acceptodds
Under review as a conference paper at ICLR 2027

IRIS: Visual-Semantic Binding for Diffusion Image Watermarking

Abstract

In-generation watermarking provides a promising approach to tracing diffusion-generated images, but recent work shows such watermarks can be transplanted onto arbitrary images because they are agnostic to the visual content being watermarked. Prior content-aware methods bind the watermark only to prompt-level semantics, leaving it loosely coupled to the visual content it watermarks. We present IRIS, a training-free in-generation watermarking framework that binds each watermark to the visual semantics of the generated image. This introduces two challenges: (1) deriving the watermark requires the visual semantics of the generated image, yet embedding that watermark during generation may itself alter those semantics, creating a circular dependency, and (2) verification should be sensitive to semantic changes while resilient to common image distortions. To address the first challenge, IRIS derives the watermark from the visual semantics of a preview. It embeds the watermark as a small perturbation in the Fourier domain of the latent only during the final few sampling steps, thereby preserving those semantics. To address the second challenge, IRIS constructs distortion-robust semantic codes from augmented visual representations. At detection, IRIS derives the expected watermark from the query image and verifies it against the inverted latents. Extensive experiments show that IRIS reduces the mean black-box forgery attack success rate from over 96% for content-agnostic methods and 11–61% for prior content-aware methods to 2.1%, while keeping high detectability and generalizing to SD3.5 and FLUX.2.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
image watermarking, in-generation watermarking, diffusion models, responsible AI

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_XFn9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Approach

Generation: First, a preview image is generated, its semantic code obtained by CLIP embedding and subsequent SimHash projection. 
In order to embed the watermark, we return to a previous latent (e.g. at step 45, out of 50), and embed the semantic code into the frequency representation of the latent. Then we re-generate the remaining few steps but now from the watermarked latent.

Detection: the image’s semantic code is recovered by preprocessing+CLIP+SimHash. A candidate set of semantic codes is constructed that enables some tolerance to most sensitive bit flips in the semantic code. Candidate set consists of every code with at most four bit flips and every code obtained by flipping any of the 9 most uncertain semantic code bits. An expected watermark is generated for every candidate code. Note that the generated watermark patterns are independent from each other if the code differs in even one bit. In the second branch, in parallel, the image is inverted partially for a number of steps. The detection score is computed over multiple inversion steps using an alignment metric with the expected watermark. In particular, a normalized alignment score is computed for every candidate pattern against every inverted time step. The scores are averaged over the timesteps to obtain an alignment score over each candidate. The decision is then done by checking how strongly the best-matching candidate matches to the image, compared to the other candidates. If one candidate’s score is sufficiently higher than the median, the image is considered watermarked.

Preprocessing: a particularly important aspect appears to be the preprocessing, which is done to obtain more stable semantic codes. The image is copied 8 times, and on each copy, a transformation is applied (JPEG, additive noise, or blur). Then the copies are canonicalized by normalizing to mean luminance of 0.5, and by applying blur and JPEG25 compression on every image. Each preprocessed copy goes into CLIP and the CLIP embeddings are averaged before being used in semantic code construction. This is done both during generation and verification (in the semantic code branch).

Experiments are done on SD2.1-base, SD3.5 and Flux.2-klein, comparing against TreeRing, RingID, GS and PRC, and SEAL and SemBind. From the results it appears that the proposed method has high detectability, and substantially lower forgeability than SEAL and SemBind. However, it has generally lower robustness than SemBind+GS, Gaussian Shading and RingID. Ablations are provided. Without preprocessing, the robustness is poor. Interestingly, reading out only at step 45 (and not averaging over 20 inverted steps) reduces robustness, especially under Gaussian Noise.

Critical strengths

S1: interesting problem, trying to address forgeability of in-generation watermarks.

S2: reasonable approach. Reusing earlier timesteps makes sense to solve the circular dependency incurred by conditioning on image semantics. Tricks for improving robustness (preprocessing and candidate sets) appear to work.

Critical weaknesses

W1: Robustness is subpar. TPR of <99% with an FPR of 1% is not great and much lower than old baselines like RingID or GS, or SemBind+GS. TPR of 93% for Gaussian Noise is particularly weak. Improved resistance to the evaluated forgery attacks makes up for this to some degree, but still raises concerns whether the poor robustness is acceptable. For some unspecified reason, SEAL is omitted from robustness comparison.

W2: Missing ablations: There is no ablation on at what timestep the watermark is embedded, and no ablation for which steps are most important for detection (except step 45). The fact that in the last denoising steps, the reported latent estimates become highly aligned, raises the question whether we can’t simply apply the watermark on the final latent. Another important missing analysis concerns the preprocessing. The paper describes generating 8 augmented views, which are then “canonicalized” using heavy blur and jpeg compression. Given that the perturbations done before this are for a large part absorbed by canonicalization, it could be that we don’t need 8 augmented views and only need the canonicalization part on a single image. So I believe the papers needs to check for canonicalization only (no augmentations before canonicalization) as well as a few settings on the number of augmented views (1, 4, 8, 16).

W3: writing needs to be improved to improve better understanding of the paper. Also, some claims in intro aren’t correct. SemBind does not rely on prompt or caption like SEAL. Also, the paper is stating that previous methods didn’t try binding semantics and a few paragraphs later it talks about SEAL etc. The earlier claim needs to be properly scoped. Moreover, there are other presentation issues, like the non-standard placement of tables, with text wrapping around the tables (not sure this is permissible in ICLR). Also, the presentation of numbers in the tables is potentially confusing. Usually bold font is reserved for pointing out the best results in a column whereas in this paper, it simply marks the IRIS row (which is already clearly marked in blue).

Questions for authors

Q1: Why is SEAL not included in the robustness comparison?

Minor remarks: I am not sure that table formatting and their placement in the text is acceptable. Table caption is below the table.

4 comments
AuthorsOpenReview ↗

Response to Reviewer XFn9 [Part 1]

W1: Robustness is subpar. TPR of <99% with an FPR of 1% is not great and much lower than old baselines like RingID or GS, or SemBind+GS. TPR of 93% for Gaussian Noise is particularly weak. Improved resistance to the evaluated forgery attacks makes up for this to some degree, but still raises concerns whether the poor robustness is acceptable. For some unspecified reason, SEAL is omitted from robustness comparison. Why is SEAL not included in the robustness comparison?

We sincerely thank Reviewer XFn9 for providing these thoughtful comments and questions.

We would like to note that IRIS improves over both content-aware and content-agnostic methods. Compared to content-aware methods, IRIS pays a smaller cost in robustness for a significant gain in forgery resistance. Compared to content-agnostic methods, IRIS achieves comparable detectability when using a key-only pattern and maintains improved visual fidelity (the watermarked image stays visually identical to the unwatermarked image).

1. IRIS achieves an improved trade-off between robustness to distortions and resistance to watermark transplantation. The table below reports the average T@1%FPR over the seven distortions in Section 5.2, Table 2 and and the 1−1-Forgery ASR from Section 5.2, Table 1. The last column reports the worst of the two. Relative to SemBind+GS, IRIS gives up 2.2 points of distortion robustness and gains 43.8 points of forgery resistance.

MethodAvg. T@1%FPR, distortions (Table 2, 7 types)1 − Forgery ASR (Table 1, mean)MeanWorst case
IRIS0.9720.9790.9760.972
SemBind+TR0.5140.9030.7090.514
SemBind+GS0.9940.5410.7680.541
SEAL0.7910.3500.5710.350

2. The oracle mark detectability of IRIS achieves robustness approaching that of RingID and GS. As shown in Section 5.2, Table 2, IRIS oracle pattern detection rates (T@1%F) are 0.99–1.0. In our response to Reviewer 2DoM, when evaluating WAVES removal attacks, IRIS without semantic binding also achieves robustness comparable to GS and RingID. We note that IRIS achieves this robust oracle mark detectability while maintaining high visual fidelity, which GS and RingID do not.

3. Robustness comparison with SEAL. We evaluated SEAL under the same protocol as the other baselines in Table 2, but we omitted its row by oversight. We have added this row to Table 2 in the revision. SEAL is less robust than IRIS under every distortion.

MethodJPEGBlurNoiseBrightCropRot.Regen.
SEAL0.7520.6180.6380.9840.9860.9540.606
IRIS0.9760.9940.9320.9940.9920.9760.940
AuthorsOpenReview ↗

Response to Reviewer XFn9 [Part 2]

w2: Missing ablations: There is no ablation on at what timestep the watermark is embedded… whether we can’t simply apply the watermark on the final latent. … missing analysis concerns the preprocessing. … check for canonicalization only (no augmentations before canonicalization) as well as a few settings on the number of augmented views (1, 4, 8, 16).

We sincerely appreciate these thoughtful comments. We added ablations on the embedding timestep and the image preprocessing to the revision.

1. Ablation of watermark embedding timestep.

embedding start step i0i_0clean (T@1%FPR)noise 0.1 (T@1%FPR)noise 0.1 (code recovery)noise 0.1 (oracle pattern)PSNR
200.8980.7420.8640.97223.9
300.9700.8420.9420.98026.6
400.9980.8960.9860.98829.3
45 (default)1.0000.9320.9821.00031.3
471.0000.9240.9900.99232.5
49 (final latent)1.0000.8280.9940.97835.1

Findings: Starting the embedding later than step 45 yields a T@1%FPR of 1.0 on clean images and a higher PSNR, but lower robustness under Gaussian noise (σ\sigma = 0.1).

We choose i0i_0 as the earliest embedding start that reaches a clean T@1%FPR of 1.0. Later starts bind the mark more tightly to the image content and give a higher PSNR, but the mark passes through fewer denoising steps and becomes less robust under noise (rows 47–49). Earlier starts lose both clean detectability and fidelity (rows 20–40).

2. Ablation of image preprocessing.

We evaluate the effectiveness of image preprocessing by examining the code recovery rate and Hamming distance of embedded and extracted semantic codes.

PreprocessingMetricQ90Q70Q50Q30Q20
Raw CLIP (no preprocessing)Recovery0.9800.9140.8560.7420.654
Raw CLIP (no preprocessing)Hamming ↓1.562.433.023.724.14
Canonicalization only (no augmentation)Recovery0.9980.9980.9940.9860.952
Canonicalization only (no augmentation)Hamming ↓0.791.121.281.461.96
Canon. + K = 1 augmented viewRecovery1.0000.9900.9880.9820.944
Canon. + K = 1 augmented viewHamming ↓1.001.311.411.782.14
Canon. + K = 4Recovery1.0001.0001.0000.9940.984
Canon. + K = 4Hamming ↓0.530.720.911.101.64
Canon. + K = 8 (ours)Recovery1.0001.0001.0001.0000.996
Canon. + K = 8 (ours)Hamming ↓0.370.620.800.951.54
Canon. + K = 16Recovery1.0001.0000.9981.0000.988
Canon. + K = 16Hamming ↓0.290.490.710.851.48

Findings: Canonicalization alone effectively increases code recovery rate and reduces Hamming distance. Averaging over augmented views further improves these two metrics under JPEG compression. A single random view (K = 1) is slightly worse than canonicalization alone, because one random augmentation adds variance that is only averaged out with several views.

AuthorsOpenReview ↗

Response to Reviewer XFn9 [Part 3]

W3 & remarks: writing needs to be improved to improve better understanding of the paper. Also, some claims in intro aren’t correct. SemBind does not rely on prompt or caption like SEAL. Also, the paper is stating that previous methods didn’t try binding semantics and a few paragraphs later it talks about SEAL etc. The earlier claim needs to be properly scoped. Moreover, there are other presentation issues, like the non-standard placement of tables, with text wrapping around the tables (not sure this is permissible in ICLR). Also, the presentation of numbers in the tables is potentially confusing. Usually bold font is reserved for pointing out the best results in a column whereas in this paper, it simply marks the IRIS row (which is already clearly marked in blue).

We sincerely thank you for these careful questions and comments that help us improve this paper. We clarify that SemBind uses a prompt-conditioned corpus to train its semantic code extractor. We have revised the claim in the introduction and fixed the issues in table and data presentation in the revision.

1. SemBind’s semantic binding is prompt-level. The reviewer is right that SemBind does not take the prompt or a caption as input: its semantic masker operates on the embedding of an auxiliary generated image. Our description “prompt-level” refers to the granularity of the binding, not the input. The masker of SemBind is trained with contrastive learning on a large prompt-conditioned corpus so that it yields "near-invariant codes for the same prompt and near-orthogonal codes across prompts" [1, Abstract and Sec. 3.1]. The watermark is therefore bound to the prompt-level cluster that an image belongs to. To avoid this misreading, we have clarified the introduction: it now states that SemBind derives its code from an auxiliary image’s embedding trained for prompt-level clustering, and we removed the sentence “The prompt is the only image-related quantity available before generation starts, so the binding is forced to stop there.”

[1] Zhang, Xin, et al. “SemBind: Binding Diffusion Watermarks to Semantics Against Black-Box Forgery Attacks.” Forty-third International Conference on Machine Learning. 2026.

2. Scope of the claim in the introduction. We agree that the claim was not properly scoped. The second paragraph of the introduction stated that “existing in-generation watermarking methods” are content-agnostic. In contrast, the third paragraph then introduced SEAL and SemBind as methods that do bind the watermark to semantics. We have revised the second paragraph so that the content-agnostic claim applies only to the methods it describes: it now reads “most existing in-generation watermarking methods, such as Tree-Ring and Gaussian Shading, … embed content-independent marks in the initial noise”, and attributes their vulnerability to the fact that “their watermarks are agnostic to the content”. The third paragraph then positions SEAL and SemBind as content-aware methods whose binding is loose because the semantics come from an auxiliary image, which is the limitation IRIS addresses.

3. Presentation. We replaced the wrapped tables with standard floating tables with titles above, and we removed the bold font from the IRIS rows. Bold now marks the best value in each column, and the IRIS row is highlighted only by the colored background.

Best regards,

The Authors

AuthorsOpenReview ↗

Final Request for Feedback Before the Discussion Closes

Dear Reviewer XFn9,

With less than three hours remaining in the discussion period, we sincerely ask you to take a final look at our rebuttal.

We have carefully addressed your concerns through additional robustness evaluations, comprehensive ablation studies on embedding timing and preprocessing, and substantial improvements to the manuscript's writing and presentation.

Despite our previous reminders, we have not yet received your feedback. We would be extremely grateful if you could briefly confirm whether our responses have addressed your concerns and reconsider your assessment accordingly.

We sincerely hope to have the opportunity to hear from you before the discussion closes.

Thank you very much for your time and consideration.

Best regards,
The Authors

Reviewer_2DoMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the critical vulnerability of watermark transplantation (forgery attacks) in diffusion image watermarking, where content-agnostic watermarks can be extracted and unlawfully transferred onto arbitrary unwatermarked images. The authors propose IRIS, a training-free, in-generation watermarking framework that binds the embedded watermark directly to the visual semantics of the generated image. To overcome circular dependencies, IRIS generates a clean preview to extract visual features via CLIP/DINO, computes a resilient binary code using keyed SimHash, and embeds an image-specific Fourier pattern during the final few denoising steps. At detection, a local candidate set of semantic codes is derived from the query image to verify alignment against inverted latents.

Critical strengths

a. Effective Defense Against Watermark Transplantation : IRIS successfully tackles a major security threat by binding the watermark payload directly to image-level visual semantics via keyed SimHash, dropping forgery ASR significantly. b. Minimal Visual Impact & Generalization : By restricting Fourier-domain perturbation to the final sampling steps, IRIS maintains high visual fidelity and generalizes well to modern flow-matching models.

Critical weaknesses

a. Lack of Standardized Robustness Benchmark Evaluation : While IRIS tests existing forgery attacks, it restricts its robustness evaluation to a hand-picked suite and completely omits established community-wide Robust Watermark Benchmarks such as WAVES (ICML'24), WaterBench (ICLR'25), MarkDiffusion (JMLR'26), WARP (ACM MM'26), or classical Stirmark 3.0/4.0. Furthermore, it lacks evaluation against adversarial attacks (e.g., WAVES...). b. Vulnerabilities in Adaptive Attacks & Theoretical Security Limits : On Surrogate Key Evaluation and Theoretical Robustness: While surrogate keys function primarily as embedding parameters rather than strict cryptographic secrets, relying solely on independent random seeds to demonstrate robustness provides limited theoretical guarantees. From a statistical learning perspective, further discussion or analysis on whether an adaptive adversary could bypass or approximate this mapping via query-based strategies would strengthen the security claims. On the Transparency of the Gray-Box Attack Loss (Eq. 9): Since the loss formulation for the gray-box embedding-matching attack (Equation 9) is placed in the Appendix, providing a more detailed intuition and derivation for its design choices—specifically explaining the rationale behind the multi-task formulation and the sensitivity of the hyperparameter λ\lambda—is vital for transparency, clarity, and reproducibility. Over-Idealized and Insufficient Rigor in Proposition 1 : The theoretical derivation and assumptions made in Proposition 1 are overly ideal and lack sufficient mathematical rigor to support the claimed security bounds under practical distribution shifts.

Questions for authors

a. Benchmark & Adversarial Evaluation : How does IRIS perform under comprehensive benchmark suites (e.g., WAVES, WaterBench) and specialized adversarial attacks? b.Adaptive Query-Based Strategies & Loss Derivation : Can the authors elaborate on how an adaptive adversary might approximate the semantic mapping via query-based strategies, and provide a clearer step-by-step derivation for Eq. 9 along with sensitivity analyses for λ\lambda? c. Formalizing Proposition 1 : Can you relax the ideal assumptions in Proposition 1 to provide a more rigorous proof?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 2DoM [Part 1]

We would like to sincerely thank Reviewer 2DoM for providing the insightful review and constructive comments.

W1 & Q1: Lack of Standardized Robustness Benchmark Evaluation & Adversarial Evaluation: ... omits established community-wide Robust Watermark Benchmarks such as WAVES (ICML’24), WaterBench (ICLR’25), MarkDiffusion (JMLR’26), WARP (ACM MM’26), or classical Stirmark 3.0/4.0. Furthermore, it lacks evaluation against adversarial attacks (e.g., WAVES…). How does IRIS perform under comprehensive benchmark suites (e.g., WAVES, WaterBench) and specialized adversarial attacks?

The distortions evaluated in the paper (JPEG compression, Gaussian noise, Gaussian blur, and brightness) are among the WAVES [1] distortion types. In response to reviewer 2seY, we added evaluations of rotation and crop-resize following the WAVES protocol. All baselines are evaluated with their official implementations (Appendix C), which MarkDiffusion [2] reports that its unified implementation reproduces the results of these official implementations for Tree-Ring, RingID, Gaussian Shading, PRC and SEAL, so our baseline numbers are comparable with that toolkit.

As a supplement, we added evaluations under WAVES removal attacks: regeneration (Regen-KLVAE, Rinse-2x, Rinse-4x) and adversarial embedding (Adv-CLIP, Adv-VAE).

Robustness under WAVES removal attacks (TPR @ 1% FPR, mean over five strengths), forgery resistance, and their equally weighted mean

MethodRegen-KLVAEAdv-CLIPAdv-VAERinse-2xRinse-4xMean TPR (5 attacks) ↑1 − Forgery ASR ↑Equally weighted mean ↑
Tree-Ring0.9990.9870.0380.8880.8020.7430.0000.371
RingID1.0001.0000.6100.9860.9750.9140.0020.458
Gaussian Shading1.0001.0000.7551.0001.0000.9510.0000.476
PRC0.9600.9680.1900.7680.6860.7140.0160.365
SemBind + Tree-Ring0.9050.8380.0980.4320.3610.5270.9030.715
SemBind + GS0.9990.9900.7350.9960.9940.9430.5410.742
SEAL0.7310.8600.3400.2100.1140.4510.3500.401
IRIS (default; visual-semantic bound)0.9960.9530.2950.7890.6130.7290.9790.854
IRIS (fixed pattern, no binding)1.0001.0000.5280.9990.9920.9040.2100.557

Forgery ASR is the mean over Imprint-Forgery and Reprompting from the separate forgery evaluation (Table 1; Table 3 for the fixed-pattern control). The last column is the mean of the two preceding columns.

Findings:

(1) IRIS has the highest equally weighted mean of robustness under the five WAVES attacks and forgery resistance (0.854), ahead of SemBind+GS (0.742) and SemBind+Tree-Ring (0.715). The content-agnostic methods are the most robust (mean TPR 0.714–0.951) but are transplanted almost always (forgery ASR ≥ 0.984). IRIS keeps the mark detectable under attacks that preserve visual content while rejecting transplanted marks.

(2) With a fixed pattern (no semantic binding), the IRIS carrier and detector outperform Tree-Ring, which also uses a single latent channel, in all five evaluations, and approach RingID and Gaussian Shading in four. This control shows that the robustness gap of the default IRIS comes from the semantic code rather than from the carrier. The fixed pattern, however, loses resistance to watermark transplantation (Imprint ASR 0.580, Table 3).

[1] An, B., Ding, M., Rabbani, and et al. (2024). WAVES: Benchmarking the Robustness of Image Watermarks. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:1456-1492.

[2] Pan, Leyi, et al. "Markdiffusion: An open-source toolkit for generative watermarking of latent diffusion models." Journal of Machine Learning Research 27.204 (2026): 1-22.

AuthorsOpenReview ↗

Response to Reviewer 2DoM [Part 2 (1/2)]

W2 & Q2: Adaptive Query-Based Strategies & Loss Derivation: Can the authors elaborate on how an adaptive adversary might approximate the semantic mapping via query-based strategies, and provide a clearer step-by-step derivation for Eq. 9 along with sensitivity analyses for ? Formalizing Proposition 1: Can you relax the ideal assumptions in Proposition 1 to provide a more rigorous proof?

1. Query-based approximation of the mapping. IRIS is designed so that even if an adversary knew the keyed hyperplanes, a forgery would still have to keep its visual representation within a small angle of the source image, far smaller than what caption- or prompt-level binding permits, i.e., a tighter semantic binding. Appendix B bounds candidate-set inclusion by this angle, and Appendix I.2 measures how close the gray-box attack gets in practice. We discuss two query models.

(1) Without a detection oracle: the adversary observes watermarked images but no detector decisions. HMAC-based pattern derivation prevents direct observation of the semantic-code labels. Whether the mapping can be learned from images alone remains an open question.

(2) With a detection oracle: the adversary could move a watermarked source toward a target and search for the point where the decision flips. A decision change does not necessarily mark a SimHash hyperplane crossing, since candidate-set changes also affect the ratio statistic. Even so, what the adversary can learn is of limited use: Table 21 shows that code collisions are governed by the CLIP angle to the source, dropping from 0.429 at 15°–20° to 0.007 beyond 30°. A successful forgery must therefore stay within about 20° of the source, i.e., visually close to it. Oracle-assisted attacks are outside the present evaluation.

Our threat model is the first query model (no detection oracle). We added this discussion of the detection oracle to Appendix I.

2. Derivation of Eq. 8 (gray-box loss; Eq. 9 in the submitted version) and sensitivity analysis.

Notation. Let xsx_s be a source watermarked image. Let xtx_t be a target unrelated real image. Let δ\delta be a pixel perturbation applied to the target image. IA(n)(⋅)\mathcal I_A^{(n)}(\cdot) denotes nn-step DDIM inversion with a surrogate model with its surrogate VAE encoder, which is differentiable for small nn. z^s=IA(50)(xs)\hat z_s=\mathcal I_A^{(50)}(x_s) is the adversary’s estimate of the source’s initial noise. e(⋅)e(\cdot) is the public CLIP ViT-B/32 image embedding, ℓ2\ell_2-normalised. The adaptive adversary knows the image preprocessing mechanism and uses canonv(⋅)\mathrm{canon}v(\cdot) as a differentiable approximation of our canonicalization for a random view vv, using the same processing operations. The adversary’s goal is to pass off the unrelated target image as watermarked.

Step-by-step construction of Eq. 8. We construct a gray-box loss based on the Imprint-Forgery [1] objective function and add a semantic term adapted to the visual-semantic binding mechanism.

(1) Transplant the mark. We make a target image carry the source watermark by pushing the target’s surrogate inversion toward the source’s inverted noise. The objective term is written as Limp(δ)=∣IA(10)(xt+δ)−z^s∣22\mathcal L_{\mathrm{imp}}(\delta)=|\mathcal I_A^{(10)}(x_t+\delta)-\hat z_s|_2^2. Against content-agnostic watermarks, this term alone succeeds (Table 1, ASR ≥96\ge 96%).

(2) Forge the semantic binding. We forge the semantic binding between the mark and the image’s visual representation by minimizing the cosine distance between the CLIP representations of the target and the source images’ preprocessed representations. The objective term is written as Lsem(δ)=1−cos⁡(eˉA(xt+δ),eˉA(xs))\mathcal L_{\mathrm{sem}}(\delta)=1-\cos\big(\bar e_A(x_t+\delta),\bar e_A(x_s)\big).

(3) Joint objectives. The two objectives have different scales, where Limp\mathcal L_{\mathrm{imp}} is a squared latent distance and Lsem∈[0,2]\mathcal L_{\mathrm{sem}}\in[0,2]. We divide the Imprint term by its value at δ=0\delta=0, so that it equals 1 at initialization and is independent of the latent scale. The joint objectives are written as follows: L(δ)=∣IA(10)(xt+δ)−z^s∣22∣IA(10)(xt)−z^s∣22+β,(1−cos⁡(eˉA(xt+δ),eˉA(xs))),∣δ∣∞≤ϵ.\mathcal L(\delta)=\frac{|\mathcal I_A^{(10)}(x_t+\delta)-\hat z_s|_2^2}{|\mathcal I_A^{(10)}(x_t)-\hat z_s|2^2}+\beta,\Big(1-\cos\big(\bar e_A(x_t+\delta),\bar e_A(x_s)\big)\Big),\qquad |\delta|\infty\le\epsilon . Here, β\beta sets how much of the ϵ\epsilon-budget the adversary spends on optimizing semantic binding. We revise the hyperparameter sensitivity from λ\lambda to β\beta because λ\lambda denotes embedding strength in the main text.

(4) Optimisation. We use Projected sign-gradient descent (PGD) to optimize the joint objectives, with step α=1/255\alpha=1/255, 150 iterations, ϵ=16/255\epsilon=16/255, followed by clipping to the valid pixel range. The update term is written as δ←Πϵ(δ−α,sign∇δL)\delta\leftarrow\Pi_{\epsilon}\big(\delta-\alpha,\mathrm{sign}\nabla_\delta\mathcal L\big). An attack succeeds if ρ⋆\rho^\star exceeds the threshold calibrated at 1% FPR at the final step.

AuthorsOpenReview ↗

Response to Reviewer 2DoM [Part 2 (2/2)]

(Continued)

Sensitivity to β\beta. We optimize the joint objectives with β∈0.1,1,10,100\beta\in{0.1,1,10,100} and in the semantic-only limit (β→∞\beta\to\infty, Imprint term dropped), using 50 source–target pairs each, the same ϵ=16/255\epsilon=16/255, and 150 PGD steps.

β\betacanonical angle to xsx_s (final, mean / min)raw CLIP angle (mean)Hamming to msm_sASR r3L7 / r4L9PSNR to xtx_t
0.143° / 25°62°50/50 / 1/5026.8 dB
1 (submission)37° / 19°62°20/50 / 1/5027.0 dB
1031° / 17°62°24/50 / 3/5027.4 dB
10029° / 13°62°25/50 / 7/5027.6 dB
∞\infty (no Imprint term)29° / 14°61°10/50 / 1/5027.6 dB

Findings. (1) The semantic encoder is a residual attack surface in semantic binding methods. A gray-box adaptive attacker can approximate the distance computed based on a semantic encoder through gradient optimization. (2) Semantic binding is highly effective at limiting gray-box attackers’ success rate. Even when the entire budget is dedicated to optimizing the semantic distance, the average semantic angle achieved by adversarial images is 29°, far from the boundary defined by the hyperplane.

Mitigation. This residual attack is an adversarial example against CLIP. Replacing the encoder with an adversarially fine-tuned CLIP (TeCoA [2] / FARE [3]) is a drop-in change within the training-free framework and directly raises the cost of the gray-box attack. We will update the revised manuscript to include content on semantic encoder-based gray-box adaptive attacks.

[1] Müller, Andreas, et al. “Black-box forgery attacks on semantic watermarks for diffusion models.” 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025.

[2] Mao, C., Geng, S., Yang, J., Wang, X., & Vondrick, C. (2023). Understanding zero-shot adversarial robustness for large-scale models. In The Eleventh International Conference on Learning Representations (ICLR 2023).

[3] Schlarmann, C., Singh, N.D., Croce, F. & Hein, M… (2024). Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings for Robust Large Vision-Language Models. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:43685-43704.

3. Replace Proposition 1 with a code-collision lemma. We reduce Proposition 1 to a lemma bounding code collisions, which follows from the SimHash geometry alone, and support forgery resistance with measured quantities instead. The lemma states: if the SimHash hyperplanes are sampled independently of both representations, then for any query image whose centred preprocessed representation is at angle α\alpha to that of the source, Pr⁡[ms∈M(m^f)]≤Pr⁡[Bin(ℓ,α/π)≤max⁡(r,L)]\Pr[m_s\in\mathcal M(\hat m_f)]\le\Pr[\mathrm{Bin}(\ell,\alpha/\pi)\le\max(r,L)]. We do not use the lemma to assert security on its own. It excludes direct testing of the transplanted source pattern whenever the forged image's code differs from the source code by more than the candidate rule allows, and bounds the probability of staying within this limit by Pr⁡[Bin(ℓ,α/π)≤max⁡(r,L)]\Pr[\mathrm{Bin}(\ell,\alpha/\pi)\le\max(r,L)]; whether another candidate's pattern still passes is measured empirically (Appendix I.2). In the revision, the lemma, its proof and the decomposition of forgery success into candidate-set inclusion and pattern acceptance are given in Appendix B (Code-collision Bound).

Best regards,

The Authors

Reviewer_2seYOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IRIS, a training-free watermarking method that derives a keyed Fourier pattern from a generated preview's visual features and embeds it during late denoising steps. Detection compares patterns derived from candidate semantic codes with the inverted image latent. Experiments report reduced forgery success and resilience to several image distortions.

Critical strengths

The problem formulation accurately captures an important challenge: binding a watermark to the generated content while preserving verification under ordinary image distortions.

Critical weaknesses
  1. Content-dependent in-generation watermarking has already been explored by SEAL and SemBind. The paper distinguishes visual binding from caption- or prompt-level binding, but the incremental contribution needs clearer justification. In particular, the evaluation should isolate the benefit of this tighter binding from the changes to embedding timing and the detection statistic.
  2. Stabilizing the semantic code and preserving the embedded pattern under geometric transformations are different requirements. Deriving a pseudorandom pattern from augmented image features does not itself make that pattern geometrically robust. Table 2 reports strong oracle-pattern recovery under the tested conditions, but rotation is limited to 5 degrees and the meaning of "95% cropping" is unclear. These results do not sufficiently establish the claimed geometric resilience or explain which design provides it.
Questions for authors

Which design ensures the pattern's robustness under geometric transformations? Please define the cropping protocol and separately report semantic-code recovery and oracle-pattern detection under stronger rotations and crop-resize transformations.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 2seY [Part 1 (1/2)]

We would like to sincerely thank Reviewer 2seY for providing these insightful and thoughtful comments.

w1: …The paper distinguishes visual binding from caption- or prompt-level binding, but the incremental contribution needs clearer justification. … the evaluation should isolate the benefit of this tighter binding from the changes to embedding timing and the detection statistic.

1. Justification of visual binding distinct from caption- or prompt-level binding

IRIS is distinct from prior work in both (1) design objective: tighter semantic binding, and (2) design components: a robust visual representation, image-specific Fourier-domain mark, later embedding timing, a normalized correlation as mark evidence, and a ratio statistic on a candidate set for detection decision. Together, IRIS achieves a better trade-off between watermark detectability and watermark transplantation resistance, which is indicated by the mean column of Table 1: Main results.

To better justify that the design objective of visual-semantic binding is distinct from caption- or prompt-level binding, we quantify semantic binding tightness by the largest semantic distance between the reference image xrefx_{ref} (from which the semantic for a mark is derived) and the watermarked image xwmx_{wm} (in which the mark is embedded). Ideal semantic binding tightness can distinguish semantic change from non-semantic change, an essential property for preventing a transplanted mark from bypassing verification.

We measure the semantic distance as Δ=1−cos⁡(ϕ(xref),ϕ(xwm))\Delta=1-\cos(\phi(x_{ref}), \phi(x_{wm})), where ϕ\phi is a fixed semantic encoder. We use CLIP ViT-B/32 as the encoder. Over the same 500 DiffusionDB prompts and seeds, the largest CLIP distance is 0.595 for SEAL, 0.506 for SemBind+TR, 0.576 for SemBind+GS, and 0.023 for IRIS. These distances indicate that the detectors need to accept these values; thus, distances for these bindings may exceed them. For reference, two images with different prompts average 0.476, and an image and its JPEG-25 compressed version average 0.058. This indicates that IRIS achieves tighter semantic binding than prior work.

Figure 4 shows visual evidence: the visual content between the reference image (unwatermarked) and the watermarked images is close in IRIS but different in SEAL and SemBind, indicating that IRIS’ visual-semantic binding yields a smaller distance than existing caption- or prompt-level methods.

We added justification for the tightness of semantic binding in Appendix A Discussion of Semantic Binding Tightness in the revision.

AuthorsOpenReview ↗

Response to Reviewer 2seY [Part 1 (2/2)]*

(Continued)

2. Isolating the benefit of tighter semantic binding from embedding timing and detection statistics.

Thank you for this perspective question. IRIS achieve a tight semantic binding through three collaborating components: (1) semantically derived pattern, (2) embedding strategy with strength and timing, and (3) detection strategy with alignment score and candidate searching.

(1) Isolating the benefit of semantic binding from changes to embedding timing.

An ablation study replaces the semantically bound mark with a fixed mark (no semantic binding) and keeps other components unchanged. The results are as follows:

TPR / ASR @1% FPRCleanNoise 0.1JPEG 25Blur 3Imprint ↓Reprompt ↓
IRIS (t0 = 45, visual binding)1.0000.9320.9760.9940.0200.022
t0 = 45, fixed pattern1.0000.9901.0001.0000.5801.000

Findings: Late embedding time alone does not prevent watermark transplantation.

Another aspect is that late embedding timing contributes to tighter semantic binding because it keeps the reference image and the watermarked image visually similar. Our justification of the embedding-timing selection in the response to reviewer XFn9-w2 provides more detail.

(2) Isolating the benefit of semantic binding from detection statistics.

IRIS detection has three levels: (i) the alignment score st(m)s_t(m), the normalized correlation (Eq. 4) between the pattern derived from a semantic code mm and one inverted latent, which quantifies the extent to which a latent may carry the mark; (ii) the detection score s(m)s(m), which averages st(m)s_t(m) over the D=20D=20 inverted latents (Algorithm 2) to aggregate watermark evidence; and (iii) the detection statistic ρ∗=max⁡ms(m)/medianms(m)\rho^*=\max_m s(m) / \mathrm{median}_m s(m) over the candidate set, which is compared against a threshold to decide whether the image carries a valid watermark.

Appendix C.8 ANALYSIS OF THE DETECTION STATISTIC ablates (i) and (iii): the alignment score is replaced by two L1L1 variants, and the ratio statistic by the best candidate score alone, giving six statistics in total. All six detect clean and distorted watermarks, but only the joint use of normalized correlation and the ratio statistic rejects forged images. Embedding a watermark creates a structured energy shift in the latent space, and watermark transplantation makes the target image carry a similar energy shift, leading to a common rise across all candidates. Normalized correlation makes these rises the same scale, and the ratio statistic cancels this common shift because it only asks whether one candidate stands out. For (ii), Section 5.4, Table 3 reports that only using the single signal at step 45 as detection score: T@1%F drops from 1.000/0.932/0.976/0.994 to 0.994/0.862/0.956/0.968 under clean/noise/JPEG/blur, while the forgery ASRs stay at 0.017/0.010 (default 0.020/0.022). Multi-step averaging therefore improves detectability but does not affect forgery resistance. In contrast, the fixed-pattern experiments above indicate that, without semantic binding, detection statistics alone cannot prevent watermark transplantation.

AuthorsOpenReview ↗

Response to Reviewer 2seY [Part 2]

w2 & question: … rotation is limited to 5 degrees and the meaning of “95% cropping” is unclear. … Which design ensures the pattern’s robustness under geometric transformations? Please define the cropping protocol and separately report semantic-code recovery and oracle-pattern detection under stronger rotations and crop-resize transformations.

1. Clarifying typo. We fixed and clarify the notation: throughout the paper, “Crop x%” denotes a crop that removes x% of the image area; the distortion table evaluates Crop 5%. We corrected the mislabelled “95%” in the revision. The design objectives of IRIS do not explicitly include semantic invariance for geometric transformations.

2. Evaluation on strong geometric transformation. The table below shows an evaluation of IRIS under strong rotations and crop transformations. Rotation: 9°–45° about the image centre on the same 512×512 canvas. Crop-resize: a square box at a random position that removes 10%–50% of the image area, resized back to 512×512. Thresholds are set at 1% / 0.1% FPR on 1,500 unwatermarked images. IRIS detects watermarks in geometrically transformed images after registering the images (restoring the rotation angle or image position so the remaining mark aligns with the mark to be detected). For crop-resize, we extract the semantic code from the cropped image and detect the watermark pattern in the registered image.

IRIS under geometric transformation (TPR @ 1% FPR)

TransformationStrengthSemantic code recoveredOracle-pattern detectionIRIS with transform registration TPR
Rotation9 / 18 / 27 / 36 / 45 deg0.955 / 0.900 / 0.790 / 0.775 / 0.7401.000 / 1.000 / 1.000 / 1.000 / 1.0000.955 / 0.900 / 0.790 / 0.780 / 0.745
Crop-resize†10 / 20 / 30 / 40 / 50 % removed0.930 / 0.820 / 0.660 / 0.550 / 0.4351.000 / 1.000 / 1.000 / 1.000 / 1.0000.930 / 0.820 / 0.660 / 0.565 / 0.435

† The registration column uses the deployed detector given the true rotation angle or crop box. For crop-resize the semantic code is read from the un-registered cropped image and the pattern is detected on the registered image.

Findings: The detection rate decreases under strong geometric transformation, mainly because geometric transformations cause semantic changes. The oracle-pattern detection rates remain 1.0 across the various strengths.

In conclusion, we agree that detecting watermarks under geometric transformation is an important research question. In the context of semantic binding, it requires determining the extent to which geometric transformation should be treated as a semantic consistency design mechanism to derive the mark from geometric-transformation-invariant features. We leave it as future work. We added evaluations and discussion of geometric transformation in Appendix in revision.

Best regards,

The Authors

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to all reviewers & Summary of revisions

We sincerely thank all the reviewers for their insightful comments and valuable suggestions, which have helped us improve the paper.

This paper introduces IRIS, a training-free visual-semantic binding watermarking framework for diffusion images. It achieves tighter semantic binding than prior prompt- or caption-level binding and improves the evaluated trade-off between detectability under distortions and resistance to watermark transplantation. We evaluate detectability, forgery resistance, distortion robustness, design ablation, computational efficiency, extension to flow-matching models, and adaptive attacks.

The reviewers expressed positive views on our work. Reviewer 2seY noted that "the problem formulation accurately captures an important challenge: binding a watermark to the generated content while preserving verification under ordinary image distortions." Reviewer 2DoM highlighted our "Effective Defense Against Watermark Transplantation," noting that IRIS addresses "a major security threat" while "dropping forgery ASR significantly," and recognized that it "maintains high visual fidelity and generalizes well to modern flow-matching models." Reviewer XFn9 also described the approach as "reasonable" and noted that "reusing earlier timesteps makes sense to solve the circular dependency."

The reviewers also raised constructive concerns, and we have made every effort to address all of them. Below is a summary of the revision.

  • Standardized robustness benchmark (Reviewers 2DoM and XFn9). We added five WAVES removal and adversarial attacks. IRIS achieves an improved trade-off between robustness to distortions and resistance to watermark transplantation.

    Added → Appx. E.1: WAVES Removal and Adversarial Attacks

  • Evaluation under strong geometric transformations (Reviewer 2seY). We added rotations up to 45° and crop-resize removing up to 50% of the image area. After registration, the decrease in detection closely tracks reduced code recovery.

    Added → Appx. E.2: Strong Geometric Transformations

  • Isolating the benefit of visual-semantic binding from embedding timing and the detection statistic (Reviewer 2seY). We added an the fixed-pattern ablation. Late embedding alone and the detection statistic alone do not prevent transplantation. They together with visual-semantic bound mark reject forged images.

    Added → Appx. D.2: Isolating Visual-Semantic Binding from Embedding Timing and the Detection Statistic

  • Discussion of semantic binding tightness (Reviewer 2seY). We added a discussion that distinguishes visual-semantic binding from caption- or prompt-level binding. For caption- or prompt-level binding, the reference-to-watermarked shift overlaps with the shift of Reprompt forgeries, which motivates IRIS's design objective of narrowing this shift.

    Added → Appx. A: Discussion of Semantic Binding Tightness

  • Ablations of embedding step and preprocessing (Reviewer XFn9). We added these ablations. Among the tested settings, step 45 is the earliest with 1.0 T@1%FPR clean detection and has the highest T@1%FPR under noise. Canonicalization and multi-view averaging improve code stability. Increasing KK from 8 to 16 slightly lowers Hamming distance.

    Added → Appx. D.1: Embedding Step

    Additional results → Appendix C.7: Analysis of Visual Semantic Representations

  • Adaptive adversaries: query models, gray-box objective and loss-weight sensitivity (Reviewer 2DoM). We distinguish image-only observations from detection-oracle access. We explained the gray-box objective construction: combining normalized pattern-transplantation and semantic-matching losses. Across β∈0.1,1,10,100,∞\beta\in \\{0.1,1,10,100,\infty \\}, the public encoder remains an attack surface, but IRIS effectively restricts ASR remains at or below 14% across the tested loss weights.

    Revised → Appx. I: Adaptive Attacks

  • Rigor of Proposition 1 (Reviewer 2DoM). We replaced Proposition 1 with a code-collision lemma for fixed representations and independent isotropic hyperplanes. We use the lemma only to bound candidate inclusion under its assumptions, and assess the remaining contribution to forgery success empirically.

    Revised → Appx. B: Code-collision Bound (Lemma 1)

  • Writing and presentation (Reviewer XFn9). We clarified SemBind's prompt-level binding granularity, scoped the content-agnostic claim in the introduction, and revised the table presentation. We also aligned terminology and experimental descriptions across the appendices and responses.

    Revised → Sec. 1 and all tables

In responses to reviewers, we have added over 200 new experimental results to address the above concerns. The revisions are highlighted in green\textrm{green}.

The reviewers’ valuable suggestions have been extremely helpful in improving our paper. We are happy to answer any further questions.

We look forward to your feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Request of Reviewer's attention and feedback

Dear Reviewer,

As a gentle reminder, the rebuttal period is drawing to a close, and we would be grateful to hear whether our response has addressed your concerns. We look forward to your feedback and would be happy to address any further questions you may have.

Thank you again for your constructive feedback and continued support, which have inspired further improvements to our paper.

Best regards,

The Authors

AuthorsOpenReview ↗

Final Request for Your Feedback Before the Discussion Deadline (<20 hours)

Dear Reviewer XFn9,

We sincerely thank you again for your thoughtful review and constructive suggestions.

With less than 20 hours remaining in the discussion period, we would be extremely grateful if you could kindly review our rebuttal and share your feedback.

In response to your concerns, we have conducted additional robustness evaluations, provided new ablation studies on embedding timing and preprocessing, and substantially improved the paper's writing and presentation. We have incorporated all these improvements into our revised manuscript.

Could you please let us know whether our responses have addressed your concerns? If any questions remain, we would be more than happy to address them promptly before the discussion closes.

Your feedback and support are particularly important to us at this final stage of the discussion. We sincerely hope to hear from you before the deadline.

Thank you again for your valuable time and consideration. We wish you all the best with your research.

Best regards,
The Authors

AuthorsOpenReview ↗

Final Request for Consideration of Our Rebuttal

Dear Area Chair,

With less than two hours remaining in the discussion period, we would like to respectfully seek your consideration of our rebuttal.

We have made every effort to address all three reviewers' concerns, providing detailed point-by-point responses, over 200 additional experimental results, and substantial manuscript revisions. Unfortunately, despite our repeated and respectful reminders, none of the reviewers has responded to our rebuttal.

We fully understand that reviewers are very busy. However, we sincerely hope that our extensive rebuttal and additional evidence can be carefully considered in your independent assessment of the paper's merits and contributions.

We would also be deeply grateful if our proactive efforts to address the reviewers' concerns, together with the absence of follow-up discussion from the reviewers, could be taken into account in your final evaluation.

Thank you very much for your time, understanding, and consideration. We sincerely appreciate your efforts in ensuring a fair and thorough review process.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.