acceptodds
Under review as a conference paper at ICLR 2027

Distortion-Free Semantic Watermarks for Black-Box Generative Models

Abstract

We present a provably distortion-free semantic watermarking algorithm that requires only black-box sampling access to a generative model and black-box access to a semantic embedding. A watermark is called distortion-free if, averaged over its secret detection key, it leaves the generator's output distribution unchanged; this provides a formal guarantee that watermarking does not reduce output quality. Conventional such watermarks are often brittle to transformations. This is especially problematic for images, where benign operations such as compression, cropping, or resizing substantially change the entire pixel representation. To this end, semantic watermarks introduce the requirement that watermark detection should be robust to transformations that preserve content. Previous methods combining these properties rely heavily on specific white-box properties of the underlying model: either to autoregressive text generation or to diffusion-specific access, such as control of the initial noise or diffusion inversion. In contrast, our construction is fully generic: It only interacts with the generative model and the semantic embedding, a function that determines the notion of “preserving content” under which detection should be robust, via sample and query access. We then instantiate this framework for image generation. We characterize the two competing properties required of the semantic embedding: it should remain stable under transformations that preserve an image’s content, but should distinguish independent generations from the same prompt. We train an image encoder directly for this tradeoff. The resulting encoder improves separation while maintaining stability over off-the-shelf alternatives and enables strong aggregate detection while remaining robust to various transformations. We demonstrate the effectiveness of our watermark scheme on both open-source and closed models, which is possible only due to restriction to black-box access.

Primary area:
generative models
Keywords:
distortion-free watermarking, semantic watermarking, generative models, diffusion models, semantic encoder, semantic entropy, AI-generated content detection, Black box watermrking, Black box sampling

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_6XezOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper adapts sampling-based LLM watermarking to propose a black-box, key-averaged distortion-free semantic watermark, applied to text-to-image models. The robustness is achieved by the use of a semantic feature extractor (plus LSH) which is robust to content-preserving image manipulation, while distortion-freeness comes from grouping candidates by LSH code and a keyed selection rule that preserves the 1/N1/N probability. In a nutshell, this is a zero-bit watermarking method (no payload, but a statistical binary detection problem), while for image generation models, an important branch is multi-bit methods (with payload, need to extract the watermark and calculate bit error rate, e.g., Gaussian Shading, but with white-box requirement). With the advantage of only requiring black-box access, one price paid is generating NN candidate outputs per released image (and a larger NN does not necessarily guarantee a better result), while another price or limitation is weak single-image detection. The authors should discuss whether a multi-bit extension is feasible and at what cost in NN.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The idea is new in the image generation regime.
  3. The theoretical justification and experiments are adequate to support the claim.
  4. Limitations are honestly stated (single image, adaptive attack).
Critical weaknesses
  1. The core mechanism combines bundle sampling, black-box sample-based watermarking, and LSH. Beyond the change of modality and a fine-tuned encoder, the technical difference is unclear.

  2. The threat model is unclear, which leads to uncertainty on how the method is deployed (see question 2 for detail).

  3. The scheme is zero-bit (no payload), single-image detection is weak, and each image costs NN generations.

Questions for authors
  1. The method requires discretizing the semantic embedding into codes, and the authors adopt LSH without justifying it over other binarization alternatives (there are many). This choice will potentially affect Theorem 4 and it is worth ablation.

  2. The threat model is not clarified. The paper does not state who deploys the scheme (model owner, API service, or end user), who holds the key κ\kappa, who performs detection, or whether EE is public or secret. If the owner deploys it, it makes much less sense to treat the model as a black box. Instead, the owner should have white-box access and be motivated to leverage it to design a stronger watermark with the best fidelity, rather than restricting itself to black-box access in the first place.

  3. The authors should discuss whether a multi-bit extension is feasible and at what cost in NN.

Reviewer_XtC9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a distortion-free semantic watermarking method for black-box generative models. Instead of modifying the generated image, the method generates multiple candidates and selects one based on a score dependent on a secret key. The authors also train a semantic encoder to distinguish different generations while remaining robust to content-preserving transformations.

Critical strengths
  • I find the idea interesting and novel in the image watermarking setting, especially since it only requires black-box access to the generator.
Critical weaknesses
  • My main concern is the practical detection setting. The method relies heavily on aggregating multiple images. But how would this work in realistic attribution scenarios? The authors evaluate multiple images generated under the same key, but what happens when images come from different keys? What if a batch contains both watermarked and unwatermarked images, or images from multiple sources? Can it distinguish between different sources? Additionally, it requires generating multiple complete images to release one, and reliable detection requires aggregating several released images. This seems expensive for a practical watermarking system.

  • I am not fully convinced by the semantic aspect of the method. The authors train the encoder to distinguish independently generated images from the same prompt while keeping edited versions of an image close together. But these objectives seem to conflict in some cases. I would like to see a clearer definition and evaluation of what counts as preserving semantic content.

  • The evaluated generative edits are constrained to preserve the main subject and scene. What happens if someone changes the main content but the watermark remains detectable? I would also like to see attacks that try to remove the watermark while preserving the original content. I think these attacks are important. The authors acknowledge the lack of adaptive attack evaluation, but I still see it as an important limitation.

Questions for authors

I find the proposed approach interesting. I would appreciate clarification on the points raised in the weaknesses section. This could change my overall assessment.

Additional question:

  • The paper claims a general framework for generative models, but all watermarking experiments are on images. I understand the sampling procedure is model-agnostic, but the semantic encoder seems central to the method working. How well would this actually generalize to other modalities?
2 comments

Thank you for the detailed rebuttal. I find the main contribution interesting. The rebuttal has addressed most of my concerns and made the contribution more convincing. I still think adaptive robustness is an important limitation, but I appreciate that the authors acknowledge it.

Overall, I view the paper more positively now and am happy to raise my score.

Thank you for following up and for updating your rating.

As discussed, we added an adaptive removal experiment with a public semantic encoder (Appendix G.11). At 20% nominal FPR, encoder-guided edits removed detection in 65.5% of initially detected images, compared with 19.4% for random edits from the same screened pool.

Reviewer_XQnrOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a distortion-free semantic watermark that needs only sampling access to the generator and query access to a semantic encoder. For each prompt, N candidate images are generated, embedded, and hashed to M-bit codes with random-hyperplane LSH. A secret key assigns each code a pseudorandom score, and the sampler favors high-scoring candidates. Because collisions among candidate codes are accounted for, every candidate is selected with probability exactly 1/N on average over keys, so the output distribution is unchanged. Detection reads the keyed score of a suspect image’s code and averages scores over a batch of images. The paper proves distortion-freeness, an expected score gap that grows with the number of distinct codes, and a TPR ≤ N·FPR ceiling for any black-box sampler.

Critical strengths
  1. The framework is general and theoretically grounded. It is distortion-free, semantic and fully black-box, so it applies to closed APIs, where most image watermarks cannot be used. The proofs are clean. The concept of semantic entropy and the TPR ≤ N·FPR ceiling clarify what any black-box sampling watermark can and cannot achieve.
  2. The encoder analysis is careful and well motivated. The paper explains why standard encoders are a poor fit, since they map same-prompt generations close together. The trained encoder improves code diversity on SD2.1 and two FLUX variants not used in training (Table 3), and it gives the best end-to-end AUROC in the encoder ablation (Appendix G.2). The paper is also honest about its limits.
Critical weaknesses
  1. The most relevant baselines are missing. The compared methods (Tree-Ring, Gaussian Shading, SEAL) need noise access and inversion, so they do not share the paper’s access model. They also dominate clean single-image detection: 100% TPR at 1% FPR, against 12.2% for the best configuration here (Table 6). The comparisons that matter for this setting are absent:
  • Post-hoc watermarks such as TrustMark, VINE or StegaStamp. These also need only the output and work on any closed API. The paper never shows when its method is preferable to watermarking the API output after generation.
  • PRC (Gunn et al., 2025), the closest distortion-free and undetectable image watermark. It is discussed but not compared.
  • SemBind, T2SMark, SLICE and NoisePrints, all cited as related work. SemBind is the natural comparison for the claimed advantage against reprompting forgery.
  1. The robustness evaluation does not reflect realistic operating points or attacks.
  • All robustness results are at a 20% clean FPR. Under geometric edits, TPR is about 45% against an FPR of about 20% (Table 8), which is a weak separation. No results are given at 1% or 5% FPR, or for batches under edits.
  • Regeneration attacks (VAE compression, diffusion regeneration, CtrlRegen) are not evaluated. A semantic watermark could in principle survive them better than noise-based marks, so this would be the method’s strongest case, and it is missing.
  • Adaptive attacks are not evaluated, as Appendix H.4 acknowledges. With only 4 to 12 code bits, an attacker may be able to push an embedding across an LSH boundary, or keep a high-scoring code while changing content.
  • The generative edit set has only 80 images.
  1. The quality, practicality and API claims are weakly supported.
  • The fixed-key FID comparison (4.71 to 5.19 against 5.42 to 6.88) has no clean-versus-clean FID floor on 10,000 images. Without it, it is unclear whether these differences are meaningful.
  • Each released image needs up to 32 full generations, but no wall-clock cost is reported.
  • The GPT Image experiment uses only 50 prompts. For B=5 there are only ten clean batches, which the authors note cannot support low FPR estimates.
  • Batch detection assumes B independently generated images from one source, which should be stated more clearly as a deployment assumption.
Questions for authors
  1. How does the method compare with a post-hoc watermark (for example TrustMark or VINE) applied to the same black-box outputs, in quality, single-image detection and robustness?
  2. How does it compare with PRC and SemBind under the same generator?
  3. What are the robustness results at 1% and 5% FPR, and for batches after edits?
  4. How does the watermark perform under VAE and diffusion regeneration attacks?
  5. What is the FID between two independent clean sets of the same size, so the fixed-key FID values can be interpreted?
  6. What is the wall-clock cost per released image for N=32?
2 comments

Thank you for reviewing our paper. Changes to the revised manuscript are highlighted in blue. We address your questions in two comments.

How does the method compare with a post-hoc watermark (for example TrustMark or VINE) applied to the same black-box outputs, in quality, single-image detection and robustness?

How does it compare with PRC and SemBind under the same generator?

We stress that our method differs qualitatively from these watermarking schemes. Post-hoc methods such as TrustMark and VINE embed the watermark by modifying the generated image, which can change the output distribution. Our method samples (N) candidates and releases one unchanged, preserving the generator’s output distribution when averaged over keys (Theorem 3). The adoption of SynthID-Text in Gemini and a version of it in Claude illustrates the practical interest in distribution-preserving watermarking (Dathathri et al., 2024; Anthropic, 2026).

Among the image watermarking methods discussed here, ours combines selection among complete samples with a distribution-preservation guarantee using only sampling access. PRC and SemBind require access to the generator’s initial diffusion latents. This matters for services such as Runway when they offer third-party generators through APIs that return only completed images. Our method uses sampling access alone and releases the selected image unchanged. For comparison on a common Stable Diffusion 2.1 backbone, we chose Tree-Ring as a representative initial-noise watermark, Gaussian Shading for its distortion-free construction, and SEAL for its semantic watermarking approach. We evaluate each with its native generation and detection procedures.

What are the robustness results at 1% and 5% FPR, and for batches after edits?

We added post-edit detection results at 1% and 5% nominal clean FPR, including results for batches of 5 and 10 edited images (Appendix F.2). We interpret the detection score of a particular output as statistical evidence rather than a “watermarked” or “clean” bit. While a single score may provide limited evidence, aggregating scores across multiple outputs can provide a much stronger detection signal.

For example, at 5% nominal clean FPR, the single-image TPR for (N,M)=(32,12) went from 44.9% before editing to 40.4% after appearance changes and compression. Averaging scores across ten images subjected to the same transformation increased post-edit TPR to 92.3% at the same nominal FPR.

How does the watermark perform under VAE and diffusion regeneration attacks?

Following your question, we evaluated our watermark under regeneration attacks. We used the same images as in our robustness experiments: 200 COCO-2017 prompts generated with Stable Diffusion 2.1. We applied four attack families:

  • Learned compression (Zhao et al., 2024): the models of Ballé et al. (2018) and Cheng et al. (2020) from CompressAI, at quality levels 1, 3 and 6.
  • VAE round trips: encoding and decoding with the Stable Diffusion 2.1 VAE and with the SDXL VAE.
  • Diffusion regeneration: SDEdit with Stable Diffusion 2.1 and an empty prompt, at noise strengths 0.1, 0.2 and 0.4 (Zhao et al., 2024).
  • CtrlRegen+ (Liu et al., 2025): the released weights and settings, at noising steps 0.5 and 1.0.

All detectors use thresholds frozen on unattacked clean images at 20% FPR. We report DINOv2 similarity to show how much each attack changes the image. For comparison, we also evaluated SEAL, the semantic baseline, which derives its key from the image caption rather than from a visual representation.

TPR / paired retention at 20% FPR (%). Our results average 25 key replicates.

Attack            DINOv2 cos  Ours (8,4)   Ours (16,8)  Ours (32,12)  SEAL
None (clean TPR)  –           49.6         74.7         85.0          100.0
Ballé et al. q=1  0.89        40.5 / 76.8  54.8 / 70.4  55.4 / 63.7   78.0 / 78.0
Cheng et al. q=1  0.90        40.1 / 76.3  56.1 / 72.4  59.0 / 68.0   82.5 / 82.5
SD 2.1 VAE        1.00        48.3 / 96.4  73.4 / 97.5  83.2 / 97.5   100.0 / 100.0
SDXL VAE          0.99        46.9 / 93.2  71.6 / 94.6  80.6 / 94.1   99.5 / 99.5
SDEdit 0.2        0.96        42.4 / 82.4  61.5 / 79.7  64.6 / 74.6   63.5 / 63.5
SDEdit 0.4        0.90        39.4 / 74.3  52.0 / 66.5  52.4 / 59.9   25.0 / 25.0
CtrlRegen+ 0.5    0.95        41.3 / 79.5  59.2 / 76.8  62.0 / 71.5   50.0 / 50.0
CtrlRegen+ 1.0    0.92        39.4 / 75.1  54.0 / 69.4  56.0 / 64.0   28.0 / 28.0

What is the FID between two independent clean sets of the same size, so the fixed-key FID values can be interpreted?

Thank you for this suggestion. The FID between two independent sets of 10,000 unwatermarked generations is 4.64. We added this clean-versus-clean reference to the revision (Table 1) and report relative FID, where 1.000 denotes the clean-versus-clean level. Our three configurations reach a relative FID of 1.017–1.120, compared with 1.170–1.483 for the baselines

What is the wall-clock cost per released image for N=32?

Following your question, we measured the cost on one NVIDIA L40S with SD2.1-base (50 DPM-Solver++ steps, 512×512, fp16), averaging over 20 prompts after warm-up. As expected, the cost of releasing a watermarked image grows linearly with N . Although generation is more expensive, our detection is much cheaper than the baselines’, requiring a single encoder forward pass (≈8.5 ms per image) instead of a 50-step diffusion inversion (≈0.8–2 s). We discuss this limitation in Section 8 and Appendix H.2. A promising direction for diffusion models is to select among candidates before denoising finishes, which would avoid generating all N images in full. We leave this for future work.

Method            Generation (s)  Detection (s)  Total (s)
No watermark      0.99            —              0.99
Ours, N=8         6.20            0.0085         6.21
Ours, N=16        12.80           0.0085         12.81
Ours, N=32        26.05           0.0085         26.06
Tree-Ring         4.35            3.46           7.81
Gaussian Shading  4.37            3.51           7.88
SEAL              2.66            2.05           4.71

We would be happy to answer any further questions.

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

This paper proposes a distortion-free semantic watermarking method that requires only black-box sampling access to a generative model and query access to a semantic encoder. For each prompt, it generates N candidate images, converts them into short binary codes using the encoder and locality-sensitive hashing, and selects one based on a secret-keyed score. Experiments on Stable Diffusion evaluate detection, robustness to image edits, image quality, and output diversity. The paper also demonstrates the method on the GPT Image API using only sampling access.

Critical strengths
  1. The method treats both the generator and the encoder as black-box oracles without modifying the generation process. This design enables watermarking through a closed commercial API.
  2. The paper makes a clear and useful point that distortion-free semantic watermarking needs variation among same-prompt generations in the encoder space and backs it with an encoder trained for this tradeoff.
  3. The paper provides theoretical guarantees for the proposed method.
Critical weaknesses

W1. Each watermarked image requires generating N candidates, while reliable detection requires a batch of images from the same source. In practice, a detector may receive only one image at a time, without a reliable way to group images by source. The paper treats batch detection as an evaluation setting rather than a requirement, and the assumed source attribution is not addressed.

W2. Batch thresholds are calibrated empirically, with most results reported at a 20% false-positive rate (FPR). The theoretical guarantees do not provide finite-sample error bounds under a fixed key. Unlike the compared baselines, the method lacks a practical way to set thresholds with guaranteed false-positive control.

W3. The experiments use batches that are either fully watermarked or fully clean, with disjoint prompts. In practice, batches may contain mixed and correlated images, with an unknown proportion of watermarked images. Under a fixed key, clean images in high-scoring regions of the encoder space may be consistently flagged, leading to systematic false positives.

W4. The distortion-free guarantee holds only when averaged over keys, while actual deployment uses a fixed key. The FID results in Table 1 lack a clean-versus-clean reference, making it difficult to distinguish distribution shifts from sampling variance. Same-prompt diversity is measured in DINOv2 space, which the paper itself shows is insensitive to the variation exploited by the watermark. The consistent FID improvement with increasing N in Table 10 also suggests a possible distribution shift under a fixed key.

Questions for authors

Q1. How practical is the method when only individual images are available and images cannot be reliably grouped by source?

Q2. How should the detection threshold be set to guarantee a target false-positive rate under a fixed key and finite batch size?

Q3. How does batch detection perform on mixed and correlated images? Could a fixed key cause systematic false positives for certain image distributions?

Q4. Can the authors provide clean-versus-clean FID comparisons and evaluate diversity in the watermark encoder's space to better validate distortion-freeness under a fixed key?

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

We thank all reviewers for their valuable feedback.

We introduce a distortion-free, sampling-based watermarking method for generative models and evaluate it primarily on image generation. To our knowledge, this is the first watermarking method that uses any model as merely a black-box sampler fully generically and any definition of “semantics” fully generically as a black-box access to an arbitrary “semantic embedder”. As we require only black-box access to the generator and a semantic encoder, we are fully agnostic to specific architectures as well as varying interpretations of the “semantics” or “meaning” that the watermark should be robust with respect to. Averaged over the secret key, our method preserves the generator’s output distribution.

We are glad that the reviewers recognized the novelty of our black-box approach (fnTE, XQnr, XtC9, 6Xez), appreciated its theoretical guarantees and clean proofs (fnTE, XQnr, 6Xez), and highlighted how our analysis improves the understanding of the fundamental capabilities and limitations of black-box watermarking (XQnr). Reviewers also appreciated our analysis of semantic diversity and the effectiveness of our trained encoder (fnTE, XQnr).

Based on the reviewers’ feedback, we uploaded a revised version of the paper. For convenience, we highlighted all new and modified sections in blue.

Below, we summarize the main clarifications and additions in the revision.

(A) Practical detection and mixed-source batches

Some reviewers raised concerns about the need to aggregate multiple images and the assumption that all images in a batch carry the same watermark.

We clarified that our detector produces a statistical score rather than a definitive binary decision for each image. Aggregating scores across multiple outputs increases detection reliability, similarly to aggregating evidence across tokens in language-model watermarking.

We also added experiments with batches containing both watermarked and unwatermarked images, as well as images carrying different secret keys (Appendix G.6). Our results demonstrate that mixed-source batches retain a meaningful detection signal, with detection strength depending on the fraction of images carrying the tested watermark.

(B) Semantic similarity and robustness

By semantic similarity, we mean preserving the content of a particular image. As intended, our watermark remains detectable after content-preserving edits ( like jpeg), and its signal drops when an edit changes the content, for example by replacing the main subject. The evaluated baselines retained a similar signal after both types of edits (Appendix G.9).

(C) Additional experiments

We compared alternative semantic encoders and discussed how the choice of encoder affects robustness. We also compared discretization methods and evaluated our approach on text-to-audio generation.

We address the remaining points in our individual replies. We look forward to hearing from the reviewers and would be happy to address any remaining questions during the discussion period.

Area_Chair_Bm8NOpenReview ↗

Please, start a discussion

Dear reviewers,

The discussion period is about to end and the authors have responded to your reviews. This is a friendly reminder to check these responses and respond if necessary.

@XtC9: Thanks for acknowledging the rebuttal.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.