency and memory consumption.

Critical strengths

The paper addresses a practically important bottleneck in diffusion watermark verification and makes a useful distinction between accurate noise reconstruction and successful watermark recovery. The encoder-only design preserves existing watermark embedders and verifiers while substantially reducing computational cost: the reported extraction latency is 19.0 ms, with 0.45 GiB peak memory. The evaluation is relatively comprehensive, covering diffusion and flow-matching backbones, multiple watermarking schemes, distortion severity, and compatible-backbone transfer. The supplementary experiments, particularly the frozen-VAE baseline and additional RingID and PRC results, strengthen the empirical case. Overall, the work presents a promising engineering approach to efficient watermark verification.

Critical weaknesses

(1)The paper excludes OSI [1] because its sign-classification formulation is incompatible with some latent-space verifiers. However, this does not preclude a comparison on Gaussian Shading, which both approaches support. Such a comparison is important for establishing the advantage over task-specific learned extraction, even if OSI cannot cover every watermarking scheme. Moreover, the initial-noise regression baseline should receive the same augmentation and magnitude-weighting components as ZeroStep to isolate the contribution of trajectory supervision. [1] dOSI: One-step Inversion Excels in Extracting Diffusion Watermarks (2)The “crop” operation retains a square region at its original coordinates and zeros the surrounding pixels. This primarily evaluates information removal rather than the spatial misalignment introduced by actual cropping and resizing. Although severity sweeps and leave-one-distortion-out experiments are valuable, they do not establish robustness to realistic spatial transformations or compound processing. Actual crop-and-resize operations, rotation, and representative combinations such as resizing followed by JPEG compression would provide stronger evidence for deployment robustness. (3)The manuscript should specify the value of lambda_w, the exact encoder modules receiving LoRA, and whether gradients propagate through the prediction-dependent magnitude weights. It should also clarify whether trajectory states are normalized before regression, since their scales can affect the effective supervision. Finally, the online efficiency gains are convincing, but trajectory generation, caching, and encoder adaptation introduce offline costs that are not quantified. Reporting these costs would clarify the practical trade-off, particularly when adaptation is required for additional watermark schemes or backbones.

Questions for authors

Please refer to the weakness.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We thank the reviewer for the constructive feedback. We address each comment below.

Re: Baseline Comparisons

Comparison with learned extraction methods. To further assess its performance against learned extraction methods, we compare ZeroStep with FARI and OSI on Gaussian Shading across different SD backbones. FARI and OSI still require a UNet forward pass, whereas ZeroStep requires only a single VAE encoder pass, with zero denoiser NFE. The results have been added to the revised manuscript.

Comparison with FARI and OSI on Gaussian Shading across SD backbones.

BackboneMethodNFECleanAvg.JPEGR.CropR.DropResizeG.BlurM.BlurG.NoiseS&PBright
SDXL-TurboFARI10.99980.95110.99030.91150.91610.98710.93700.99220.98010.93540.9102
SDXL-TurboOSI11.00000.95230.96200.93500.92700.98500.96900.98700.96200.97900.8647
SDXL-TurboZeroStep01.00000.97910.99120.98380.98670.99060.93600.99440.99460.97400.9607
SD v3.5FARI10.99950.95150.96710.97130.94690.97560.93840.98520.87650.91910.9833
SD v3.5OSI10.99930.95460.93020.95360.92900.95380.93060.97650.94950.99330.9749
SD v3.5ZeroStep01.00000.99110.99240.99950.99630.99260.96190.99410.99700.99050.9954

Re: Ablation of Training Objectives

Contribution of adaptive trajectory selection and magnitude weighting. To isolate the contributions of adaptive trajectory selection and magnitude weighting, we conduct additional ablations using initial noise regression and different trajectory sampling strategies. The results demonstrate that magnitude weighting improves both initial noise regression and trajectory-based supervision, while adaptive trajectory selection provides further gains.

Ablation on training objectives.

VariantAdaptive Traj.Mag.TR ↑GS ↑
Initial Noise MSE––0.9970.9883
Initial Noise MSE + Magnitude Weighting–✓0.9970.9911
Uniform Trajectory––0.9970.9924
Recoverability Trajectory✓–0.9980.9956
Recoverability Trajectory + Magnitude Weighting✓✓0.9990.9983
AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

Re: Additional Attacks and Security Robustness

Additional attacks and security robustness. We further evaluate robustness under true crop and resize, as well as compound distortions, to assess generalization beyond the original training distribution. We also include three representative regeneration attacks, Regen-Diff, Rinse-2xDiff, and CtrlRegen. All methods are evaluated on the same attacked images using the original GS verifier.

Robustness under compound distortions and regeneration attacks on SD v2.1 with Gaussian Shading.

AttackDDIM-50DDIM-1FARIZeroStep
R.Crop + Resize0.95220.84790.95830.9997
Resize + JPEG0.98180.96430.99210.9991
G.Blur + M.Blur0.98130.94810.99350.9995
Regen-Diff0.94860.83000.95470.9671
Rinse-2xDiff0.92070.79740.94110.9584
CtrlRegen0.89530.72510.90920.9204

ZeroStep achieves the highest bit accuracy across all evaluated attacks. However, its resistance to attacks remains largely dependent on the underlying design of the base watermarking methods, as ZeroStep serves as a plug-and-play latent extraction framework to enhance their detection performance.

Re: Offline Training Cost and Implementation Details

Offline adaptation costs. We report the complete offline adaptation costs, including trajectory generation, caching, and LoRA training, to assess the practical deployment overhead.

Offline adaptation costs across backbones and watermark schemes.

BackboneWatermarkTrajectory Gen. (s)Caching (s)LoRA Training (s)Total Time (s)Cache Size (GiB)
SD v2.1Tree-Ring299.05354.82402.491056.368.057
SD v2.1GS300.76350.76402.661054.188.057
SDXL-TurboTree-Ring49.85311.01402.03762.897.446
SDXL-TurboGS50.09311.84402.25764.187.447
SD v3.5MTree-Ring1204.79349.46402.321956.5711.353
SD v3.5MGS1206.08348.12402.341956.5411.353

These costs are incurred only during offline adaptation and are not required at verification time. The total adaptation time is approximately 17.6 minutes for SD v2.1, 12.7 minutes for SDXL-Turbo, and 32.6 minutes for SD v3.5 Medium. Further details on the hyperparameters and gradient computation have been added to the revised appendix.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer Xsfp,

We appreciate your insightful feedback. We have carefully considered your comments and provided a point-by-point response, and we hope it addresses your concerns.

We are very keen to clarify any remaining points and look forward to your feedback.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.