betaFeedback

WorksComputer Science

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 change the entire pixel representation. Semantic watermarks therefore require detection to be robust to content-preserving transformations. 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.

Introduction

Generative models can produce large amounts of text and images for which the original source is difficult to determine. Watermarking addresses this problem by embedding a signal that can later be detected using a secret key. Such a signal can be used to reliably attribute content to a particular model or service.

A watermark should not noticeably degrade the model’s output. Christ et al. [11] introduced a theoretical framework to guarantee the preservation of quality via the notion of cryptographic undetectability: without a (secret) watermark detection key, no efficient observer should be able to distinguish watermarked outputs from ordinary model outputs; in particular, no benchmark may show any degradation. A weaker variant of the above notion is distortion-freeness, under which the marginal distribution of a (single) watermarked output, averaged over the secret key, is identical to the original output distribution. This approach was subsequently adopted in practical systems. DeepMind introduced a configurable distortion-free version of SynthID-Text and deployed it in Gemini ([12]). Anthropic later adopted a version of SynthID-Text for Claude ([1]). The resulting deployments illustrate an important advantage of distortion-free watermarking: the quality-preservation guarantee circumvents reevaluation for every benchmark or class of prompts. Distribution preservation, however, does not guarantee another ideal property of watermarking: robustness to edits. A watermark may be detectable in the original output but disappear after the output is slightly modified. The problem with editing is highly relevant for images. Even mundane operations such as JPEG compression, cropping, resizing, or applying a filter can change essentially the entire pixel representation of the image. A watermark tied directly to that pixel representation may therefore disappear even though the content of the image remains the same. This motivates semantic watermarking, in which detection should be robust to transformations that preserve the content represented by the image.

Watermarking the semantic meaning of a model’s output is also arguably far more robust to watermark removal attempts. Usually, some of the semantic components of an output are requested in the prompt, but some of them are randomly chosen by the model during generation. An adversary that only sees the output does not know which semantic components were inherent to the prompt and which were not, so they should not be able to remove the model’s semantic choices without possibly degrading the intended meaning of the output substantially. This comes in sharp contrast to syntactic attacks, such as paraphrasing of sorts, that completely change the syntax without hurting the intended meaning of the output.

Existing semantic watermarking methods typically rely on specific “white-box” properties of the underlying generation process. For text, this is less restrictive because many models share an autoregressive token-by-token interface. Image watermarks, on the other hand, often require control over a particular diffusion latent space or noise distribution, as well as a model-specific inversion procedure. Such access is unavailable through closed image-generation APIs and any change in the generation architecture may render the watermark scheme inapplicable.

We introduce the first watermarking framework that is simultaneously distortion-free, semantically robust, and uses only black-box access with respect to both the generative model and the semantic embedding, making it applicable to general generative models. We access the generative model only as an oracle that returns independent samples and the semantic embedding only as an oracle that maps a completed output to a vector, without assuming how either oracle is implemented.

Our algorithm samples several candidate outputs, maps them to semantic codes depending on their semantic embeddings, and selects one using a secret keyed function. By accounting for the number of candidates that received each code, we ensure that the selected output has exactly the generator’s original marginal distribution when averaged over the key and the algorithm’s randomness. This holds for every generator and semantic embedding. If an edit preserves the semantic representation, the watermark is likely to survive (Appendix § 17). We prove these guarantees for any two oracles, without requiring a particular embedding architecture or model architecture. This generality comes at a cost: single-output detection has a non-negligible false-positive rate and limited true-positive rate. This limitation is inherent to black-box output selection: selecting among NN independent candidates gives single-output detection rates satisfying TPR ≤N\le N FPR (Appendix § 22.§ 1). We therefore focus on settings where multiple outputs can be attributed to the same source, such as images uploaded by one social-media profile or assignments submitted by one student. Aggregating the watermark detection scores across this batch of outputs enables high-confidence detection.

To apply our framework to images, we need a semantic embedding that distinguishes different images generated from the same prompt, while remaining stable under edits that preserve an image’s content. Standard semantic encoders may intentionally treat such independent generations from the same prompt as similar because they depict the same high-level concept. For watermarking, however, we need to observe the variation among the generated outputs that remains distinguishable after editing. Otherwise, selecting one output rather than another cannot introduce a detectable semantic watermark. This requires randomness in the meaning of the generated content itself, beyond randomness in its pixel representation; we call this semantic entropy and derive a lower bound on our algorithm’s detection probability in terms of this entropy. We train an image encoder specifically to distinguish outputs generated by the same model from the same prompt using different random seeds, while keeping each image close to edited versions of it. We compare the resulting encoder with existing image embeddings, evaluating both its stability under content-preserving transformations and its ability to distinguish independent generations across several image-generation models.

We evaluate detection from individual images and batches, robustness to common transformations, image quality, and same-prompt diversity. With 32 candidates, we obtain an AUROC of 0.920 for one image and 1.000 for ten. Even when reusing a fixed key, our method achieves the lowest FID relative to unwatermarked generations among the evaluated methods, while retaining essentially the same Vendi diversity. We also apply our framework to OpenAI’s closed GPT Image API, using only sampling access, and obtain an AUROC of 0.978 for batches of five images with eight candidates per output.

Our main contributions are therefore:

  1. We give a black-box, distortion-free semantic watermarking framework for general generative models, requiring only sampling access to the generator and a semantic-embedding oracle.

  2. We characterize how the semantic encoder affects detection. We use semantic entropy to quantify the encoder’s ability to distinguish independent generations from the same prompt, and prove a lower bound on the true-positive rate that increases with semantic entropy.

  3. For diffusion-image generation, we train an encoder and evaluate image quality, same-prompt diversity, detection, and robustness to content-preserving transformations.

Related works

Early language-model watermarks embedded statistical signals during autoregressive generation. Kirchenbauer et al. [28] promoted a keyed subset of tokens during decoding. Later work strengthened the distributional guarantees: Christ et al. [11] formalized cryptographic undetectability, Vaikuntanathan & Zamir [41] introduced bundle sampling that exactly preserves the output distribution, Kuditipudi et al. [29] used Gumbel-based sampling, and Bahri & Wieting [3] obtained distortion-free watermarking using only sampled text, and observed that shared meanings among candidates may weaken the watermark. MUSE selects among samples from black-box tabular generators but is not distortion-free [13].

Preserving the output distribution does not ensure robustness to edits. Robust text schemes withstand substitutions and paraphrasing, including a construction based on pseudorandom error-correcting codes [29, 11]. Semantic methods make the signal depend on meaning: SemStamp uses locality-sensitive hashing, k-SemStamp uses clustering, and PMark considers multiple channel constraints (Hou et al., 2024a;b; Huo et al., 2026). These methods exploit autoregressive decisions; ours scores outputs, aggregating evidence across independent generations.

Image watermarking has largely relied on diffusion internals. Tree-Ring inserts Fourier patterns into the initial noise, while Gaussian Shading maps watermark information to Gaussian latents [43] (Yang et al. 2024b). Gunn et al.’s image construction uses pseudorandom error-correcting codes for multi-query undetectability and distinguishes Gaussian Shading’s single-image guarantee from fixed-key reuse, where they report reduced same-prompt diversity [17]. T2SMark adds per-generation randomness to relax fixed-codeword restrictions (Yang et al., 2025). These methods require access to the initial noise, which is typically unavailable through image-generation APIs, and diffusion inversion for detection. NoisePrints avoids inversion but requires the generation seed [16]. We instead select among completed outputs.

Images also make the motivation for semantic watermarking clear: a watermark should survive pixel-level changes that preserve content, but not changes to the content itself. Reprompting attacks expose the latter risk by generating different content that retains the original watermark [32]. SEAL encodes caption-derived information in the initial noise [2], although captions may omit appearance or composition. SemBind learns prompt-consistent codes (Zhang et al., 2026), while SPLICE separates semantic factors [14]. We instead use direct image representations that remain stable under content-preserving edits yet distinguish independent same-prompt generations, without modifying or inverting generation.

Preliminaries

In this section, we introduce the notation used by our method, with additional formal definitions and discussion provided in Appendix § 12. Let G:P×S→XG:\mathcal{P} \times\mathcal{S} \to\mathcal{X} be a conditional generator, where p∈Pp \in\mathcal{P} is a prompt, s∈S={0,1}ks \in\mathcal{S} = \{0,1\}^{k} is a seed, and X\mathcal{X} is the output space. For a fixed prompt, we write X=G(p,S)X = G(p,S) for S∼Unif⁡(S)S \sim\operatorname{Unif}(\mathcal{S}), with distribution PG(⋅∣p)P_G(\cdot\mid p). A semantic encoder E:X→RdE:\mathcal{X} \to\mathbb{R}^{d} maps an output to a semantic representation. For a nonzero representation, let

z(x)=E(x)∥E(x)∥2,dE(x,x′)=1−⟨z(x),z(x′)⟩.z(x) = \frac{E(x)}{\lVert E(x)\rVert_2}, \qquad d_E(x,x') = 1 - \langle z(x), z(x')\rangle.

We use an MM-bit random-hyperplane locality-sensitive hash HR:Rd→CH_R : \mathbb{R}^d \to C, where C={0,1}MC = \{0,1\}^M, which partitions the representation space into regions identified by binary codes. Here, RR denotes the sampled hyperplane normals. The binary representation also provides a natural notion of proximity between hash regions: for c,c′∈Cc,c' \in C, the Hamming distance dH(c,c′)d_H(c,c') is the number of bits on which they differ. The parameter rhamr_{\mathrm{ham}} denotes the maximum number of boundary-uncertain bits considered during detection (§ 5). For a key κ∈K\kappa\in\mathcal{K}, where K\mathcal{K} is the secret-key space, a keyed pseudorandom function Fκ:C→[0,1]F_\kappa: C \to[0,1] assigns each hash code a secret score.

Semantic Encoders for Distortion-Free Watermarking

Distortion-freeness imposes an inherent entropy requirement. If a prompt deterministically produces one output, any distortion-free scheme must return it, making watermarked and nonwatermarked outputs indistinguishable. More generally, Christ, Gunn, and Zamir showed that the detection power of distortion-free watermarking is necessarily tied to generation entropy [11].

Semantic robustness creates a stronger limitation: a generator may have high output entropy while producing outputs equivalent under the chosen semantics. For example, the same scene viewed from slightly different viewpoints. A robust semantic representation should map them to one semantic region, leaving no variation from which to introduce a detectable signal. We therefore require sufficient variation in the semantic representations of outputs sampled from the same model and prompt. We refer to this variation as semantic entropy. The definition below quantifies this variation at a chosen semantic resolution.

The semantic embedding determines both detection power and robustness. Detection power grows with the semantic entropy of the generated outputs (Theorem 4), and robustness depends on whether an edit keeps an image close to its original representation in the semantic space (Appendix § 17). Our framework is not specific to images or to a particular embedding architecture. For our image instantiation, however, the embedding must satisfy two complementary requirements: an image and its content-preserving transformations should remain close in the embedding space, while independent same-prompt generations should remain distinguishable.

Black-box encoder access also lets users tailor what counts as a content-preserving transformation. For example, a layout-based representation may favor robustness to style changes, whereas a style-based representation may favor robustness to spatial transformations. Appendix § 21.§ 4 demonstrates these contrasting behaviors across three style edits and three spatial transformations, without changing the watermarking algorithm.

However, the chosen representation must also distinguish independent same-prompt generations. Most semantic encoders are designed to separate images associated with different concepts or prompts, rather than different generations of the same prompt. Yet changing the random seed can substantially alter content and composition even when the prompt remains fixed (Xu et al., 2025). We therefore train the encoder to capture these instance-level differences.

Figure 3 and Figure 4 in Appendix § 18 illustrate the representation requirements and training pairs: transformations of one image should preserve its representation, whereas an independent same-prompt generation may receive a different representation. Existing representations are not explicitly optimized for both requirements: CLIP emphasizes image–text alignment, DINOv2 invariance across augmented views, and TIPS perceptual similarity. We therefore fine-tune DINOv2 with a contrastive objective that treats content-preserving transformations as positives and different-seed generations from the same prompt as negatives. Appendix § 18 provides architecture and training details.

The two requirements of a semantic encoder

Figure 3. The two requirements of a semantic encoder. (a) An SDXL image, content-preserving transformations of that image, and an independent generation from the same prompt. (b) Mean cosine distance between different-seed generations divided by mean cosine distance under transformations. Higher values indicate stronger separation relative to transformation sensitivity.

Training-pair construction using SDXL generations

Figure 4. Training-pair construction using SDXL generations for the prompt the feeling of writer’s block. (a) An image and a semantics-preserving transformation form a positive pair and are pulled together. (b) Independent generations from the same prompt but different seeds 2 and 5 form a negative pair and are pushed apart.

Unlike exact-output entropy, which treats every distinct image as a different outcome, semantic entropy measures how diverse independent same-prompt generations are in the encoder space. For a fixed semantic radius ε\varepsilon, semantic entropy is high when two independent generations are unlikely to lie within distance ε\varepsilon of one another.

Definition 1 (Semantic entropy). Fix ε∈[0,2]\varepsilon\in[0,2], and let X,X′X, X' ∼i.i.d.\overset{\mathrm{i.i.d.}}{\sim} PG(⋅∣p)P_G(\cdot\mid p) be two independent generations for prompt pp. We define

HsemE,ε(G∣p)=−log⁡Pr⁡X,X′[dE(X,X′)≤ε].H_{\mathrm{sem}}^{E,\varepsilon}(G \mid p) = -\log\Pr_{X,X'}[d_E(X,X') \le\varepsilon].

The choice of ε\varepsilon determines what counts as meaningfully different. If ε\varepsilon is very small, semantic entropy can be high simply because two generations are rarely almost identical, even when they remain semantically very similar. A larger ε\varepsilon gives a stronger test: high semantic entropy then means that independent generations remain separated even at a meaningful semantic radius.

Because enumerating all seeds is impractical, given a prompt pp, we estimate semantic entropy from nn independent samples X1,…,Xn∼PG(⋅∣p)X_1,\ldots,X_n \sim P_G(\cdot\mid p) and report H^semE,ε\widehat{H}_{\mathrm{sem}}^{E,\varepsilon}. We estimate the probability inside the logarithm by the fraction of pairs whose distance is at most ε\varepsilon:

P^ε=(n2)−1∑1≤i<j≤n1{dE(Xi,Xj)≤ε},H^semE,ε=−log⁡P^ε.\widehat{P}_{\varepsilon} = \binom{n}{2}^{-1} \sum_{1 \le i < j \le n} \mathbf{1}\{d_E(X_i,X_j) \le\varepsilon\}, \qquad\widehat{H}_{\mathrm{sem}}^{E,\varepsilon} = -\log\widehat{P}_{\varepsilon}.

Appendix § 19 reports how estimated semantic entropy varies with ε\varepsilon. Across the evaluated semantic radii 0.3≤ε≤10.3 \le\varepsilon\le1, our encoder achieves the highest estimated semantic entropy among the evaluated encoders for every tested generator.

Semantic entropy captures the representation-level variability available for watermarking. The next section shows how it yields a detectable signal while preserving the generator’s output distribution.

Method

We watermark a black-box generator by selecting which of its outputs to return. Given a prompt, we independently sample NN candidate images, embed each image using the semantic encoder EE, and map its normalized embedding z(x)z(x) to a semantic region c=HR(z(x))c = H_R(z(x)) using a fixed LSH function HRH_R. For a chosen code length MM, our scheme samples MM independent uniform unit hyperplane normals RR once, independently of the secret key, and reuses the resulting HRH_R for generation and detection. We assign each semantic region a pseudorandom score Fκ(c)F_\kappa(c), determined by a secret key κ\kappa that we reuse across generations and during detection. We then use these scores to select the output image. By assigning the same score to images in the same semantic region, we aim to make the watermark robust to non-semantic transformations. In particular, when a transformation preserves an image’s semantic region, it also preserves its score and therefore the detector’s decision.

The pseudorandom scoring function allows us to construct a distortion-free selection rule. When all NN candidates belong to distinct semantic regions, each region has an equal probability of receiving the highest score over the random secret key, under the ideal-PRF assumption (Assumption 17). Thus, selecting the highest-scoring candidate gives each candidate probability 1/N1/N. As we prove in Appendix § 14, this uniform-selection property is sufficient to guarantee distortion-freeness after averaging over the secret key. However, multiple candidates may share the same semantic region, breaking this simple argument. We first address this problem with representative sampling (Algorithm 3), an intuitive construction inspired by bundle sampling that selects one representative uniformly from each semantic region and compensates for the remaining candidates through an additional random-selection step. We describe the full construction in Appendix § 13.

Representative sampling is not the only distortion-free selection rule. We therefore compare several sampling rules in Appendix § 21.§ 3 and find that Gumbel-based selection (Kuditipudi et al., [29]) achieves the strongest empirical detection performance among the methods we evaluate. We consequently replace representative sampling with the Gumbel-based sampler in our final generation algorithm. Algorithm 1 and Figure 1 illustrate this procedure, and Appendix § 14 proves the distortion-freeness of both sampling rules.

Method overview showing watermarked generation, encoding and hashing, grouping by code, and distortion-free selection

Figure 1. Method overview. Candidates are encoded, hashed, and grouped by semantic code; each occupied code cc has a keyed score Fκ(c)F_{\kappa}(c) and ncn_c members. Gumbel selection (Algorithm 1) returns a uniformly random member of the code maximizing log⁡Fκ(c)/nc\log F_{\kappa}(c)/n_c. Averaged over keys, code cc wins with probability nc/Nn_c/N, so each candidate is returned with probability 1/N1/N and the output follows the generator’s distribution (Theorem 3). With a fixed key, the preference for particular codes supplies the signal the detector tests. Other distortion-free rules, such as representative sampling in Appendix § 13, can replace the selection stage.

Algorithm 1 Watermarked generation with Gumbel selection

Input: Prompt pp and number of candidates NN

Parameters: Generator GG, encoder EE, and secret key κ\kappa

  1. for i=1,…,Ni=1,\ldots,N do

  2. Draw Si∼Unif⁡(S)S_i \sim\operatorname{Unif}(\mathcal{S}) independently; set xi←G(p,Si)x_i \leftarrow G(p,S_i)

  3. Set ci←HR(z(xi))c_i \leftarrow H_R(z(x_i))

  4. end for

  5. CN←{ci:i∈[N]}C_N \leftarrow\{c_i:i\in[N]\}

  6. Form Ic←{i:ci=c}I_c \leftarrow\{i:c_i=c\} and nc←∣Ic∣n_c \leftarrow|I_c| for each c∈CNc\in C_N

  7. c⋆←arg⁡max⁡c∈CNlog⁡Fκ(c)/ncc^\star\leftarrow\arg\max_{c\in C_N}\log F_{\kappa}(c)/n_c, breaking ties uniformly

  8. Draw i⋆∼Unif⁡(Ic⋆)i^\star\sim\operatorname{Unif}(I_{c^\star})

  9. return xi⋆x_{i^\star}

For detection, we require access only to the semantic encoder and the keyed scoring function, without access to the original generator. Our basic detector (Algorithm 2) computes the semantic score of an image and declares it watermarked if the score is at least τ\tau. For batch detection, we average the image scores before applying the threshold.

Algorithm 2 Exact-code detection from one or more images

Input: Images y1,…,yBy_1,\ldots,y_B, where B≥1B\geq1

Parameters: Encoder EE, secret key κ\kappa, and threshold τ\tau

  1. for j=1,…,Bj=1,\ldots,B do

  2. cj←HR(z(yj))c_j \leftarrow H_R(z(y_j))

  3. sj←Fκ(cj)s_j \leftarrow F_{\kappa}(c_j)

  4. end for

  5. return 1{B−1∑j=1Bsj≥τ}\mathbf{1}\left\{B^{-1}\sum_{j=1}^{B}s_j\geq\tau\right\}

However, non-semantic transformations may change the semantic region of an image near a region boundary, potentially changing its score. To address this limitation, we introduce a boundary-aware detector that also considers nearby semantic regions when computing the detection score. We use this detector in all our experiments and provide its full description in Appendix § 13.

Having introduced our generation and detection algorithms, we next establish their theoretical guarantees. In § 6, we prove that our generation algorithm is distortion-free and analyze how the detector’s performance depends on the semantic entropy of the generator.

Theoretical Guarantees

We now establish the theoretical guarantees of our watermarking scheme. First, we prove that our generation algorithm is distortion-free for any black-box generative model and semantic encoder, after averaging over the secret key. Second, we establish the detector’s false-positive rate and prove that sufficiently high semantic entropy guarantees strong detection for an appropriate choice of parameters. Finally, we establish guarantees for batch detection and robustness to transformations that approximately preserve the semantic representation.

Theorem 3 (Distortion-free generation). Let GG be any conditional generative model and EE any encoder for which Algorithm 1 is well-defined. Under the ideal-PRF assumption (Assumption 17), the watermarked generator defined by Algorithm 1 is distortion-free (Definition 10): for every prompt pp,

EK∼Unif⁡(K)[PGKwm(⋅∣p)]=PG(⋅∣p).\mathbb{E}_{K \sim\operatorname{Unif}(\mathcal{K})}\left[P_{G_K^{\mathrm{wm}}}(\cdot\mid p)\right] = P_G(\cdot\mid p).

We prove Theorem 3 in Appendix § 14. This ensures that, regardless of the semantic distribution of the outputs, the watermarking generator’s output is also distributed according to the original model’s distribution when we average over the random choice of the key.

We next establish a lower bound on the detection advantage of our watermarking scheme and analyze how it depends on the semantic entropy of the generator GG and encoder EE for a given prompt pp, as well as on three configurable parameters:

  • Number of candidates NN: The number of black-box generator calls per watermarked output. Larger NN increases the potential detection advantage, at a generation cost that scales linearly with NN.

  • Number of hash bits MM: Increasing MM reduces hash collisions, improving the detection bound but potentially decreasing robustness, since small changes in the semantic representation are more likely to change the hash code.

  • Detection threshold τ\tau: Controls the tradeoff between false-positive and true-positive rates.

The guarantee also depends on the semantic entropy HsemE,ε(G∣p)H_{\mathrm{sem}}^{E,\varepsilon}(G \mid p), defined in § 4. Higher semantic entropy strengthens our detection guarantee.

Theorem 4 (Semantic entropy and detection). Fix a prompt pp and a semantic radius ε∈[0,2]\varepsilon\in[0, 2]. Generate N≥2N \ge2 independent candidates and apply Algorithm 1. Under the ideal-PRF assumption, for single-image detection, the exact-code detector (Algorithm 2) at threshold τ∈[0,1]\tau\in[0, 1] satisfies

FPR⁡(τ)=1−τ\operatorname{FPR}(\tau) = 1 - \tau

and

TPR⁡(N)(τ)≥1−τ+1−qM,ε(p)1+(N−1)qM,ε(p)(τ−τN).\operatorname{TPR}^{(N)}(\tau) \ge1 - \tau+ \frac{1 - q_{M,\varepsilon}(p)}{1 + (N - 1)q_{M,\varepsilon}(p)}\left(\tau- \tau^N\right).

where

qM,ε(p)=e−HsemE,ε(G∣p)+(1−e−HsemE,ε(G∣p))(1−arccos⁡(1−ε)π)M.q_{M,\varepsilon}(p) = e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G \mid p)} + \left(1 - e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G \mid p)}\right)\left(1 - \frac{\arccos(1 - \varepsilon)}{\pi}\right)^M.

Both rates average over the secret key and random hyperplanes for the fixed prompt pp.

Theorem 4 shows that higher semantic entropy strengthens the guaranteed detection advantage without increasing the false-positive rate. For a suitable choice of parameters, sufficiently high semantic entropy therefore guarantees a substantial separation between watermarked and unwatermarked outputs. These guarantees average over the secret key and random hyperplanes. We provide the proof in Appendix § 15.

Finally, we prove in Appendix § 16 that batch detection reduces both error probabilities exponentially with the number of independently generated outputs, provided their expected clean and watermarked scores lie on opposite sides of the threshold for a fixed key. In Appendix § 17, we establish that transformations preserving semantic similarity are likely to preserve the hash code and hence the watermark score.

We next evaluate our method empirically. We first examine image quality and diversity under fixed-key reuse, since our distortion-free guarantee averages over the secret key (§ 7.1). We then evaluate single-image and batch detection and robustness to content-preserving transformations (§ 7.2), and compare semantic entropy across different encoders (§ 7.3).

Experiments and Results

Experimental setup. We conduct the watermarking experiments using Stable Diffusion 2.1 ([40, 37]) and COCO-2017 prompts ([30, 8]) held out from semantic encoder training. Our black-box method selects among complete candidate images. Because sampling-based watermarks primarily target language while image methods typically access the generation process, no directly comparable image-watermarking baseline shares the same sampling-only access model. We therefore compare with Tree-Ring ([43]), Gaussian Shading (Yang et al., 2024b), and SEAL ([2]), three closely related methods for watermarking diffusion-generated images. We evaluate all methods on a common generation backbone using each method’s native generation and detection procedures. For representation analysis, we compare our encoder with DINOv2 ([33]), CLIP ([36]), DINOv3 ([38]), and TPIPS ([42]) on independent generations from SD2.1 and two FLUX variants ([5]) and content-preserving perturbations of real images.

In addition to the main experiments, which test watermarking quality, detectability, and robustness, Appendix § 21 presents several extensions. We demonstrate closed-API use with GPT Image through black-box sampling (Appendix § 21.§ 5) and examine our method in a different modality, audio (Appendix § 21.§ 6). We also evaluate detection in mixed batches (Appendix § 21.§ 8), test whether edits that replace the main object remove the watermark signal (Appendix § 21.§ 10), and examine how the choice of semantic encoder affects robustness (Appendix § 21.§ 4).

Generation Quality and Diversity

Our distortion-free guarantee matches the original generator distribution after averaging over secret keys; it does not require each fixed key to preserve that distribution. Because deployment typically reuses one key across many generations, we empirically evaluate fixed-key quality and diversity. Prior work shows that fixed-key reuse can reduce same-prompt diversity for some initial-noise watermarking methods ([17] (Yang et al. 2025)). We therefore measure fixed-key distributional deviation and same-prompt diversity; Appendix § 21.§ 7 reports prompt alignment.

We evaluate Stable Diffusion 2.1 Base at 512×512512 \times512. We measure distributional deviation by computing FID between each evaluated population and its matched unwatermarked population, using 10,000 generations from 1,000 COCO-2017 prompts at 10 generations per prompt. To assess same-prompt diversity, we generate 60 images for each of 60 prompts and measure the relative Vendi Score in a frozen DINOv2-base feature space, independent of our semantic encoder. Each configuration uses an unwatermarked reference set with the same model, prompts, and sampling settings.

Table 1 reports both quantities. Across our three configurations, matched-population FID ranges from 4.75 to 5.22, averaged across 10 key selections, compared with 5.42–6.88 for the baselines. Relative to the clean-versus-clean reference of 4.64, these ranges are 1.024–1.125 and 1.168–1.483, respectively. Same-prompt diversity remains within 1.3% of clean generation for our method and is unchanged for Tree-Ring, whereas Gaussian Shading reduces it by 9.1% and SEAL by 13.2%. Together, these results show that, among the evaluated methods, our method has the smallest measured distributional shift under fixed-key reuse while retaining near-clean same-prompt diversity. Appendix § 21.§ 7 reports complementary FID against real COCO-2017 images and image-prompt alignment results.

Watermark Detection and Robustness

We evaluate detection on clean outputs and after common content-preserving transformations.

Clean images and image sets. Using 1,000 held-out COCO-2017 prompts, we evaluate matched watermarked and unwatermarked generations for (N,M)∈{(8,4),(16,8),(32,12)}(N, M) \in\{(8, 4), (16, 8), (32, 12)\} and set sizes B∈{1,5,10}B \in\{1, 5, 10\}. We report AUROC and TPR at empirical FPRs of 1%, 5%, and 20%. Tree-Ring, Gaussian Shading, and SEAL provide native single-image references; Appendix § 20.§ 1 reports the ROC curves and aggregation protocol.

Table 2 shows that aggregation improves detection: for (N,M)=(32,12)(N, M) = (32, 12), TPR@20% rises from 0.920 for one image to 0.999 for five and 1.000 for ten. While native baselines are stronger for single-image detection because they modify the generation process, our black-box method instead selects among NN complete outputs and preserves its distortion-free guarantee. Consequently, it faces a theoretical single-image ceiling, TPR≤N FPRTPR \le N\ FPR, inherent to black-box sampling (Appendix § 22).

MethodB=1B = 1B=5B = 5B=10B = 10
AUROCTPR@20%AUROCTPR@20%AUROCTPR@20%
Ours (N=8,M=4)(N = 8, M = 4)0.8020.6420.9380.9050.9600.941
Ours (N=16,M=8)(N = 16, M = 8)0.8790.8150.9910.9940.9980.998
Ours (N=32,M=12)(N = 32, M = 12)0.9200.9200.9980.9991.0001.000
Tree-Ring1.0001.000––––
Gaussian Shading1.0001.000––––
SEAL1.0001.000––––

Table 2. Clean detection on 1,000 matched prompts. TPR@20% is the true-positive rate read at 20% FPR from the mean ROC curve. A dash means that only single-image detection was evaluated for that native baseline.

We also test the framework through OpenAI’s closed GPT Image API, using eight candidates per output. With only sampling access to the generator, we obtain an AUROC of 0.815 for individual images and 0.978 for batches of five (Appendix § 21.§ 5).

Robustness under common transformations. To approximate deployment conditions, we evaluate images after transformations commonly encountered when they are stored, shared, or edited: appearance changes, compression, geometric transformations, and localized edits produced by a generative image-editing model. For the configurations (N,M)∈{(8,4),(16,8),(32,12)}(N, M) \in\{(8, 4), (16, 8), (32, 12)\}, we compare with the baseline methods in the single-image setting. We report both detection after transformation and retention, defined as the fraction of images detected before transformation that remain detected afterward. This distinction separates robustness from the initial detection strength of each method. Appendix § 20.§ 2 details transformations, detector configurations, calibration, and settings.

Figure 2 shows different patterns across transformation types. The baselines are stronger under appearance changes and compression, whereas all three of our configurations retain more initial detections under geometric transformations. For (N,M)=(16,8)(N, M) = (16, 8) and (32,12)(32, 12), detection after geometric transformation reaches 48.9% and 48.2%, above Tree-Ring (46.7%), Gaussian Shading, and SEAL. Thus, detection after transformation and retention provide complementary views of robustness. Table 10 and Table 11 in Appendix § 20.§ 2 report TPR and retention at 1% and 5% nominal clean FPR for single images and batches of edited images.

Robustness under appearance and compression, geometric transformations, and image-editing-model edits

Figure 2. Robustness at 20% nominal clean FPR. Filled and open markers show edited and clean detection; retention is the fraction of initial detections preserved after editing. Panels (a,b) cover appearance/compression and geometric transformations; panel (c) covers 80 generative edits. Colors mark the best and second-best values. Appendix § 20.§ 2 gives detector configurations, per-setting results, and FPRs.

FPRMethodAppearance/compression CleanAppearance/compression EditedAppearance/compression Ret.Geometric CleanGeometric EditedGeometric Ret.Generative edits CleanGenerative edits EditedGenerative edits Ret.
1%Ours (N=8N=8, M=4M=4)0.00.0–0.00.0–0.00.0–
Ours (N=16N=16, M=8M=8)7.87.187.27.83.842.08.85.057.1
Ours (N=32N=32, M=12M=12)14.613.086.614.65.934.213.87.554.5
Tree-Ring100.0100.0100.0100.019.519.5100.017.517.5
Gaussian Shading100.0100.0100.0100.01.81.8100.07.57.5
SEAL100.096.096.0100.01.21.2100.00.00.0
5%Ours (N=8N=8, M=4M=4)8.47.990.08.45.354.012.512.5100.0
Ours (N=16N=16, M=8M=8)32.629.990.432.617.348.531.227.584.0
Ours (N=32N=32, M=12M=12)49.244.389.049.221.740.855.037.568.2
Tree-Ring100.0100.0100.0100.026.626.6100.030.030.0
Gaussian Shading100.0100.0100.0100.011.511.5100.017.517.5
SEAL100.096.096.0100.01.21.2100.00.00.0

Table 10. Robustness at 1% and 5% nominal clean FPR (all values in %). Columns report TPR before editing (clean), post-edit TPR, and paired retention conditioned on original detection. Appearance/compression and geometric results are unweighted means over 14 and six settings, respectively; generative results summarize 80 edits. Detector settings are given in Table 5.

BFPRMethodAppearance/compression CleanAppearance/compression EditedAppearance/compression Ret.Geometric CleanGeometric EditedGeometric Ret.
B=5B=51%Ours (N=8N=8, M=4M=4)62.152.282.762.122.133.6
Ours (N=16N=16, M=8M=8)92.377.883.992.323.425.1
Ours (N=32N=32, M=12M=12)98.982.583.398.921.621.8
5%Ours (N=8N=8, M=4M=4)85.877.890.185.845.251.4
Ours (N=16N=16, M=8M=8)99.090.291.199.041.541.8
Ours (N=32N=32, M=12M=12)100.090.390.4100.036.736.7
20%Ours (N=8N=8, M=4M=4)97.093.796.397.072.574.2
Ours (N=16N=16, M=8M=8)100.097.097.1100.065.965.9
Ours (N=32N=32, M=12M=12)100.096.596.5100.060.160.1
B=10B=101%Ours (N=8N=8, M=4M=4)91.983.590.391.942.245.4
Ours (N=16N=16, M=8M=8)99.993.393.499.938.038.0
Ours (N=32N=32, M=12M=12)100.093.393.3100.033.933.9
5%Ours (N=8N=8, M=4M=4)98.294.796.298.265.666.6
Ours (N=16N=16, M=8M=8)100.097.597.5100.057.557.5
Ours (N=32N=32, M=12M=12)100.096.896.8100.050.150.1
20%Ours (N=8N=8, M=4M=4)99.899.099.199.885.585.6
Ours (N=16N=16, M=8M=8)100.099.399.3100.077.277.2
Ours (N=32N=32, M=12M=12)100.099.099.0100.071.071.0

Table 11. Batch detection after editing for our configurations (all values in %). Each batch contains BB images from different prompts that received the same transformation. FPRs are nominal, with thresholds calibrated on additional unedited negatives.

For the 80 localized edits made with a generative image editor, (N,M)=(16,8)(N, M) = (16, 8) achieves the highest TPR after editing (70.0%), followed by (32,12)(32, 12) (66.2%), both exceeding Tree-Ring (60.0%), Gaussian Shading (37.5%), and SEAL (10.0%). (16,8)(16, 8) also has the highest retention (84.8%). In contrast, (32,12)(32, 12) detects the most images before editing (88.8%) but keeps only 74.6% of those detections. This matches the tradeoff in Section 6: more hash bits strengthen detection but make the code more sensitive to edits.

Semantic Representation Analysis

Our watermarking method relies on a semantic representation with two complementary properties: content-preserving transformations should produce similar representations, while independent generations should remain distinguishable. We analyze whether the learned encoder satisfies these properties, its transfer across generators, and how its geometry relates to watermark retention.

We compare our encoder with DINOv2-base, CLIP ViT-B/32, DINOv3 L/16, and TPIPS-Embed. To evaluate same-prompt diversity, we generate 16 images with different random seeds for each of 200 COCO-2017 prompts held out from encoder training, using three generators: SD2.1 and two FLUX variants. Because the encoder was trained only on SDXL (Podell et al. [35]) images, this also tests transfer to unseen generators. Appendix § 19 gives full protocol details.

We quantify these properties using mean pairwise cosine similarity, which measures how closely representations are clustered. For content-preserving transformations, higher cosine similarity indicates greater invariance. For independent same-prompt generations, lower cosine similarity indicates greater diversity in the encoder’s representation. We also report semantic entropy at ε=0.5\varepsilon= 0.5, counting pairs with cosine similarity at least 0.5 as close.

Table 3 compares how the encoders represent independent generations from the same prompt. Our encoder achieves the lowest mean cosine similarity and highest estimated semantic entropy at ε=0.5\varepsilon= 0.5 for all three evaluated generators, none of which were used for training. Figure 5 in Appendix § 19 shows how the entropy comparison changes with ε\varepsilon.

DINOv2CLIPDINOv3TPIPSOurs
Cos.↓H^\hat{H} ↑Cos.↓H^\hat{H} ↑Cos.↓H^\hat{H} ↑Cos.↓H^\hat{H} ↑Cos.↓H^\hat{H} ↑
FLUX distilled0.8080.0330.9020.0000.7830.0540.8260.0000.4310.739
FLUX base0.7300.0930.8730.0010.7200.1080.7620.0010.2971.104
SD2.10.6340.2520.8630.0010.6510.2170.6450.0390.2221.345

Table 3. Same-prompt diversity in encoder space. Mean pairwise cosine similarity and estimated semantic entropy at ε=0.5\varepsilon= 0.5, across 16 independent generations from the same prompt. Lower cosine similarity and higher estimated semantic entropy indicate greater diversity in the encoder’s representation. Appendix H provides the complete protocol and stability results under content-preserving transformations.

Figure 5. Semantic entropy as a function of ε\varepsilon. Each panel reports H^semE,ε\widehat{H}_{\mathrm{sem}}^{E,\varepsilon} for one generator. Higher is better: at a fixed ε\varepsilon, higher curves indicate greater semantic diversity. Dashed lines mark the table radius, ε=0.5\varepsilon=0.5. Panels show ε∈[0,1]\varepsilon\in[0,1] and entropy in [0,5][0,5].

Appendix § 19 reports two further encoder analyses. First, it measures stability under content-preserving transformations; here, our encoder and DINOv3 are strongest overall: ours under JPEG compression and resizing, and DINOv3 under Gaussian blur and mild crop-and-color transformations. Second, it examines the discrete hash codes that the watermark actually uses. For independent generations of the same prompt, our encoder produces more distinct codes, and the proof of Theorem 4 shows that more occupied codes increase detection probability. Under content-preserving transformations, the opposite is desirable, and our encoder keeps relatively few distinct codes across all four transformation families, so an edited image tends to keep its original code. Appendix § 21.§ 2 isolates the encoder’s effect in the complete watermarking pipeline, where ours achieves the highest mean AUROC in all three evaluated configurations.

Encoder similarity closely tracks watermark retention. Across the 20 transformation settings, the Spearman correlation between encoder similarity and retention ranges from 0.988 to 0.998 for our configurations, compared with 0.747 to 0.779 for the baseline watermarks. This is consistent with our geometric analysis: higher cosine similarity makes the original and transformed images more likely to receive the same LSH code and clean keyed score (Lemma 22 in Appendix § 17). Figure 14 in Appendix § 21 provides the complete visualization. Consequently, an encoder that better matches human judgments of whether an edited image still shows the same scene, while still separating independent generations, could directly improve detection and robustness. Appendix § 21.§ 4 further demonstrates that replacing the encoder changes which transformations preserve watermark detection, allowing robustness to be tailored to application-specific image properties.

Figure 14. Encoder similarity and watermark retention. Each point pairs mean original–edited cosine similarity in our learned embedding with paired detection retention for one of 20 edit settings. Marker shape indicates the edit category, and color indicates the method. Spearman rank correlations are computed separately for each method. Detection thresholds are calibrated on clean images and frozen across edits.

Organization of the Appendices

We organize the appendices to provide the technical foundations, proofs, and additional experimental evidence supporting our results. Appendix A introduces the formal definitions and assumptions underlying our framework, while Appendix B provides the full algorithmic specifications. We then establish our theoretical guarantees. In Appendix C, we prove that our generation algorithm is distortion-free for any black-box generator and semantic encoder. In Appendix D, we show how increasing semantic entropy strengthens the detection guarantee. In Appendix E, we prove that aggregating independently generated outputs improves detection reliability by reducing detection errors exponentially. In Appendix F, we establish robustness guarantees for transformations that approximately preserve semantic representations.

To instantiate our framework for images, we train a semantic encoder that separates independent generations from the same prompt while remaining stable under content-preserving transformations. Appendix G describes the training procedure, and Appendix H evaluates the resulting representation. We show that our encoder achieves higher estimated semantic entropy than the evaluated alternatives across multiple image generators, while maintaining stability under common image transformations.

We further investigate the effectiveness and generality of our method through additional experiments. Appendix I provides the complete detection and robustness evaluation protocols used in experiments, additional results at low false-positive rates, and examples of image transformations. In Appendix J, we examine several design choices and extensions. We relate encoder similarity to watermark retention (Appendix J.1). To motivate our algorithmic and encoder choices, we compare our trained encoder against off-the-shelf alternatives and Gumbel sampling against other distortion-free selection rules (Appendices J.2 and J.3). We also compare how different embedding discretizers spread candidates across codes and preserve those codes under edits (Appendix J.9). We show that our encoder improves watermark detection and that Gumbel sampling achieves stronger detection than the alternative sampling rules we evaluate. We further demonstrate the flexibility of our black-box framework in two settings. First, by changing the semantic encoder, we can target different notions of robustness without modifying the watermarking algorithm (Appendix J.4). Second, we demonstrate that our method works with closed-API image generators without accessing their internals. In particular, our GPT Image experiment achieves an AUROC of 0.978 when aggregating five images (Appendix J.5). We also watermark text-to-audio outputs (Appendix J.6) and evaluate detection in batches that mix images from different keys or include unwatermarked images (Appendix J.8). Finally, we evaluate generation quality and prompt alignment under fixed-key reuse, finding the smallest measured distributional shift among the evaluated methods while maintaining nearly unchanged prompt alignment (Section 7.1 and Appendix J.7).

Appendix J.10 compares detection on watermarked and clean images after edits that replace the main subject, scene, or action.

Finally, Appendix K discusses the limitations of our approach. In particular, we establish an inherent limitation of black-box sampling: when selecting one of NN unmodified candidate outputs, the single-image true-positive rate cannot exceed NN times the false-positive rate (Appendix K.1). We also discuss computational costs, fixed-key reuse, adaptive attacks, and extensions beyond images.

Limitations and Future Work

Our approach has three main limitations, each pointing to future work:

  • Single-image detection. Our current image implementation provides weaker clean single-image detection than methods that directly modify the diffusion process (Table 2). This tradeoff matters most when every image must be detected independently and internal model access is available. Our method instead prioritizes black-box applicability, semantic binding, and distribution preservation, while aggregation substantially strengthens detection (Table 2). More generally, any method that selects one of NN independent, unmodified generator outputs satisfies TPR≤NFPR\mathrm{TPR} \leq N\mathrm{FPR} for single-image detection (Theorem 23 in Appendix K.1). Thus, achieving constant TPR at exponentially small FPR requires exponentially many generator queries. This O(N)O(N) ceiling belongs to black-box sampling, not our construction.

  • Semantic encoder. A better semantic encoder could improve detection within this limit and strengthen robustness. It must keep content-preserving transformations close while separating independent generations; ours improves this balance over the evaluated alternatives (Appendix § 21.§ 2).

  • Computational cost. Our strongest configuration generates 32 complete candidates per released image, which may be impractical for large-scale or latency-sensitive applications. Early-selection methods may reduce this cost by using intermediate denoiser activations to predict quality or reusing low-resolution semantic features across denoising steps (Guo et al., [18]; Habibian et al., [19]; Parmar et al., [34]). If these signals predict the final semantic code, unpromising candidates could be discarded early.

In Appendix § 22 we discuss these limitations at length as well as further open problems.

Conclusion

We introduced a black-box, distortion-free semantic watermarking framework requiring only sampling access to a generator and query access to a semantic encoder. Our analysis connects detectability to semantic entropy, while experiments demonstrate aggregate detection and robustness to common transformations without compromising generation quality or diversity. These results establish semantic rejection sampling as a general black-box foundation for generative watermarking. Stronger encoders and more efficient sampling remain directions for future work.

AI Use Statement, Reproducibility and Acknowledgments

Generative AI assistants were used to support software development and analysis workflows, literature discovery, and the editing and formatting of manuscript text. They were also used as a source of feedback while refining the presentation of the method, experimental protocol, and theoretical exposition. We independently designed the study, ran and checked the experiments, verified code and mathematical statements, and take full responsibility for the paper’s content. The images produced by the generative models evaluated in this work are experimental outputs generated under the stated protocols, rather than content produced by these assistants.

The main paper specifies the sampling-based watermarking method, its theoretical assumptions, and the experimental setup. Our code is available at https://github.com/avivyossef/distortion-free-semantic-watermark.

The research of OZ and AY is supported by grants from the AISI Alignment Project and the Israeli Ministry of Science.

Formal Preliminaries and Definitions

Generative Models and Sampling

Definition 5 (Conditional generative model). Let P\mathcal{P} be the set of all possible prompts, let S={0,1}k\mathcal{S} = \{0,1\}^{k} be the set of all possible seeds, and let X\mathcal{X} be the output space. A conditional generative model is a mapping

G:P×S→X.G: \mathcal{P} \times\mathcal{S} \to\mathcal{X}.

We model S\mathcal{S} as a finite set of seeds, as is the case in practical implementations, whereas the prompt space P\mathcal{P} may be infinite. For any fixed prompt p∈Pp \in\mathcal{P} and seed s∈Ss \in\mathcal{S}, the output G(p,s)G(p,s) is deterministic.

Definition 6 (I.i.d. sample from a conditional generative model). Fix a prompt p∈Pp \in\mathcal{P}. Let

S1,…,Sn∼i.i.d.Unif⁡(S),Xi=G(p,Si).S_1,\ldots,S_n \mathrel{\overset{\mathrm{i.i.d.}}{\sim}} \operatorname{Unif}(\mathcal{S}), \qquad X_i = G(p,S_i).

Any realization (x1,…,xn)(x_1,\ldots,x_n) of (X1,…,Xn)(X_1,\ldots,X_n) is called an i.i.d. sample from GG conditioned on the prompt pp.

Definition 7 (Encoder and semantic encoder). Let Z=Rd\mathcal{Z} = \mathbb{R}^{d} be a dd-dimensional representation space. An encoder is a mapping

E:X→Z.E: \mathcal{X} \to\mathcal{Z}.

For x∈Xx \in\mathcal{X}, we call E(x)E(x) the representation of xx. We call EE a semantic encoder if its representations capture the semantic content of the outputs.

The precise properties required of a semantic encoder in our framework are introduced in § 4.

Watermarking Generative Models

Definition 8 (Keyed watermarking scheme). Let G\mathcal{G} be a conditional generative model and let K\mathcal{K} be a finite key space. A keyed watermarking scheme for G\mathcal{G} is a pair

W=(Gwm,Detect⁡),\mathcal{W} = \left(G^{\mathrm{wm}}, \operatorname{Detect}\right),

where, for every key κ∈K\kappa\in\mathcal{K}:

  1. The watermarked generator is a possibly randomized mapping

Gκwm:P×SN→X.G_{\kappa}^{\mathrm{wm}}:\mathcal{P} \times\mathcal{S}^{N} \to\mathcal{X}.

Given a prompt and NN seeds, it produces a watermarked output.

  1. For every batch size B≥1B \ge1, the detector is a mapping

Detect⁡κ,τ(B):XB→{0,1}.\operatorname{Detect}_{\kappa,\tau}^{(B)}:\mathcal{X}^{B} \to\{0,1\}.

Given a batch of images x1:B=(x1,…,xB)x_{1:B} = (x_{1},\ldots,x_{B}) and a fixed detection threshold τ∈[0,1]\tau\in[0,1], it outputs 1 when the batch is detected as watermarked under the key κ\kappa, and 0 otherwise. The case B=1B = 1 corresponds to single-image detection.

Definition 9 (Black-box watermarking scheme). A watermarking scheme W=(Gwm,Detect⁡)\mathcal{W} = (G^{\mathrm{wm}},\operatorname{Detect}) is black-box with respect to G\mathcal{G} if GwmG^{\mathrm{wm}} and Detect⁡\operatorname{Detect} may interact with G\mathcal{G} only by evaluating G(p,s)\mathcal{G}(p,s), and otherwise have no access to its internal implementation.

Definition 10 (Distortion-free watermarking). Let PG(⋅∣p)P_{\mathcal{G}}(\cdot\mid p) denote the distribution of G(p,S)\mathcal{G}(p,S), where S∼Unif⁡(S)S \sim\operatorname{Unif}(\mathcal{S}). For each key κ∈K\kappa\in\mathcal{K}, let PGκwm(⋅∣p)P_{G_{\kappa}^{\mathrm{wm}}}(\cdot\mid p) denote the output distribution of the watermarked generator for prompt pp. The watermarking scheme is distortion-free if

EK∼Unif⁡(K)[PGKwm(⋅∣p)]=PG(⋅∣p)\mathbb{E}_{K \sim\operatorname{Unif}(\mathcal{K})}\left[P_{G_{K}^{\mathrm{wm}}}(\cdot\mid p)\right] = P_{\mathcal{G}}(\cdot\mid p)

for every prompt p∈Pp \in\mathcal{P}.

This is a key-averaged, single-output guarantee: it requires equality after averaging over a uniformly sampled key KK, but not for every fixed key.

Hashing and Pseudorandom Functions

Definition 11 (Hash function). A hash function is a mapping h:Z→Ch:\mathcal{Z} \to\mathcal{C}, where C\mathcal{C} is a finite code space. Two representations z,z′∈Zz,z' \in\mathcal{Z} collide under hh if h(z)=h(z′)h(z) = h(z').

Definition 12 (Locality-sensitive hash family). Let (Z,d)(\mathcal{Z},d) be a metric space. A distribution H\mathcal{H} over hash functions h:Z→Ch:\mathcal{Z} \to\mathcal{C} is (r,cr,p1,p2)(r,cr,p_{1},p_{2})-locality-sensitive if c>1c > 1, p1>p2p_{1} > p_{2}, and for every z,z′∈Zz,z' \in\mathcal{Z},

d(z,z′)≤r⟹Pr⁡h∼H[h(z)=h(z′)]≥p1,d(z,z') \le r \Longrightarrow\Pr_{h \sim\mathcal{H}}[h(z) = h(z')] \ge p_{1},

whereas

d(z,z′)≥cr⟹Pr⁡h∼H[h(z)=h(z′)]≤p2.d(z,z') \ge cr \Longrightarrow\Pr_{h \sim\mathcal{H}}[h(z) = h(z')] \le p_{2}.

We use the standard notion of locality-sensitive hashing [26]. In our method, H\mathcal{H} is instantiated with the MM-bit random-hyperplane LSH construction for angular distance, equivalently cosine similarity, following Charikar [7].

Definition 13 (Random-hyperplane locality-sensitive hash). Let R=(rˉ1,…,rˉM)\mathcal{R} = (\bar{r}_{1},\ldots,\bar{r}_{M}) consist of MM unit hyperplane normals sampled independently and uniformly from the unit sphere. For a normalized representation z∈Rdz \in\mathbb{R}^{d}, the random-hyperplane LSH is the code

HR(z)=(1{⟨rˉ1,z⟩≥0},…,1{⟨rˉM,z⟩≥0})∈{0,1}M.H_{\mathcal{R}}(z) = \left( \mathbb{1}\{\langle\bar{r}_{1},z\rangle\ge0\},\ldots,\mathbb{1}\{\langle\bar{r}_{M},z\rangle\ge0\} \right) \in\{0,1\}^{M}.

Thus, each bit indicates on which side of its corresponding hyperplane the representation lies.

Definition 14 (Hamming distance and Hamming ball). Following Hamming [20], the Hamming distance between two codes c,c′∈{0,1}Mc, c^{\prime} \in\{0,1\}^{M} is the number of bit positions on which they differ:

dH(c,c′)=∑j=1M1{cj≠cj′}.d_{\mathrm{H}}(c,c^{\prime}) = \sum_{j=1}^{M} \mathbf{1}\{c_j \ne c_j^{\prime}\}.

For an integer r∈{0,…,M}r \in\{0,\ldots,M\}, the radius-rr Hamming ball centered at cc is

BH(c,r)={c′∈{0,1}M:dH(c,c′)≤r}.\mathcal{B}_{\mathrm{H}}(c,r) = \{c^{\prime} \in\{0,1\}^{M}: d_{\mathrm{H}}(c,c^{\prime}) \le r\}.

Definition 15 (Boundary-restricted Hamming neighborhood). For a normalized representation zz and boundary threshold δ\delta, let

Jδ(z)={j∈[M]:∣⟨rˉj,z⟩∣≤δ}J_{\delta}(z) = \{j \in[M] : |\langle\bar{r}_j,z\rangle| \le\delta\}

be the set of hash bits whose corresponding hyperplanes lie within the boundary margin. Let Flip⁡(c,S)\operatorname{Flip}(c,S) denote the code obtained by flipping the bits of cc indexed by SS. For an integer r∈{0,…,M}r \in\{0,\ldots,M\}, define

Nδ,r(z)={Flip⁡(HR(z),S):S⊆Jδ(z),∣S∣≤r}.\mathcal{N}_{\delta,r}(z) = \{\operatorname{Flip}(H_{\mathcal{R}}(z),S) : S \subseteq J_{\delta}(z), |S| \le r\}.

Every code in Nδ,r(z)\mathcal{N}_{\delta,r}(z) lies within Hamming distance rr of HR(z)H_{\mathcal{R}}(z). Unlike the complete Hamming ball, however, this neighborhood permits changes only to bits associated with nearby hyperplane boundaries.

We use a pseudorandom function (PRF) to assign a secret score to each hash code, following Goldreich et al. [15].

Definition 16 (Keyed scoring function). Let K\mathcal{K} be the key space and let C={0,1}M\mathcal{C} = \{0,1\}^{M} be the code space. A keyed scoring family is a collection of functions

{Fκ:C→[0,1]}κ∈K.\{F_{\kappa} : \mathcal{C} \to[0,1]\}_{\kappa\in\mathcal{K}}.

For a secret key κ\kappa, the value Fκ(c)F_{\kappa}(c) is the normalized secret score assigned to code cc.

Assumption 17 (Ideal-PRF scoring model). Let K∼Unif⁡(K)K \sim\operatorname{Unif}(\mathcal{K}). For every collection of distinct codes c1,…,cLc_1,\ldots,c_L, the scores

FK(c1),…,FK(cL)F_K(c_1),\ldots,F_K(c_L)

are independent and uniformly distributed over [0,1][0,1]. Equivalently, in the theoretical analysis, FKF_K is modeled as a uniformly random function from C\mathcal{C} to [0,1][0,1]. In particular, distinct codes receive independent scores, and score ties occur with probability zero.

A practical implementation can obtain normalized scores from the finite bit-string output of a cryptographic PRF. The exact equalities in our theoretical results refer to the ideal model above; a secure PRF provides the corresponding computational interpretation.

Algorithmic Specification

In Section 5, we introduced our watermarking algorithm using Gumbel selection (Figure 1). Here, we first introduce a simpler sampling rule to explain how we preserve distortion-freeness when candidates share semantic codes. We then explain why we replace this rule with Gumbel selection in our final algorithm. Finally, we describe the boundary-aware detector we use in our experiments.

When all candidates have distinct semantic codes, we can simply return the image with the highest keyed score. Under the ideal-PRF assumption (Assumption 17), each candidate has probability 1/N1/N of receiving the highest score after averaging over the secret key. However, when multiple candidates share a code, they also share its score. Selecting the highest-scoring code and then choosing uniformly among its candidates would favor candidates belonging to smaller groups.

We first address this problem with a simpler construction inspired by bundle sampling (Vaikuntanathan & Zamir, [41]), which we call representative sampling. We choose one representative uniformly from each group of candidates sharing a code and let the representatives compete according to their keyed scores. We then compensate for the remaining candidates through an additional uniform-selection branch.

More precisely, let Ic={i∈[N]:ci=c}I_c = \{i \in[N] : c_i = c\} denote the group of candidates sharing code cc. We choose one representative from each group to form UU, and let D=[N]∖UD = [N] \setminus U. With probability ∣U∣/N|U|/N, we return the highest-scoring representative. Otherwise, we select uniformly from DD. Algorithm 3 specifies the complete procedure. In Appendix § 14, we prove that this rule selects every candidate with probability 1/N1/N after averaging over the secret key and sampler randomness.

Algorithm 3 Representative sampling

Input: Prompt pp and number of candidates NN

Parameters: GG, EE, and secret key κ\kappa

  1. Sample independent seeds S1,…,SNS_1, \ldots, S_N — Candidate seeds

  2. for i=1,…,Ni = 1, \ldots, N do

  3. xi←G(p,Si)x_i \leftarrow G(p, S_i) — Generate candidate

  4. ci←HR(z(xi))c_i \leftarrow H_R(z(x_i)) — Assign its LSH code

  5. end for

  6. CN←{ci:i∈[N]}C_N \leftarrow\{c_i : i \in[N]\} — Occupied codes

  7. U←∅U \leftarrow\varnothing

  8. for all c∈CNc \in C_N do

  9. Ic←{i∈[N]:ci=c}I_c \leftarrow\{i \in[N] : c_i = c\} — Candidates sharing code cc

  10. Sample Rc∼Uniform⁡(Ic)R_c \sim\operatorname{Uniform}(I_c) — Choose one representative

  11. U←U∪{Rc}U \leftarrow U \cup\{R_c\}

  12. end for

  13. D←[N]∖UD \leftarrow[N] \setminus U — Remaining candidates

  14. Sample Z∼Bernoulli⁡(∣U∣/N)Z \sim\operatorname{Bernoulli}(|U|/N) — Preserve uniform marginal selection

  15. if Z=1Z = 1 then

  16. A←arg⁡max⁡i∈UFκ(ci)A \leftarrow\arg\max_{i \in U} F_\kappa(c_i) — Highest keyed score

  17. i⋆∼Uniform⁡(A)i^\star\sim\operatorname{Uniform}(A) — Break ties uniformly

  18. else

  19. i⋆∼Uniform⁡(D)i^\star\sim\operatorname{Uniform}(D) — Select a remaining candidate

  20. end if

  21. return xi⋆x_{i^\star}

Representative sampling provides an intuitive distortion-free construction, but its uniform-selection branch ignores the keyed scores of the remaining candidates. We therefore replace it with Gumbel selection (Kuditipudi et al., [29]), which accounts for code multiplicities while using keyed scores throughout code selection. Specifically, we select the code maximizing log⁡Fκ(c)/nc\log F_\kappa(c)/n_c, where nc=∣Ic∣n_c = |I_c|, and then sample uniformly from its candidates, as specified in Algorithm 1. Both rules coincide when all candidates have distinct codes. In Appendix § 21.§ 3, we compare the two rules and tournament sampling on identical candidate pools and show that Gumbel selection achieves stronger detection performance. We therefore use Gumbel selection in our final algorithm and experiments. In Appendix § 14, we prove that both sampling rules preserve the original generator distribution.

Boundary-aware detection. We now describe how we improve robustness to content-preserving transformations. Such transformations may move an image’s semantic representation across a nearby hyperplane and change its hash code, even when the image retains its semantic content. To address this problem, we identify hash bits whose representations lie close to their corresponding hyperplane boundaries. We then evaluate nearby codes obtained by flipping up to rhamr_{\mathrm{ham}} of these bits and use the highest keyed score for detection.

When rham=0r_{\mathrm{ham}} = 0, we recover the exact-code detector from Algorithm 2. In all our experiments, including clean-image detection, we use rham=1r_{\mathrm{ham}} = 1 and calibrate detection thresholds on clean images. By restricting the search to near-boundary bits, we avoid evaluating the complete Hamming ball. We formally define the boundary-restricted neighborhood in Appendix § 12 and specify the complete detection procedure for both individual images and batches in Algorithm 4.

Algorithm 4 Boundary-aware detection from one or more images
Input: Images x1,…,xBx_1,\ldots,x_B, where B≥1B \ge1
Parameters: Encoder EE, boundary threshold δ\delta, Hamming radius rhamr_{\mathrm{ham}}, secret key κ\kappa, and detection threshold τ\tau
Notation: Flip(c,S)(c,S) denotes code cc with the bits indexed by SS flipped.
1: for b=1,…,Bb=1,\ldots,B do
2: zb←E(xb)/∥E(xb)∥2\quad z_b \leftarrow E(x_b)/\lVert E(x_b)\rVert_2
3: cb←HR(zb)\quad c_b \leftarrow H_R(z_b)
4: Jδ(zb)←{j∈[M]:∣⟨rˉj,zb⟩∣≤δ}\quad J_\delta(z_b) \leftarrow\{j \in[M] : |\langle\bar r_j,z_b\rangle| \le\delta\}
5: Nδ,rham(zb)←{Flip⁡(cb,S):S⊆Jδ(zb),∣S∣≤rham}\quad\mathcal{N}_{\delta,r_{\mathrm{ham}}}(z_b) \leftarrow\{\operatorname{Flip}(c_b,S) : S \subseteq J_\delta(z_b), |S| \le r_{\mathrm{ham}}\}
6: sb←max⁡c′∈Nδ,rham(zb)Fκ(c′)\quad s_b \leftarrow\max_{c' \in\mathcal{N}_{\delta,r_{\mathrm{ham}}}(z_b)} F_\kappa(c')
7: end for
8: return 1{1B∑b=1Bsb≥τ}\mathbb{1}\left\{\frac{1}{B}\sum_{b=1}^{B}s_b \ge\tau\right\}

Table 4.

We use Algorithm 4 for both single-image and batch detection in our experiments. For B=1B=1, we evaluate an individual image; for larger batches, we average the boundary-aware scores across independently generated outputs. In Appendix E, we prove that batch detection reduces both error probabilities exponentially with batch size when the expected clean and watermarked scores lie on opposite sides of the threshold for a fixed key.

Proof of Distortion-Free Generation

We now prove the distortion-free guarantee stated in Theorem 3. Our goal is to show that our watermarking algorithm preserves the original generator’s distribution after averaging over the secret key, regardless of the generator or semantic encoder. We first prove that representative sampling (Algorithm 3, Appendix B) and Gumbel selection (Algorithm 1, Section 5) both select every candidate with probability 1/N1/N, even when candidates share semantic codes. We then show that this uniform-selection property guarantees distortion-free generation.

Conditional Uniform Candidate Selection

We begin with representative sampling (Algorithm 3), which handles candidates sharing semantic codes by choosing one representative from each group and compensating for the remaining candidates through a uniform draw.

Lemma 2 (Conditional uniform candidate selection).

Let X1,…,XN∼i.i.d.PG(⋅∣p)X_1,\ldots,X_N \mathrel{\mathop{\sim}^{\mathrm{i.i.d.}}} P_G(\cdot\mid p), and let I⋆∈[N]I^\star\in[N] be the index selected by Algorithm 3. Under the ideal-PRF assumption (Assumption 17), for every fixed candidate batch x1:Nx_{1:N} and every i∈[N]i \in[N],

EK∼Unif⁡(K)[Pr⁡(I⋆=i∣X1:N=x1:N,K)]=1N.\mathbb{E}_{K\sim\operatorname{Unif}(\mathcal{K})}\left[\Pr(I^\star=i\mid X_{1:N}=x_{1:N},K)\right]=\frac{1}{N}.

Proof. Fix a candidate batch x1:Nx_{1:N} and its codes c1:Nc_{1:N}. Let L=∣U∣L=|U| denote the number of occupied codes, and let ni=∣Ici∣n_i=|I_{c_i}| denote the number of candidates sharing candidate ii’s code.

Candidate ii becomes the representative of its group with probability 1/ni1/n_i. We choose the representative branch with probability L/NL/N. Under the ideal-PRF assumption, the LL distinct occupied codes receive independent, uniformly distributed scores, so each code wins with probability 1/L1/L. The representative branch therefore selects candidate ii with probability

1ni⋅LN⋅1L=1Nni.\frac{1}{n_i}\cdot\frac{L}{N}\cdot\frac{1}{L}=\frac{1}{Nn_i}.

:::`

If L<NL < N, candidate ii belongs to DD with probability (ni−1)/ni(n_i - 1)/n_i. We choose the complementary branch with probability (N−L)/N(N-L)/N and select uniformly among the N−LN-L remaining candidates. This branch selects candidate ii with probability

ni−1ni⋅N−LN⋅1N−L=ni−1Nni.\frac{n_i-1}{n_i}\cdot\frac{N-L}{N}\cdot\frac{1}{N-L}=\frac{n_i-1}{Nn_i}.

Adding both contributions, we obtain

Pr⁡(I⋆=i∣X1:N=x1:N)=1Nni+ni−1Nni=1N.\Pr(I^\star=i\mid X_{1:N}=x_{1:N})=\frac{1}{Nn_i}+\frac{n_i-1}{Nn_i}=\frac{1}{N}.

where we average over the secret key and sampler randomness.

If L=NL=N, every candidate has a distinct code, so the representative branch alone selects each candidate with probability 1/N1/N.

We next establish the same uniform-selection property for Gumbel selection (Algorithm 1), which we use in our final watermarking algorithm. Unlike representative sampling, Gumbel selection accounts for code multiplicities while using keyed scores throughout code selection.

Fix a candidate batch and let Ic={i:ci=c}I_c=\{i:c_i=c\}, with nc=∣Ic∣n_c=|I_c|. Under the ideal-PRF assumption, the scores FK(c)F_K(c) are independent and uniformly distributed over [0,1][0,1] for distinct codes. Consequently, the random variables

Tc=−log⁡FK(c)ncT_c=\frac{-\log F_K(c)}{n_c}

are independent exponential random variables with rates ncn_c. By the exponential-minimum sampling identity [29], the code minimizing TcT_c, equivalently maximizing log⁡FK(c)/nc\log F_K(c)/n_c, wins with probability

Pr⁡(c⋆=c)=nc∑c′∈CNnc′=ncN.\Pr(c^\star=c)=\frac{n_c}{\sum_{c'\in C_N}n_{c'}}=\frac{n_c}{N}.

We then select uniformly among the ncn_c candidates sharing the winning code. Therefore, each candidate ii has selection probability

Pr⁡(I⋆=i∣X1:N=x1:N)=nciN⋅1nci=1N,\Pr(I^\star=i\mid X_{1:N}=x_{1:N})=\frac{n_{c_i}}{N}\cdot\frac{1}{n_{c_i}}=\frac{1}{N},

where we again average over the secret key and sampler randomness.

We have thus established conditional uniform candidate selection for both representative sampling and Gumbel selection, even when multiple candidates share the same semantic code.

Distortion-Free Generation

We now prove Theorem 3 using the conditional uniform-selection property established in Section C.1. We first apply it to Gumbel selection (Algorithm 1) and then conclude that the same guarantee holds for representative sampling.

Theorem 3 (Distortion-free generation). Let GG be any conditional generative model and EE any encoder for which Algorithm 1 is well-defined. Under the ideal-PRF assumption (Assumption 17), the watermarked generator defined by Algorithm 1 is distortion-free (Definition 10): for every prompt pp,

EK∼Unif⁡(K)[PGwm(⋅∣p)]=PG(⋅∣p).\mathbb{E}_{K\sim\operatorname{Unif}(\mathcal{K})}\left[P_G^{\mathrm{wm}}(\cdot\mid p)\right]=P_G(\cdot\mid p).

Proof. Fix a prompt pp and draw X1,…,XNX_1,\ldots,X_N independently from PG(⋅∣p)P_G(\cdot\mid p). Let I⋆I^\star be the index selected by Algorithm 1, and set Xwm=XI⋆X^{\mathrm{wm}}=X_{I^\star}. Fix any measurable set of outputs AA. By conditioning on the candidate batch and applying the uniform-selection property of Algorithm 1 proved in Section C.1, we obtain

EK∼Unif⁡(K)[PGwm(A∣p)]=Pr⁡(Xwm∈A∣p)=EX1:N[∑i=1N1{Xi∈A}Pr⁡(I⋆=i∣X1:N)]=EX1:N[1N∑i=1N1{Xi∈A}]=1N∑i=1NPr⁡(Xi∈A∣p)=PG(A∣p).\begin{aligned} \mathbb{E}_{K\sim\operatorname{Unif}(\mathcal{K})}\left[P_G^{\mathrm{wm}}(A\mid p)\right] &=\Pr(X^{\mathrm{wm}}\in A\mid p)\\ &=\mathbb{E}_{X_{1:N}}\left[\sum_{i=1}^{N}\mathbb{1}\{X_i\in A\}\Pr(I^\star=i\mid X_{1:N})\right]\\ &=\mathbb{E}_{X_{1:N}}\left[\frac{1}{N}\sum_{i=1}^{N}\mathbb{1}\{X_i\in A\}\right]\\ &=\frac{1}{N}\sum_{i=1}^{N}\Pr(X_i\in A\mid p)\\ &=P_G(A\mid p). \end{aligned}

Therefore, the watermarked output follows the original generator distribution after averaging over the secret key.

We also prove distortion-freeness for representative sampling (Algorithm 3) by applying the same argument to its uniform-selection property established in Section C.1.

High Semantic Entropy and Detection

We want to understand when high semantic entropy leads to strong detection in our construction. Detection benefits from candidates occupying distinct codes, since each occupied code supplies another keyed score. We therefore use semantic entropy to bound code agreement, turn this into a lower bound on expected code occupancy, and then derive the detection guarantee.

Detection rates. We evaluate detection for a fixed prompt pp, declaring an output watermarked when its exact-code score is at least τ\tau. The false-positive rate FPR⁡(τ)\operatorname{FPR}(\tau) measures how often an independent unwatermarked sample X∼PG(⋅∣p)X\sim P_G(\cdot\mid p) is declared watermarked. The true-positive rate TPR⁡(N)(τ)\operatorname{TPR}^{(N)}(\tau) measures how often the detector correctly identifies a watermarked output from Algorithm 1, using NN candidates for the same prompt. We average both rates over the secret key, random hyperplanes, and generation and sampler randomness, using the same key and hyperplanes for watermarked generation and detection. Because the unwatermarked sample is independent of the key, its score is uniform under the ideal-PRF assumption. Therefore,

FPR⁡(τ)=Pr⁡(U≥τ)=1−τ,U∼Unif⁡([0,1]).\operatorname{FPR}(\tau)=\Pr(U\ge\tau)=1-\tau,\qquad U\sim\operatorname{Unif}([0,1]).

The boundary-aware detector used in our experiments also checks nearby codes and therefore requires separate false-positive calibration. These averaged rates do not give a guarantee for every fixed key or hyperplane draw.

Lemma 18 (High semantic entropy limits hash agreement). Fix a prompt pp and a semantic radius ε∈[0,2]\varepsilon\in[0,2]. For two independent generations and an MM-bit random-hyperplane LSH, the probability that their codes agree is at most

qM,ε(p)=e−HsemE,ε(G∣p)+(1−e−HsemE,ε(G∣p))(1−arccos⁡(1−ε)π)M.q_{M,\varepsilon}(p)=e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)}+\left(1-e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)}\right)\left(1-\frac{\arccos(1-\varepsilon)}{\pi}\right)^M.

Proof. Let A={dE(X,X′)≤ε}A=\{d_E(X,X')\le\varepsilon\}, and let B={HR(z(X))=HR(z(X′))}B=\{H_R(z(X))=H_R(z(X'))\}. The probability is over two independent generations and the independent random hyperplanes. By Definition 1,

Pr⁡(A)=e−HsemE,ε(G∣p),Pr⁡(Ac)=1−e−HsemE,ε(G∣p).(1)\Pr(A)=e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)},\qquad\Pr(A^c)=1-e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)}. \tag*{(1)}

For a fixed pair at distance dd, the random-hyperplane identity gives

Pr⁡R(B∣X,X′)=(1−arccos⁡(1−d)π)Mwhere d=dE(X,X′).\Pr_{\mathcal{R}}(B \mid X, X') = \left(1-\frac{\arccos(1-d)}{\pi}\right)^M \qquad\text{where } d=d_E(X,X').

This expression decreases with dd. Thus Pr⁡(B∣A)≤1\Pr(B\mid A)\le1 and Pr⁡(B∣Ac)≤(1−arccos⁡(1−ε)/π)M\Pr(B\mid A^c)\le\left(1-\arccos(1-\varepsilon)/\pi\right)^M. The law of total probability now gives

Pr⁡(B)=Pr⁡(B∩A)+Pr⁡(B∩Ac)=Pr⁡(B∣A)Pr⁡(A)+Pr⁡(B∣Ac)Pr⁡(Ac)≤Pr⁡(A)+(1−arccos⁡(1−ε)π)MPr⁡(Ac)=e−HsemE,ε(G∣p)+(1−e−HsemE,ε(G∣p))(1−arccos⁡(1−ε)π)M=qM,ε(p).\begin{aligned} \Pr(B) &= \Pr(B\cap A)+\Pr(B\cap A^c)\\ &= \Pr(B\mid A)\Pr(A)+\Pr(B\mid A^c)\Pr(A^c)\\ &\le\Pr(A)+\left(1-\frac{\arccos(1-\varepsilon)}{\pi}\right)^M\Pr(A^c)\\ &=e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)} +\left(1-e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)}\right) \left(1-\frac{\arccos(1-\varepsilon)}{\pi}\right)^M\\ &=q_{M,\varepsilon}(p). \end{aligned}

If a conditioning event has probability zero, its term is omitted.

Lemma 19 (Expected number of occupied codes). Fix a prompt pp, a semantic radius ε∈[0,2]\varepsilon\in[0,2], and N≥2N\ge2. Let X1,…,XNX_1,\ldots,X_N be independent generations from PG(⋅∣p)P_G(\cdot\mid p), and let LL be the number of distinct codes HR(z(Xi))H_R(z(X_i)) under an MM-bit random-hyperplane LSH. Then

E[L]≥N1+(N−1)qM,ε(p).\mathbb{E}[L]\ge\frac{N}{1+(N-1)q_{M,\varepsilon}(p)}.

where qM,ε(p)q_{M,\varepsilon}(p) is defined in Lemma 18. The expectation is over the candidate generations and random hyperplanes.

Proof. Write q=qM,ε(p)q=q_{M,\varepsilon}(p) and Ci=HR(z(Xi))C_i=H_R(z(X_i)), and let JJ count pairs of candidates whose codes agree. If the occupied-code multiplicities are n1,…,nLn_1,\ldots,n_L, then

J=∑i<j1{Ci=Cj}=∑j=1L(nj2),∑j=1Lnj2=N+2J.J=\sum_{i<j}\mathbf{1}\{C_i=C_j\}=\sum_{j=1}^{L}\binom{n_j}{2},\qquad\sum_{j=1}^{L}n_j^2=N+2J.

Lemma 18 and linearity of expectation give

E[J]=∑i<jPr⁡(Ci=Cj)≤(N2)q.\mathbb{E}[J]=\sum_{i<j}\Pr(C_i=C_j)\le\binom{N}{2}q.

By Cauchy–Schwarz,

N2=(∑j=1Lnj)2≤L∑j=1Lnj2=L(N+2J).N^2=\left(\sum_{j=1}^{L}n_j\right)^2\le L\sum_{j=1}^{L}n_j^2=L(N+2J).

Taking expectations and applying Jensen’s inequality to the reciprocal,

E[L]≥E[N2N+2J]≥N2N+2E[J]≥N1+(N−1)q.\mathbb{E}[L]\ge\mathbb{E}\left[\frac{N^2}{N+2J}\right]\ge\frac{N^2}{N+2\mathbb{E}[J]}\ge\frac{N}{1+(N-1)q}.

To turn code occupancy into a detection guarantee, we next compute the score distribution of the code returned by Gumbel selection. We fix the candidate codes so that only the key remains random.

Lemma 20 (Gumbel selection over occupied codes). Fix candidate codes c1:Nc_{1:N}, let CN={ci:i∈[N]}C_N = \{c_i : i \in[N]\}, and let nc=∣{i:ci=c}∣n_c = |\{i : c_i = c\}|. Draw K∼Unif⁡(K)K \sim\operatorname{Unif}(\mathcal{K}). Under the ideal-PRF assumption (Assumption 17), let c⋆c^\star be the code selected by Algorithm 1. Then for every c∈CNc \in C_N and τ∈[0,1]\tau\in[0,1],

Pr⁡(c⋆=c)=ncN,Pr⁡(FK(c⋆)≥τ∣c⋆=c)=1−τN/nc.\Pr(c^\star= c) = \frac{n_c}{N}, \qquad\Pr(F_K(c^\star) \ge\tau\mid c^\star= c) = 1 - \tau^{N/n_c}.

The probabilities average over the secret key.

Proof. Write Uc=FK(c)U_c = F_K(c). Under the ideal-PRF assumption these scores are independent and uniform on [0,1][0,1]. Gumbel selection maximizes log⁡Uc/nc\log U_c/n_c. Thus, for a fixed score Uc=u∈(0,1)U_c = u \in(0,1), code cc wins exactly when every other occupied code dd satisfies

log⁡Udnd≤log⁡unc,or equivalentlyUd≤und/nc.\frac{\log U_d}{n_d} \le\frac{\log u}{n_c}, \qquad\text{or equivalently} \qquad U_d \le u^{n_d/n_c}.

Ties have probability zero. Independence and uniformity therefore give

Pr⁡(c⋆=c∣Uc=u)=∏d≠cund/nc=u(N−nc)/nc.\Pr(c^\star= c \mid U_c = u) = \prod_{d \ne c} u^{n_d/n_c} = u^{(N-n_c)/n_c}.

Since UcU_c has uniform density, integrating over scores at least τ\tau gives

Pr⁡(c⋆=c,Uc≥τ)=∫τ1u(N−nc)/nc du=ncN(1−τN/nc).\begin{aligned} \Pr(c^\star= c, U_c \ge\tau) &= \int_\tau^1 u^{(N-n_c)/n_c}\,du \\ &= \frac{n_c}{N}\left(1-\tau^{N/n_c}\right). \end{aligned}

At τ=0\tau= 0, this gives Pr⁡(c⋆=c)=nc/N\Pr(c^\star= c) = n_c/N. Dividing the joint probability by nc/Nn_c/N then gives Pr⁡(Uc≥τ∣c⋆=c)=1−τN/nc\Pr(U_c \ge\tau\mid c^\star= c) = 1-\tau^{N/n_c}, as claimed. This is the occupied-code form of the exponential-minimum identity in Kuditipudi et al. (2024, Lemma C.1).

Theorem 4 (Semantic entropy and detection). Fix a prompt pp and a semantic radius ε∈[0,2]\varepsilon\in[0,2]. Generate N≥2N \ge2 independent candidates and apply Algorithm 1. Under the ideal-PRF assumption, for single-image detection, the exact-code detector (Algorithm 2) at threshold τ∈[0,1]\tau\in[0,1] satisfies

FPR⁡(τ)=1−τ\operatorname{FPR}(\tau) = 1-\tau

and

TPR⁡(N)(τ)≥1−τ+1−qM,ε(p)1+(N−1)qM,ε(p)(τ−τN),\operatorname{TPR}^{(N)}(\tau) \ge1-\tau+ \frac{1-q_{M,\varepsilon}(p)}{1+(N-1)q_{M,\varepsilon}(p)}\left(\tau-\tau^N\right),

where

qM,ε(p)=e−HsemE,ε(G∣p)+(1−e−HsemE,ε(G∣p))(1−arccos⁡(1−ε)π)M.q_{M,\varepsilon}(p) = e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)} +\left(1-e^{-H_{\mathrm{sem}}^{E,\varepsilon}(G\mid p)}\right) \left(1-\frac{\arccos(1-\varepsilon)}{\pi}\right)^M.

Both rates average over the secret key and random hyperplanes for the fixed prompt pp; the definitions are given above.

Proof. Write q=qM,ε(p)∈[0,1]q = q_{M,\varepsilon}(p) \in[0,1], and let LL be the number of occupied codes among the NN candidates. For each candidate ii, let ni=∣Ici∣n_i = |I_{c_i}| be the number of candidates sharing its code.

Detection conditioned on code multiplicities. Condition on the candidates and hyperplanes, and average over the independent key and sampler randomness. By Lemma 20,

Pr⁡(Detect⁡=1∣c1:N)=∑c∈CNncN(1−τN/nc)=1−τ+1N∑i=1N(τ−τN/ni).\begin{aligned} \Pr(\operatorname{Detect}=1 \mid c_{1:N}) &= \sum_{c\in C_N}\frac{n_c}{N}\left(1-\tau^{N/n_c}\right) \\ &= 1-\tau+\frac{1}{N}\sum_{i=1}^{N}\left(\tau-\tau^{N/n_i}\right). \end{aligned}

The second equality counts each occupied code once for each of its members.

Combining the bounds. For 0<τ<10 < \tau< 1, the function s↦τ−τss \mapsto\tau- \tau^s is concave and vanishes at s=1s = 1. Its chord between s=1s = 1 and s=Ns = N gives

τ−τN/ni≥N/ni−1N−1(τ−τN).\tau- \tau^{N/n_i} \ge\frac{N/n_i - 1}{N - 1}\left(\tau- \tau^N\right).

Each occupied code contributes NN to ∑i=1NN/ni\sum_{i=1}^{N} N/n_i, so this sum equals NLNL. Therefore,

Pr⁡(Detect⁡=1∣c1:N)≥1−τ+L−1N−1(τ−τN).\Pr(\operatorname{Detect} = 1 \mid c_{1:N}) \ge1 - \tau+ \frac{L - 1}{N - 1}\left(\tau- \tau^N\right).

Taking expectations over the candidates and hyperplanes and applying Lemma 19,

TPR⁡(N)(τ)≥1−τ+E[L]−1N−1(τ−τN)\operatorname{TPR}^{(N)}(\tau) \ge1 - \tau+ \frac{\mathbb{E}[L] - 1}{N - 1}\left(\tau- \tau^N\right)
≥1−τ+1N−1(N1+(N−1)q−1)(τ−τN)=1−τ+1−q1+(N−1)q(τ−τN).\begin{aligned} \ge1 - \tau+ \frac{1}{N - 1}\left(\frac{N}{1 + (N - 1)q} - 1\right)\left(\tau- \tau^N\right) \\ &= 1 - \tau+ \frac{1 - q}{1 + (N - 1)q}\left(\tau- \tau^N\right). \end{aligned}

An independent unwatermarked output has a uniform exact-code score after averaging over the key, giving FPR⁡(τ)=1−τ\operatorname{FPR}(\tau) = 1 - \tau. The endpoint cases τ=0\tau= 0 and τ=1\tau= 1 follow directly.

Batch-Detection Concentration

We now analyze how aggregating multiple independently generated outputs improves watermark detection. In Algorithm 2, we introduced batch detection by averaging the keyed scores of multiple images and comparing the result with a threshold. We show that this aggregation reduces both false-positive and false-negative probabilities exponentially with the batch size, provided the expected clean and watermarked scores lie on opposite sides of the threshold.

Unlike Theorem 4, which averages over the secret key and random hyperplanes, we analyze detection with a fixed deployed key and hash. Our result applies to both the exact-code detector (Algorithm 2) and the boundary-aware detector (Algorithm 4), since both aggregate per-image scores in [0,1][0, 1].

Lemma 21 (Batch-detection concentration). Fix a secret key κ\kappa and hash HRH_R, and let μ0\mu_0 and μwm\mu_{\mathrm{wm}} denote the expected per-image scores for unwatermarked and watermarked generations, respectively. Suppose that

μ0<τ<μwm,\mu_0 < \tau< \mu_{\mathrm{wm}},

where τ\tau is the batch-detection threshold. For a batch of BB independently generated images with the same expected score within each population, the false-positive and false-negative probabilities satisfy

Pr⁡(FP)≤exp⁡(−2B(τ−μ0)2)\Pr(\mathrm{FP}) \le\exp\left(-2B(\tau- \mu_0)^2\right)

and

Pr⁡(FN)≤exp⁡(−2B(μwm−τ)2).\Pr(\mathrm{FN}) \le\exp\left(-2B(\mu_{\mathrm{wm}} - \tau)^2\right).

Thus, both error probabilities decrease exponentially with BB.

Proof. We first consider a batch of BB independently generated unwatermarked images Y1,…,YBY_1, \ldots, Y_B. Let

Sb0=Score⁡κ,R(Yb),S‾B0=1B∑b=1BSb0.S_b^0 = \operatorname{Score}_{\kappa,R}(Y_b), \qquad\overline{S}_B^0 = \frac{1}{B}\sum_{b=1}^{B} S_b^0.

For the fixed key and hash, the scores are independent, lie in [0,1][0, 1], and have expectation μ0\mu_0. A false positive occurs when S‾B0≥τ\overline{S}_B^0 \ge\tau. Applying Hoeffding’s inequality, we obtain

Pr⁡(FP)=Pr⁡(S‾B0−μ0≥τ−μ0)≤exp⁡(−2B(τ−μ0)2).\begin{aligned} \Pr(\mathrm{FP}) = \Pr\left(\overline{S}_B^0 - \mu_0 \ge\tau- \mu_0\right) \\ &\le\exp\left(-2B(\tau- \mu_0)^2\right). \end{aligned}

We next consider BB independently generated watermarked images X1wm,…,XBwmX^{\mathrm{wm}}_1,\ldots,X^{\mathrm{wm}}_B. Define

Sbwm=Score⁡κ,R(Xbwm),S‾Bwm=1B∑b=1BSbwm.S_b^{\mathrm{wm}}=\operatorname{Score}_{\kappa,\mathcal{R}}(X_b^{\mathrm{wm}}),\qquad\overline{S}_B^{\mathrm{wm}}=\frac{1}{B}\sum_{b=1}^{B}S_b^{\mathrm{wm}}.

For the same fixed key and hash, these scores are independent, lie in [0,1][0,1], and have expectation μwm\mu_{\mathrm{wm}}. A false negative occurs when S‾Bwm<τ\overline{S}_B^{\mathrm{wm}}<\tau. Applying Hoeffding’s inequality again, we obtain

Pr⁡(FN)=Pr⁡(S‾Bwm<τ)≤Pr⁡(μwm−S‾Bwm≥μwm−τ)≤exp⁡(−2B(μwm−τ)2).\begin{aligned} \Pr(\mathrm{FN})=\Pr\left(\overline{S}_B^{\mathrm{wm}}<\tau\right) \\ &\leq\Pr\left(\mu_{\mathrm{wm}}-\overline{S}_B^{\mathrm{wm}}\geq\mu_{\mathrm{wm}}-\tau\right) \\ &\leq\exp\left(-2B(\mu_{\mathrm{wm}}-\tau)^2\right). \end{aligned}

Since μ0<τ<μwm\mu_0<\tau<\mu_{\mathrm{wm}}, both exponents are strictly negative and decrease linearly with BB. We therefore obtain exponential decay of both detection error probabilities.

We emphasize that our fixed-key mean-separation assumption is necessary for this result. The key-averaged guarantee in Theorem 4 does not imply that the expected scores are separated for every deployed key and hash. Rather, we show that whenever this separation holds, aggregating more independently generated outputs exponentially improves detection reliability.

Robustness Under Semantic Similarity

We now analyze how transformations that preserve semantic similarity affect our watermark. An edit may change many pixels while leaving the image’s semantic representation nearly unchanged. Since we assign watermark scores to LSH codes, we want to understand when such transformations preserve the original code and consequently its keyed score.

We show that the probability of preserving the hash code increases with the cosine similarity between the original and transformed representations. This provides a theoretical explanation for why our watermark can remain detectable after content-preserving transformations.

Lemma 22 (Hash-code preservation under semantic similarity). Fix two images xx and x′x' with nonzero encoder representations, and let

ρ=⟨z(x),z(x′)⟩.\rho=\langle z(x),z(x')\rangle.

Let HRH_{\mathcal{R}} consist of MM independently sampled uniform random hyperplanes, and fix a secret key κ\kappa. Then

Pr⁡R[HR(z(x′))=HR(z(x))]=(1−arccos⁡(ρ)π)M.\Pr_{\mathcal{R}}\left[H_{\mathcal{R}}(z(x'))=H_{\mathcal{R}}(z(x))\right]=\left(1-\frac{\arccos(\rho)}{\pi}\right)^M.

Whenever the codes coincide, the images receive identical exact-code keyed scores. In particular, if the exact-code detector identifies xx as watermarked, our boundary-aware detector also identifies x′x' as watermarked on this event.

Proof. For a single random hyperplane, the probability that z(x)z(x) and z(x′)z(x') receive the same bit is

1−arccos⁡(ρ)π.1-\frac{\arccos(\rho)}{\pi}.

Since we sample the MM hyperplanes independently, the probability that all hash bits agree is

(1−arccos⁡(ρ)π)M.\left(1-\frac{\arccos(\rho)}{\pi}\right)^M.

When the codes agree, their keyed scores also agree:

Fκ(HR(z(x′)))=Fκ(HR(z(x))).F_{\kappa}(H_{\mathcal{R}}(z(x')))=F_{\kappa}(H_{\mathcal{R}}(z(x))).

Our boundary-aware detector always includes the observed hash code among the codes it evaluates. Therefore, if Fκ(HR(z(x)))≥τF_{\kappa}(H_{\mathcal{R}}(z(x)))\geq\tau, then

Score⁡κ,R(x′)≥Fκ(HR(z(x′)))≥τ,\operatorname{Score}_{\kappa,\mathcal{R}}(x')\geq F_{\kappa}(H_{\mathcal{R}}(z(x')))\geq\tau,

and the boundary-aware detector identifies x′x' as watermarked.

We thus establish a code-preservation guarantee that depends on semantic similarity. The probability averages over random hyperplanes; we do not claim the same guarantee for every fixed hash or against adaptive transformations.

Semantic Encoder Training

We train the semantic encoder to satisfy two complementary requirements: semantics-preserving transformations of the same image should remain close, whereas independent generations from the same prompt should remain distinguishable. Figure 3 summarizes these requirements, and Figure 4 illustrates the corresponding training pairs.

Architecture and objective. We initialize the encoder from DINOv2-base and freeze its pretrained parameters. Rank-8 LoRA adapters [24] are applied to the query and value projections of the last four transformer blocks. The CLS representation is passed through a 768→512→256768 \to512 \to256 MLP with GELU activation and dropout 0.1, followed by ℓ2\ell_2 normalization. For normalized embeddings ziz_i and zjz_j, we use

L+=1−zi⊤zj,L−=max⁡(0,zi⊤zj).\mathcal{L}_{+} = 1 - z_i^{\top} z_j, \qquad\mathcal{L}_{-} = \max(0, z_i^{\top} z_j).

Positive pairs include Gaussian noise, Gaussian blur, JPEG compression, resampling, photometric changes, and affine transformations. Negative pairs contain generations obtained from different seeds of the same prompt.

Training details. We use 100 prompts evenly from DiffusionDB’s 2M-text-only split and the official COCO-2017 training captions. We divide them into 80 training and 20 validation prompts; neither set is used in the reported evaluation experiments. We use 64 seeds per prompt. Each optimizer update contains 16 pairs of one type, with pair types balanced across updates. We train using AdamW with learning rate 10−410^{-4}, mixed precision, gradient clipping at 1.0, and learning-rate decay of 0.95 per epoch. We select the rank-8 LoRA model with the MLP head on the validation split. The checkpoint used in Appendix § 19 was selected after 19,840 optimizer updates; the semantic-entropy results were not used for model selection.

Encoder Evaluation: Details and Additional Results

This appendix expands the semantic representation analysis in § 7.3. We provide the perturbation settings, analyze code counts under both transformations and independent generation, and conclude with a qualitative example. The main text gives the entropy definition and encoder comparison.

Stability evaluation. We evaluate stability using 200 real source images and 16 semantics-preserving transformations of each image and for every transformation family: JPEG compression, Gaussian blur, resizing and resampling, and mild cropping with color adjustment. Within each transformation family, severity increases gradually across the 16 variants: JPEG quality ranges from 35 to 95, blur radius from 0.25 to 1.75, resize scale from 0.65 to 0.95, and crop scale from 0.86 to 0.98, with color factors from 0.90 to 1.08. Every encoder is evaluated on the same source images and transformations. For each family, Table 4 summarizes all 16 severity levels together: we compute mean pairwise cosine similarity across the variants of each source image and then average the resulting per-image values over the evaluation set.

DINOv2CLIPDINOv3TPIPSOurs
JPEG compression0.9740.9590.9760.9560.985
Gaussian blur0.9220.9320.9740.8950.963
Resize–resample0.9920.9940.9950.9780.996
Mild crop + color0.9540.9640.9720.9190.937

Table 4 (PDF p. 30). Encoder stability under content-preserving transformations. Mean pairwise cosine similarity across transformed versions of the same source image. Higher values indicate greater invariance.

Sensitivity to ε\varepsilon. Figure 5 reports semantic-entropy curves using all (162)=120\binom{16}{2}=120 pairs per prompt and 200 held-out prompts per generator. For Table 3, we use the same samples at ε=0.5\varepsilon=0.5, counting pairs with cosine similarity at least 0.5 as close. Each reported entropy value is computed from the overall fraction of close pairs across the 200 prompts.

Distinct codes under perturbations. We next examine whether semantics-preserving transformations retain the same discrete hash code. For each transformation type, the analysis uses 200 source images and 16 perturbed variants of each image. We apply 8-bit random-hyperplane hashing to the corresponding embeddings and average over 200 projections, using each projection consistently across all images and encoders. Figure 6 reports the resulting code counts. Fewer distinct codes indicate that transformed versions of the same image more often retain its original code. This discrete analysis complements the continuous results above.

Figure 6. Distinct codes under semantics-preserving perturbations. Mean number of distinct 8-bit codes as additional perturbed versions are included. Each panel averages 200 source-image groups over 200 random hyperplane draws. Lower counts are desirable. Separate panels highlight the different behavior of the four perturbation types.

Distinct codes under independent generations. We perform the corresponding code-count analysis on independent generations from the same prompt. For each of SDXL and the two FLUX variants, we evaluate 200 prompts with up to 16 independently generated images per prompt. All encoders are evaluated on the same images. For every encoder, results are averaged over 200 random-hyperplane projections, with each projection fixed across all images. Figure 7 reports the corresponding code counts. More distinct codes mean fewer repeated candidate codes. The proof of Theorem 4 relates code occupancy to detection probability.

Figure 7. Distinct codes under independent generations. Mean number of distinct 8-bit codes as additional independent generations from the same prompt are included. Each panel averages results across 200 prompts and 200 random-hyperplane projections. Higher counts are desirable because they indicate fewer collisions. Separate panels show SDXL and the two FLUX variants, making transfer across generators visible.

Qualitative view of generation diversity. Figure 8 presents 16 SDXL images generated from one prompt. Although the images share the same description, they differ in layout and appearance. The cosine-similarity matrices show how DINOv2 and our encoder represent these differences.

Sixteen independent SDXL generations and pairwise cosine similarity matrices under DINOv2-base and Ours

Figure 8. Different images from one prompt. (a) Sixteen independent SDXL generations. (b) Pairwise cosine similarities under DINOv2-base and our encoder. DINOv2 assigns high similarity across the group, whereas our encoder represents more of the differences between the images. SDXL is the encoder’s training domain; the prompts are held out.

Detection and Robustness Protocols

Evaluation Setup

Prompts and datasets. We evaluate clean detection using images generated from 1,000 captions from the COCO-2017 validation set. For each prompt, we generate both watermarked and unwatermarked images using the same model and generation settings. The watermarked images form the positive population, while the unwatermarked images form the negative population used to measure the false-positive rate. These prompts are held out from semantic-encoder training; further details about the encoder-training data are provided in Appendix § 18.

Detection settings. The detector has three parameters: the boundary threshold δ\delta, the Hamming radius rhamr_{\mathrm{ham}}, and the score threshold τ\tau. Boundary-aware detection is activated when an image representation lies within margin δ\delta of at least one of the MM random hyperplanes used to construct its hash code. The activation rate is the fraction of clean watermarked images for which this occurs. When activated, the detector evaluates neighboring codes that differ in at most rhamr_{\mathrm{ham}} bits. For each (N,M)(N, M) configuration, we select δ\delta and rhamr_{\mathrm{ham}} through a parameter sweep on a separate development set drawn from the COCO-2017 training set, considering clean detection, detection after editing, and the corresponding false-positive rate. We use rham=1r_{\mathrm{ham}} = 1 throughout and choose δ\delta to obtain activation rates of 10%, 50%, and 75% for (N,M)=(8,4)(N, M) = (8, 4), (16,8)(16, 8), and (32,12)(32, 12), respectively. The sweep favors higher activation rates for larger MM: longer codes introduce more hyperplane boundaries that an edit may cross, while the larger candidate budget provides a stronger watermark signal that can support more frequent boundary correction. With δ\delta and rhamr_{\mathrm{ham}} fixed, we vary τ\tau to obtain the required FPR. The selected settings are summarized in Table 5.

(N,M)(N, M)δ\delta calibration targetrhamr_{\mathrm{ham}}
(8,4)(8, 4)10% activation1
(16,8)(16, 8)50% activation1
(32,12)(32, 12)75% activation1

Table 5. Selected detector settings. Because δ\delta is calibrated separately for each random-hyperplane frame, we report its target activation rate rather than a single numeric value.

Clean-detection experiment. We evaluate detection from either a single image or a set of B∈{5,10}B \in\{5, 10\} images generated using the same watermark key. For image sets, we average the BB per-image detection scores and compare the resulting score with τ\tau. Watermarked images form the positive population, and unwatermarked images generated from the same prompts form the negative population. For each configuration and set size, the false-positive rate is the fraction of unwatermarked images or image sets whose score exceeds τ\tau, while the true-positive rate is the corresponding fraction of watermarked images or sets. We compute ROC curves from these held-out positive and negative scores and report AUROC and TPR at empirical FPR targets, averaged over 25 combinations of five random-hyperplane frames and five watermark keys. Table 2 summarizes the AUROC and TPR@20% results, and Figure 9 and Figure 10 show the corresponding TPR/FPR curves on linear and logarithmic FPR axes.

Figure 9. Clean detection for our method at B∈{1,5,10}B \in\{1,5,10\} on a linear FPR axis. Curves are means over 25 hyperplane-frame and watermark-key combinations; Table 6 gives numerical operating points.

Figure 10. The same clean-detection curves on a logarithmic FPR axis, exposing aggregation gains at low FPR.

Table 6 reports the corresponding TPR at 1%, 5%, and 20% FPR for our configurations and the native single-image baselines.

MethodBBTPR at FPR
Ours (N=8N=8, M=4M=4)14.923.164.2
552.375.390.5
1069.785.094.1
Ours (N=16N=16, M=8M=8)18.635.081.5
586.296.499.4
1097.399.299.8
Ours (N=32N=32, M=12M=12)113.547.892.0
597.399.499.9
1099.799.9100.0
Tree-Ring1100.0100.0100.0
Gaussian Shading1100.0100.0100.0
SEAL1100.0100.0100.0

Table 6. Clean-detection TPR (%) at empirical FPR targets, averaged over 25 combinations of five hyperplane frames and five watermark keys. BB is the number of images attributed to one source; baselines use their native single-image detectors.

Robustness Protocols and Results

Ordinary edits. We evaluate robustness on 200 prompts in the single-image setting (B=1B = 1). Starting from clean watermarked and unwatermarked images, we apply the same transformation to both populations. The detector parameters δ\delta, rhamr_{\mathrm{ham}}, and τ\tau are fixed before applying the transformations and are not recalibrated on edited images. For each transformation, we report the edited TPR on watermarked images and the edited FPR on unwatermarked images. We also report paired retention: the fraction of watermarked images detected before transformation that remain detected afterward. This separates robustness to an edit from the detector’s initial clean-detection rate. The evaluation includes 20 settings spanning compression, appearance changes, noise and filtering, cropping, and translation.

DINOv2 is pretrained using augmentations including random resized crops, color changes, and Gaussian blur, and therefore already maps many such transformations to similar representations. Our fine-tuning preserves and adapts this behavior using positive pairs constructed with noise, blur, JPEG compression, resampling, photometric changes, and affine transformations, while negative pairs separate independent generations of the same prompt. The robustness evaluation includes transformations used during DINOv2 pretraining, transformations used during our fine-tuning, and transformations used during neither. Figure 12 reports the per-transformation results and marks this training exposure for each transformation. This allows us to examine generalization to held-out images and transformation settings, including transformation families not used during training.

More generally, robustness depends on the geometry of the semantic representation. For a target transformation, the method benefits from an encoder that maps the original and transformed images to nearby representations. § 7.3 compares DINOv2 with our fine-tuned encoder and shows the relationship between encoder similarity and watermark retention. For applications involving different transformations, the semantic encoder can therefore be selected or trained to maintain high cosine similarity under those transformations.

For the non-semantic generative-edit evaluation, we use 80 watermarked images and 80 corresponding unwatermarked images, with 20 examples assigned to each of four edit categories: attribute change, background cleanup, object insertion, and camera-viewpoint change. For each source image, we provide Qwen3-VL-8B-Instruct ([4]) with the original image and the fixed category-specific prompt reported in Table 7. The vision model then produces an image-specific editing instruction that follows the assigned category while preserving the main subject and scene. Qwen-Image-Edit ((Wu et al., 2025)) receives the original image together with this instruction and generates the edited image. Figure 11 shows this pipeline on one attribute-change example. We manually review every edited image to verify that the requested change is visible while the main subject and overall scene remain preserved. Representative examples from all four edit categories are shown in Figure 13.

Generative-edit evaluation protocol

Figure 11. Generative-edit evaluation protocol. For each source image, the assigned category prompt and image are given to Qwen3-VL-8B-Instruct, which produces a validated image-specific instruction. The original image and this instruction are then passed to Qwen-Image-Edit, and the edited result is evaluated using detector parameters fixed before editing. The same pipeline is applied to watermarked and corresponding unwatermarked sources. The example changes the refrigerator handles from metallic to matte black; the boxes identify the object named by the instruction, not a guaranteed pixel-local change.

Edit categoryFixed prompt constraint
Attribute changeModify only the colour or material of one visibly present, non-salient secondary object. Preserve the main subject, main action, object inventory, and scene. Do not target faces, text, logos, or watermarks.
Background cleanupRemove only one tiny, non-salient background clutter item. Preserve the main subject, main action, salient objects, and scene. Do not target faces, text, logos, or watermarks.
Object insertionAdd one unrelated small realistic cat in the background. The cat must be clearly visible but must not replace or occlude the main subject. Do not target faces, text, logos, or watermarks.
Camera-viewpoint changeChange only the camera viewpoint to a clearly different three-quarter or slightly elevated viewpoint. Preserve the same main subject, objects, actions, scene, colours, and lighting. Do not target faces, text, logos, or watermarks.

Table 7. Fixed category-specific constraints provided to Qwen3-VL-8B-Instruct when generating image-specific editing instructions.

Figure 13. Examples of ordinary edits, grouped by family across three pages: appearance and compression, geometric transformations, and edits from an image-editing model. Each row pairs the original image with its edited version or versions; parameter values are shown below the images where applicable.

All results average five hyperplane frames and five PRF keys. Paired retention matches original and edited images within each method. For ordinary edits, achieved FPR is measured on transformed negatives and each reported category is an unweighted mean over its transformation settings. Table 8 summarizes the ordinary and generative-edit results. Detailed per-transformation TPR and retention results are reported in Figure 12.

MethodAppearance/compressionGeometricGenerative edits
TPRRet.FPRTPRRet.FPRTPRRet.FPR
Ours (N=8N=8, M=4M=4)55.493.117.241.065.917.641.276.717.8
Ours (N=16N=16, M=8M=8)76.692.719.648.957.520.270.084.819.2
Ours (N=32N=32, M=12M=12)83.591.419.248.251.820.066.274.620.5
Tree-Ring100.0100.022.646.746.721.760.060.021.0
Gaussian Shading100.0100.021.627.427.423.737.537.522.0
SEAL97.997.918.714.514.519.210.010.019.5

Table 8. Robustness at 20% nominal clean FPR (all values in %). Columns report post-edit TPR, paired retention conditioned on original detection, and post-edit FPR on unwatermarked images. Appearance/compression and geometric results are unweighted means over 14 and six settings, respectively; generative results summarize 80 edits. Detector settings are given in Table 5.

The boundary-aware settings were selected on the separate development set described above. Table 9 evaluates their effect on the held-out ordinary-edit evaluation set by comparing them with exact-code detection on the same images within each (N,M)(N, M) pair.

(NN, MM)DetectorTPRRetentionFPR
(8, 4)Exact code50.5482.7016.75
(8, 4)Boundary aware51.0684.9617.33
(16, 8)Exact code65.4472.6419.13
(16, 8)Boundary aware68.2882.1419.78
(32, 12)Exact code65.6366.7018.53
(32, 12)Boundary aware72.9179.5119.47

Table 9. Boundary-aware detector ablation on the held-out ordinary-edit evaluation set at 20% nominal clean FPR. Values are unweighted means over the 20 reported transformation settings (14 appearance/compression and six geometric). TPR and FPR are measured after transformation, and retention is conditioned on original detection. Exact-code detection uses no boundary expansion; boundary-aware detection uses the settings in Table 5. Each pair uses the same images.

Best / Second
Edited TPR (%)
EditDINOv2 pretrainOurs fine-tuneOurs (8, 4)Ours (16, 8)Ours (32, 12)Tree-RingGaussian ShadingSEAL
Appearance and compression
JPEG Q90×✓58.081.490.1100.0100.099.5
JPEG Q75×✓56.480.188.3100.0100.099.5
Gaussian blur r=0.5r = 0.5✓✓58.481.890.6100.0100.0100.0
Gaussian blur r=1.5r = 1.5✓✓54.776.483.0100.0100.0100.0
Gaussian noise σ=2\sigma= 2×✓57.681.289.5100.0100.0100.0
Gaussian noise σ=6\sigma= 6×✓54.476.282.7100.0100.099.5
Brightness ×0.8✓✓58.181.389.9100.0100.0100.0
Brightness ×1.2✓✓56.779.286.4100.0100.0100.0
Contrast ×0.8✓✓58.181.390.0100.0100.0100.0
Contrast ×1.2✓✓56.979.287.0100.0100.0100.0
Median 3 × 3××57.079.987.9100.0100.0100.0
Median 5 × 5××52.571.276.3100.0100.0100.0
Salt-and-pepper 1%××50.065.168.7100.0100.091.5
Salt-and-pepper 3%××46.657.758.9100.0100.080.5
Geometric transformations
Center crop 50%✓×32.432.629.418.520.514.0
Center crop 75%✓×41.449.148.354.532.59.0
Off-center crop 75%✓×35.135.031.838.523.57.0
Off-center crop 90%✓×43.954.954.347.541.018.5
Shift 32 px×✓43.454.855.121.023.530.0
Shift 8 px×✓49.667.170.0100.023.58.5
Edits from an image-editing model
Attribute change××45.065.070.065.025.010.0
Background cleanup××45.080.090.065.050.015.0
Cat insertion××45.080.065.075.060.010.0
Camera-viewpoint change××30.055.040.035.015.05.0
Paired retention (%)
EditDINOv2 pretrainOurs fine-tuneOurs (8, 4)Ours (16, 8)Ours (32, 12)Tree-RingGaussian ShadingSEAL
Appearance and compression
JPEG Q90×✓98.299.198.9100.0100.099.5
JPEG Q75×✓95.497.296.7100.0100.099.5
Gaussian blur r=0.5r = 0.5✓✓99.399.699.3100.0100.0100.0
Gaussian blur r=1.5r = 1.5✓✓92.092.490.8100.0100.0100.0
Gaussian noise σ=2\sigma= 2×✓97.398.698.0100.0100.0100.0
Gaussian noise σ=6\sigma= 6×✓91.391.990.5100.0100.099.5
Brightness ×0.8✓✓98.498.998.6100.0100.0100.0
Brightness ×1.2✓✓95.495.994.6100.0100.0100.0
Contrast ×0.8✓✓98.398.898.6100.0100.0100.0
Contrast ×1.2✓✓95.796.195.2100.0100.0100.0
Median 3 × 3××96.496.996.3100.0100.0100.0
Median 5 × 5××87.585.683.2100.0100.0100.0
Salt-and-pepper 1%××82.978.074.8100.0100.091.5
Salt-and-pepper 3%××76.168.963.9100.0100.080.5
Geometric transformations
Center crop 50%✓×49.136.730.718.520.514.0
Center crop 75%✓×67.157.752.054.532.59.0
Off-center crop 75%✓×54.139.533.538.523.57.0
Off-center crop 90%✓×71.765.458.747.541.018.5
Shift 32 px×✓70.765.159.621.023.530.0
Shift 8 px×✓82.480.576.1100.023.58.5
Edits from an image-editing model
Attribute change××90.086.777.865.025.010.0
Background cleanup××90.0100.094.765.050.015.0
Cat insertion××81.894.181.275.060.010.0
Camera-viewpoint change××50.061.144.435.015.05.0

Table 12. Per-edit detection and paired retention for the 20 selected appearance/compression and geometric settings, followed by four edit families produced by an image-editing model, at 20% nominal clean FPR. The upper section reports edited TPR and the lower section reports paired retention; all values are percentages. The two left columns in both sections mark transformation-family exposure during DINOv2 pretraining and our encoder fine-tuning. Bold green and orange indicate the best and second-best values within each row; ties at the displayed precision share a color. Detector configurations and evaluation protocols are given in Appendix I.1.

Table 2/3. Ordinary-edit examples (2/3)

Table 3/3 (PDF p. 42). Ordinary-edit examples (3/3)

Additional Results

Encoder Similarity and Watermark Retention

We examine whether transformations that remain close to the original image in our learned representation are also more likely to preserve the watermark. We use the same 20 edit settings and configuration-specific detector settings as the robustness evaluation. For each method and edit setting, we pair the mean original–edited cosine similarity in our embedding with paired detection retention. Detection thresholds are calibrated on clean images and frozen across edits.

For our method, the Spearman correlations are 0.9880, 0.9985, and 0.9970 for (N,M)=(8,4)(N, M) = (8, 4), (16,8)(16, 8), and (32,12)(32, 12), respectively. The corresponding correlations are 0.7471 for Tree-Ring, 0.7572 for Gaussian Shading, and 0.7785 for SEAL. The stronger association for our method is consistent with its design: images that remain close in the encoder space are more likely to retain the same LSH code and clean keyed score. For the baselines, our embedding serves only as an external diagnostic and does not define their watermarking mechanisms.

Semantic Encoder Ablation for Clean Detection

To isolate the effect of the semantic encoder, we evaluate our learned encoder, DINOv2-base, CLIP ViT-B/32, DINOv3 L/16, and TPIPS-Embed (Qwen3VL-8B). Using the same 1,000 held-out COCO-2017 prompts and Stable Diffusion 2.1, we rerun our watermarking scheme and clean-image detection separately for each encoder. We evaluate (N,M)∈{(8,4),(16,8),(32,12)}(N, M) \in\{(8, 4), (16, 8), (32, 12)\} with B=1B = 1.

Each encoder uses normalized embeddings in its native dimension. We average over the Cartesian product of five random-hyperplane frames and five PRF keys, giving 25 replicates per encoder and configuration. We use the same configuration-specific detector settings as the main clean-detection evaluation: clean activation targets of 10%, 50%, and 75% for (N,M)=(8,4)(N, M) = (8, 4), (16,8)(16, 8), and (32,12)(32, 12), respectively, with Hamming radius one and five-fold clean-only calibration.

Figure 15 shows that our encoder obtains the highest mean AUROC in all three configurations: 0.802, 0.879, and 0.922. TPIPS-Embed is the strongest baseline, with mean AUROCs of 0.752, 0.849, and 0.908, respectively. At an empirical FPR of at most 0.2, our corresponding mean TPRs are 0.576, 0.808, and 0.921.

Figure 15. Semantic-encoder ablation for clean watermark detection. Mean ROC curves at B=1B = 1 for (N,M)=(8,4)(N, M) = (8, 4), (16,8)(16, 8), and (32,12)(32, 12). Using the same 1,000 held-out COCO-2017 prompts and Stable Diffusion 2.1, we rerun our watermarking scheme and clean-image detection separately for each encoder. The dotted diagonal is random guessing (TPR=FPR). Exact detection settings are given in Appendix § 20.§ 1.

Alternative Distortion-Free Sampling Rules

Several distortion-free sampling rules can use the same candidate images and semantic codes. We compare representative sampling (Algorithm 3), our deployed Gumbel rule in Algorithm 1 ([29]), and non-distortionary Tournament sampling ([12]). All three rules are distortion-free, but they use collisions among candidate codes differently.

To isolate the effect of the selection rule, we replay all three rules on the same Stable Diffusion 2.1 candidate pools and cached embeddings, keeping the candidate images and semantic representations fixed. We evaluate (N,M)∈{(8,4),(16,8),(32,12)}(N, M) \in\{(8, 4), (16, 8), (32, 12)\} on 1,000 held-out COCO-2017 prompts, with B=1B = 1 for single-image detection. Results average five random-hyperplane frames and five keys. Each rule uses the same detector configuration and clean-image threshold calibration, with its corresponding keyed score.

Table 12 shows that Gumbel sampling improves AUROC and TPR in all three configurations. At (N,M)=(32,12)(N, M) = (32, 12), for example, AUROC increases from 0.872 to 0.920, and TPR@20% increases from 0.852 to 0.919. This advantage also persists after editing: across all 51 evaluated transformations, attacked TPR increases from 0.679 to 0.726, while conditional retention remains nearly unchanged (0.787 versus 0.792). Tournament sampling is weaker than both rules for N≥16N \ge16.

(N,M)(N, M)Representative samplingGumbelTournament
AUROCTPRAUROCTPRAUROCTPR
(8, 4)0.7340.5000.8020.5760.7460.454
(16, 8)0.8210.7330.8790.8080.7620.545
(32, 12)0.8720.8520.9200.9190.7560.532

Table 12 (PDF p. 45). Clean detection under alternative distortion-free sampling rules. TPR is measured at thresholds calibrated to a 20% clean FPR target. Results use B=1B = 1 for single-image detection and average five random-hyperplane frames and five keys. Bold green and orange indicate the best and second-best values within each row.

These results show that Gumbel selection extracts a stronger detection signal than representative sampling from the same candidates. We therefore use Gumbel selection in our experiments, except where sampling rules are explicitly compared.

Black-Box Semantic Encoder Access

One strength of our method is that black-box access to the semantic encoder allows users to choose a representation suited to their desired notion of watermark preservation, without changing the watermarking algorithm. We demonstrate this flexibility using two encoders: CSD [39], which represents image style, and a depth-based encoder built from Depth Anything V2 Small (Yang et al., 2024a), which represents spatial layout. Both representations are normalized before hashing.

We hypothesize that these representations induce different watermark invariances. Style edits such as oil painting, watercolor, and pencil sketch change image appearance while approximately preserving spatial layout. They should therefore move images farther in CSD’s embedding space than in the depth-based representation, weakening the CSD-based watermark more strongly. Conversely, translation, cropping, and horizontal flipping alter spatial arrangement without directly restyling the image. We therefore expect higher retention for the CSD-based watermark than for the depth-based watermark under these transformations. These are empirical predictions: neither encoder guarantees invariance to every transformation that preserves its intended image property.

To test these predictions, we apply six transformations using the same watermarking algorithm and evaluation protocol. We use the first 100 prompts of the held-out COCO-2017 encoder-ablation panel with Stable Diffusion 2.1, (N,M)=(16,8)(N, M) = (16, 8), B=1B = 1, Hamming radius one, and a 50% clean activation target. We average detection results over five random-hyperplane frames and five keys. The style edits use InstructPix2Pix [6] with the instructions “Make it an oil painting”, “Turn it into a watercolor painting”, and “Turn it into a pencil sketch”. The spatial transformations are a 32-pixel translation per axis with black padding, a 75% off-center crop anchored at (0.65,0.35)(0.65, 0.35) and resized to the original resolution, and a horizontal flip. Candidate selection and boundary calibration use only unedited embeddings. Detection thresholds target 20% FPR on the 100 unedited clean images and remain fixed after editing; these images are also used to report clean FPR, rather than forming an independent calibration set. Mean cosine distance is 1−⟨z,z′⟩1 - \langle z, z' \rangle for normalized original and edited embeddings, averaged over all candidates and clean controls.

Table 13 reports detection, retention, and representation changes; Figure 16 illustrates the six transformations. The results show contrasting robustness profiles. Across the three style edits, CSD retains 25.2–36.9% of its initial detections, compared with 86.3–94.0% for the depth-based encoder. The spatial transformations reverse this pattern: CSD retains 55.6–89.5%, whereas depth retains 23.5–25.9%. The corresponding embedding changes follow the same pattern.

EncoderTransformationTPR (%)Retention (%)Cosine distance ↓\downarrow
CSD styleOriginal66.8——
Style changes
Oil painting28.536.90.199
Watercolor20.825.20.327
Pencil sketch23.025.80.452
Spatial transformations
Translation, 32 px44.663.60.041
Off-center crop, 75%38.955.60.059
Horizontal flip60.689.50.008
Depth layoutOriginal64.0——
Style changes
Oil painting60.794.00.009
Watercolor57.488.50.013
Pencil sketch56.286.30.021
Spatial transformations
Translation, 32 px21.324.40.504
Off-center crop, 75%19.323.50.402
Horizontal flip19.925.90.418

Table 13. Encoder-dependent watermark robustness. TPR uses thresholds calibrated at 20% FPR on unedited images and frozen after editing. Retention is conditional on initial detection; smaller cosine distance indicates greater representation stability.

Figure 16. Transformations used to test encoder-dependent robustness. Top: original, oil painting, watercolor, and pencil sketch. Bottom: original, 32-pixel translation, 75% off-center crop, and horizontal flip.

These results illustrate how users can choose an encoder according to their desired notion of watermark preservation. Consistent with Appendix § 21.§ 1, transformations that induce smaller cosine distances in the encoder’s representation also retain more watermark detections. A user seeking robustness to style transfer may choose a layout-based representation, while a user seeking a watermark tied to image style may choose CSD, preserving detection under spatial transformations such as horizontal flipping. The evaluated transformations demonstrate this flexibility without requiring access to encoder internals or changes to the watermarking algorithm.

Closed-API Black-Box Generator Access

To demonstrate that our sampling-only method can watermark outputs from a closed-API model without access to its internals, we use OpenAI’s GPT Image API (“gpt-image-2.5-sunburst”) strictly as a black-box generator. We use the first 300 held-out COCO-2017 prompts. For each prompt, we independently request eight watermark candidates and one clean control at low quality and 1024×10241024 \times1024 resolution, for 2700 API generations in total. We use (N,M)=(8,8)(N, M) = (8, 8), exact-code detection (rham=0r_{\mathrm{ham}} = 0), five independently sampled choices of the random-hyperplane LSH HRH_R, and five PRF keys. Reported curves and metrics average the resulting 25 key-hash repetitions.

GPT Image’s candidates are more clustered under our learned encoder than the open-model generations in Table 3: their mean within-prompt cosine similarity is 0.573, compared with 0.222 for SD2.1. GPT Image’s eight candidates occupy 5.419 distinct 8-bit codes on average across the 300 prompts and five choices of HRH_R, leaving substantial variation for keyed selection. Figure 17 illustrates this behavior for the prompt “Two women walk outside, both holding up umbrellas.” Its eight candidates occupy a mean of 6.6/8 distinct codes across the five choices of HRH_R.

Eight independent GPT Image generations from one prompt

Figure 17. Eight independent GPT Image generations from one prompt. All images were returned by separate API calls for “Two women walk outside, both holding up umbrellas.” Although their layouts are more similar than in the open-model generation panels, our encoder maps them to a mean of 6.6 distinct 8-bit codes across five independently sampled random-hyperplane LSH functions.

For individual images B=1B = 1, the mean AUROC is 0.815 and the mean TPR at a nominal 20% FPR is 64.6%. For batches of five B=5B = 5, we divide 300 prompts into 60 groups. Each group gives us a clean batch and a watermarked batch generated from the same five prompts. We average the five detection scores in each batch. The mean AUROC is 0.978, and the mean TPR at a nominal 20% FPR is 98.0%. The figure shows both ROC curves.

Figure 18. Closed-API watermark detection. Mean ROC curves for GPT Image with (N,M)=(8,8)(N, M) = (8, 8), averaged over five keys and five choices of HRH_R. Shading shows one standard deviation across the 25 repetitions. For B=5B = 5, each observation averages five watermarked or clean scores with identical prompt membership.

Text-to-Audio Watermarking

To demonstrate our sampling procedure in another generative domain, we apply Algorithm 1 to text-to-audio generation with AudioLDM 2 [31]. We generate 32 independent 10-second clips for each of 200 prompts from the AudioCaps test set [27] and encode them with LAION-CLAP (Wu et al., 2023). CLAP learns to align audio with text descriptions, rather than to separate independent clips generated from the same prompt or preserve their hash codes under audio processing. Nevertheless, its embeddings provide enough code variation for keyed selection and detection. We evaluate the same three (N,M)(N, M) configurations as in our image experiments, using a separately generated clean clip per prompt, five keys, and five hyperplane choices. Each prompt provides a single candidate pool, so for B>1B > 1 the detector averages the scores of BB watermarked clips generated from BB different prompts, and likewise for clean clips.

Table 14 and Figure 19 show that single-clip detection improves across the evaluated configurations, although it remains weaker than in our image experiments. Aggregating evidence over clips strengthens it substantially: with B=10B = 10, AUROC reaches 0.964, 0.998, and 1.000 for the three configurations. We also apply MP3 compression at 128 kbps, Gaussian noise at 30 dB SNR, a 3 dB reduction in gain, and a 100 ms shift, keeping the clean-calibrated thresholds fixed. At (N,M)=(32,12)(N, M) = (32, 12), these transformations retain 56.2%, 59.2%, 60.7%, and 44.5% of initially detected clips, respectively. Even the gain reduction preserves the complete 12-bit CLAP code for only 36% of selected clips. Thus, the same selection procedure yields a detectable audio watermark, while this encoder’s code instability limits robustness to benign audio processing.

(N,M)(N, M)B=1B = 1 AUROCB=1B = 1 TPR@20%B=5B = 5 AUROCB=5B = 5 TPR@20%B=10B = 10 AUROCB=10B = 10 TPR@20%
(8, 4)0.7110.3970.9030.8460.9640.956
(16, 8)0.8170.6700.9780.9830.9981.000
(32, 12)0.8620.7820.9900.9931.0001.000

Table 14. Audio watermark detection. Means over 25 key–hyperplane repetitions on 200 prompts. For B>1B > 1, each batch contains clips generated from BB different prompts. We calibrate the detector on clean clips and report the true-positive rate at a nominal 20% clean false-positive rate.

Figure 19. Audio watermark ROC curves. Mean clean-detection ROC over 25 key–hyperplane repetitions on 200 prompts for B=1B = 1, 5, and 10 clips per decision; shaded bands show one standard deviation across repetitions. The dashed line marks the 20% false-positive rate of Table 14. With 200 prompts, B=10B = 10 yields 20 batches per repetition, so its false-positive rate moves in steps of 0.05.

Generation Quality and Prompt Alignment

To assess generation quality and prompt alignment, we evaluate Stable Diffusion 2.1 Base at a resolution of 512×512512 \times512. We generate one image for each of 5,000 COCO-2017 prompts and measure FID [21] against 5,000 COCO-2017 validation images. We measure OpenCLIP similarity [9] separately on 1,000 prompts. For every watermarking configuration, we generate an unwatermarked reference set using the same model, prompts, and sampling settings.

Table 15 reports generation quality and prompt alignment. Across our three configurations, FID improves by 0.38–1.82% relative to matched clean generation, while OpenCLIP similarity changes by at most 0.03%. Tree-Ring and Gaussian Shading change FID by −0.16% and +0.05%, respectively, while SEAL changes it by +1.78%. Relative to their matched clean reference sets, the watermarked populations have comparable FID to real COCO images and nearly unchanged image-prompt similarity. These complementary diagnostics are consistent with the direct matched-population FID results in the main text.

MethodWatermarkedMatched cleanRelative change
FID (lower is better)
Ours (N,M)=(8,4)(N, M) = (8, 4)26.84026.943−0.38%
Ours (N,M)=(16,8)(N, M) = (16, 8)26.45326.943−1.82%
Ours (N,M)=(32,12)(N, M) = (32, 12)26.61526.943−1.22%
Tree-Ring24.85324.894−0.16%
Gaussian Shading24.90724.894+0.05%
SEAL25.64925.201+1.78%
OpenCLIP image prompt similarity (higher is better)
Ours (N,M)=(8,4)(N, M) = (8, 4)0.31540.31540.00%
Ours (N,M)=(16,8)(N, M) = (16, 8)0.31550.3155−0.01%
Ours (N,M)=(32,12)(N, M) = (32, 12)0.31560.3157−0.03%
Tree-Ring0.31440.3145−0.03%
Gaussian Shading0.31550.3145+0.31%
SEAL0.31380.3140−0.06%

Table 15. Distributional quality on 5,000 unique COCO-2017 validation prompts and image–prompt alignment on 1,000 prompts. Each method is compared with its matched clean control. Relative change is 100(swm/sclean−1)100(s_{\mathrm{wm}}/s_{\mathrm{clean}} - 1). This table supports the quality claim in § 7.1.

Detection under Mixed-Source Batches

We test how our detector performs in real-world batches that mix watermarked and unwatermarked images or images from different watermarking sources, and whether it provides a signal for distinguishing those sources. We start with ten images watermarked under the first key and progressively replace them, one image at a time, with images watermarked under a second, independent key. At each mixture, we test the batch under both keys. We also repeat the replacement procedure with unwatermarked images and evaluate detection under the first key.

We use Stable Diffusion 2.1 candidates for 1,000 COCO-2017 prompts and apply Algorithm 1 with the original encoder and detector settings. We evaluate all three configurations over five hyperplane frames and five key pairs. Both keys share the encoder and hyperplane frame. We calibrate thresholds on unwatermarked batches at nominal FPRs of 1%, 5%, and 20%, then keep them fixed across mixtures.

Figure 21 shows that detection decreases gradually as we replace images carrying the tested key with unwatermarked images. For (N,M)=(32,12)(N, M) = (32, 12), batches containing five watermarked and five unwatermarked images achieve detection rates of 30.2%, 60.6%, and 87.9% at the three operating points, respectively. Replacing the five unwatermarked images with images carrying K2K_2 gives similar rates of 31.3%, 59.6%, and 86.4%. Detection therefore remains informative even when only half the batch carries the tested watermark.

![Six-panel detection plot titled “Detection Under Mixed-Source Batches (B=10B = 10)”. The left column is labeled “Remaining images unwatermarked” and the right column “Remaining images under independent key K2K_2”. Rows are labeled N=8,M=4N = 8, M = 4; N=16,M=8N = 16, M = 8; and N=32,M=12N = 32, M = 12. The vertical axis is “TPR (%)”; the horizontal axis is “K1K_1-watermarked images in the batch (of 10)” with values 0 through 10. The legend shows FPR = 1%, FPR = 5%, and FPR = 20%.] (media/fig21.png) bbox=0.176,0.212,0.816,0.657

Figure 21. Detection under mixed-source batches. We vary the number of K1K_1 images in a ten-image batch; the rest are unwatermarked (left) or carry K2K_2 (right). Colors denote nominal clean-calibrated FPR. Shading shows 95% intervals from 1,000 prompt bootstrap resamples with thresholds recomputed and keys and frames fixed. Resampling ties can bias calibrated quantities slightly downward. At zero K1K_1 images, detection rates measure false alarms.

Figure 20 shows detection under both keys. At 5% nominal clean FPR per key, a batch containing ten K1K_1 images triggers only the K1K_1 detector in 91.9% of cases. As we progressively replace K1K_1 images with K2K_2 images, detection under K1K_1 decreases gradually while detection under K2K_2 increases. When the batch contains five images from each key, both detectors trigger in 36.6% of cases. Because both keys select from the same candidates, batches containing only K2K_2 images are also detected under K1K_1 slightly above the nominal rate (2.2%, 7.6%, and 22.5% at the three operating points). We therefore conclude that the detector provides a signal for distinguishing known watermarking sources at the batch level.

Figure 20. Detection under two known keys. We score the same mixed batches with K1K_1 (solid) and K2K_2 (dashed), using (N,M)=(32,12)(N, M) = (32, 12) and B=10B = 10. FPR labels refer to per-key calibration on unwatermarked batches, not family-wise or cross-key error rates. TK1T_{K_1} and TK2T_{K_2} denote detection using keys K1K_1 and K2K_2, respectively.

Choice of Embedding Discretization

Our method discretizes semantic embeddings with random-hyperplane LSH. We compare LSH with two alternatives that also map each embedding to one of 2M2^M codes. A good discretization should spread independent generations of the same prompt across many codes, because Lemma 20 ties the expected keyed score of the selected code to the multiplicities ncn_c of the occupied codes: with ideal keyed scores, the expected gap of the selected score above 12\frac{1}{2} is Δ=∑cncN+nc−12\Delta= \sum_c \frac{n_c}{N+n_c} - \frac{1}{2}, which grows as the candidates spread over more codes. A good discretization should also keep an image’s code unchanged under content-preserving transformations.

K-means assigns each embedding to the nearest of 2M2^M centroids, which we fit on 1,000 real COCO-2017 validation images. Coordinate sign takes the sign of the first MM embedding coordinates; it equals LSH with the random hyperplanes replaced by the standard basis. We average LSH over five hyperplane frames and K-means over five clustering seeds.

Only LSH guarantees that two embeddings at angle θ\theta share a bit with probability 1−θ/π1-\theta/\pi, regardless of where they lie. This relation links semantic entropy to code occupancy (Lemma 19) and encoder stability to score preservation (Lemma 22). K-means cells depend on the calibration data, and coordinate-sign collisions depend on the position of the embeddings relative to the coordinate axes.

We measure code occupancy on Stable Diffusion 2.1 generations for the 1,000 held-out COCO-2017 prompts, with up to N=320N = 320 candidates per prompt, and compute Δ\Delta from the code multiplicities of each batch. We measure code stability on the images and transformations of the robustness evaluation (Appendix § 20.§ 1) as the percentage of images whose full MM-bit code changes.

LSH occupies the most codes in every configuration (Table 16, Figure 22). The margin is small at the paper’s candidate counts: for (N,M)=(16,8)(N, M) = (16, 8), the expected gap is 0.414 for LSH, 0.408 for K-means, and 0.407 for coordinate sign. The margin grows with NN: at M=8M = 8 and N=320N = 320, LSH occupies 84.9 codes, compared with 79.0 for coordinate sign and 72.0 for K-means. In contrast, K-means changes fewer codes under every transformation family (Figure 23), for example 8.1% versus 12.3% for LSH under JPEG compression at M=8M = 8, while LSH and coordinate sign behave alike.

(N,M)(N, M)LSH (ours) UULSH (ours) Δ\DeltaK-means UUK-means Δ\DeltaCoordinate sign UUCoordinate sign Δ\Delta
(8,4)(8, 4)5.290.3095.170.3045.180.304
(16,8)(16, 8)12.940.41412.410.40812.360.407
(32,12)(32, 12)27.710.459——26.880.456

Table 16. Code occupancy at the paper configurations. UU is the mean number of occupied codes among the NN candidates of a batch, and Δ\Delta is the expected keyed-score gap of Gumbel selection (Lemma 20). We evaluate K-means for M∈{4,8}M \in\{4, 8\}.

Figure 22. Code occupancy as the number of candidates grows. Mean number of distinct codes among NN candidates for the same prompt. Shading spans the five hyperplane frames (LSH) and five clustering seeds (K-means). Higher is better.

Figure 23. Code changes under transformations at M=8M=8. Percentage of images whose full code changes, averaged over the settings in each family. Lower is better.

No discretization is best on both criteria. We use LSH because it yields the largest expected keyed-score gap and because it is the only one whose collision probability depends solely on the angle between embeddings, which our analysis requires. Randomization may also make adaptive attacks harder. If we sample the hyperplanes privately together with the key, public data reveals nothing about the code boundaries. In contrast, the coordinate axes are fixed, and K-means centroids depend mostly on the data distribution, so an attacker could approximate them by clustering similar images. We leave evaluating this to future work.

Detection after Content-Changing Edits

Our encoder is trained to preserve detection under edits that preserve the main content. We evaluate what happens when an edit replaces the main subject, scene, or action by comparing edited watermarked images with edited unwatermarked images.

We use 180 COCO-2017 prompts with Stable Diffusion 2.1, with 60 prompts per edit category. Qwen3-VL-8B-Instruct [4] writes one editing instruction per prompt from the prompt text. For each prompt, Qwen-Image-Edit (Wu et al., 2025) applies the same instruction and edit seed to the Gumbel-selected images and to an independently generated clean image. We include all renders without screening individual edits for success. We use our single-image detector settings from Table 5, with a fixed key and hyperplane frame. Tree-Ring, Gaussian Shading, and SEAL receive the same instructions and edit seeds and use their own clean-calibrated detectors.

Table 17 reports detection on both edited populations. For (N,M)=(32,12)(N, M) = (32, 12), edited TPR is 21.7–23.3%, compared with 16.7–21.7% on edited clean images. AUROC for this configuration is 0.536 for subject replacement, 0.463 for scene replacement, and 0.541 for action changes. Paired retention is 22.4–25.9% across these edits. For comparison, the separate content-preserving panel in Figure 12 reports 94.7% retention after background cleanup for the same configuration. Figure 24 shows an example from each category.

EditOurs (8, 4)Ours (16, 8)Ours (32, 12)Tree-RingGaussian ShadingSEAL
Edited TPR
Subject replacement20.015.023.350.038.33.3
Scene replacement25.023.321.740.025.011.7
Action change38.328.323.331.725.010.0
Edited clean FPR
Subject replacement16.728.316.731.723.38.3
Scene replacement15.016.716.728.318.313.3
Action change25.015.021.713.315.06.7
Paired retention
Subject replacement31.814.625.950.038.33.3
Scene replacement42.326.522.440.025.011.7
Action change54.837.025.031.725.010.0

Table 17. Detection after content-changing edits. All values are percentages at 20% nominal clean FPR. Each category has 60 prompts. Paired retention is the fraction of originally detected watermarked images that remain detected after editing. SEAL achieves 8.1% on its unedited calibration set.

Content-changing edit examples

Figure 24. Content-changing edit examples. The first prompt by panel rank in each edit category for (N,M)=(32,12)(N, M) = (32, 12). Each row shows the same instruction and seed applied to the watermarked and clean generations for that prompt.

Extended Limitations and Future Directions

This appendix expands on the limitations summarized in the main paper. We discuss single-image detection under limited black-box sampling, computational cost, fixed-key reuse, adaptive attacks, and extension beyond images.

Single-Image Detection under Limited Black-Box Sampling

Our image implementation has weaker clean single-image detection than methods that insert a signal directly into the diffusion process (Table 2). Aggregating scores across independently generated outputs substantially improves detection, in agreement with Lemma 21. However, aggregation is unavailable when each image must be judged independently.

The following theorem shows that any black-box sampling method selecting one of NN independent, unmodified generator outputs can amplify the single-image detection probability over the false-positive rate by at most a factor of NN. This O(N)O(N) ceiling belongs to black-box sampling, not our construction.

Theorem 23 (Single-image detection under limited sampling). Fix a prompt pp, secret key κ\kappa, and detector Dκ:X→{0,1}D_{\kappa} : \mathcal{X} \to\{0, 1\}. Let X1,…,XN∼i.i.d.PG(⋅∣p)X_1, \ldots, X_N \mathrel{\overset{\mathrm{i.i.d.}}{\sim}} P_G(\cdot\mid p). Consider any possibly randomized watermarking algorithm that returns XI⋆X_{I^\star}, where I⋆∈[N]I^\star\in[N] may depend on the candidates and key. For a clean image X∼PG(⋅∣p)X \sim P_G(\cdot\mid p), define

α=FPR⁡κ,p:=Pr⁡[Dκ(X)=1],TPR⁡κ,p:=Pr⁡[Dκ(XI⋆)=1].\alpha= \operatorname{FPR}_{\kappa,p} := \Pr[D_{\kappa}(X) = 1], \qquad\operatorname{TPR}_{\kappa,p} := \Pr[D_{\kappa}(X_{I^\star}) = 1].

Then

TPRκ,p≤Nα=N FPRκ,p.\mathrm{TPR}_{\kappa,p} \le N\alpha= N\,\mathrm{FPR}_{\kappa,p}.

Proof. The algorithm can return a detected image only if at least one candidate is detected. Therefore,

{Dκ(XI⋆)=1}⊆⋃i=1N{Dκ(Xi)=1}.\{D_{\kappa}(X_{I^\star}) = 1\} \subseteq\bigcup_{i=1}^{N}\{D_{\kappa}(X_i) = 1\}.
TPRκ,p≤Pr⁡[⋃i=1N{Dκ(Xi)=1}]≤∑i=1NPr⁡[Dκ(Xi)=1]=Nα.\mathrm{TPR}_{\kappa,p} \le\Pr\left[\bigcup_{i=1}^{N}\{D_{\kappa}(X_i) = 1\}\right] \le\sum_{i=1}^{N}\Pr[D_{\kappa}(X_i) = 1] = N\alpha.

This bound applies to any method selecting among NN independent, unmodified generator outputs, regardless of distortion-freeness, encoder, or scoring rule. For α>0\alpha> 0, achieving TPRκ,p≥c>0\mathrm{TPR}_{\kappa,p} \ge c > 0 requires N≥c/αN \ge c/\alpha. Thus, targeting α≤e−t\alpha\le e^{-t} requires N≥cetN \ge ce^{t}: exponentially small FPR at constant TPR requires exponentially many generator queries.

This does not contradict the exponential improvement from aggregating independently generated images in Lemma 21: the query limit concerns one sampled output, whereas batch detection accumulates evidence across multiple outputs.

Computational Cost and Early Candidate Selection

Our strongest configuration generates 32 complete candidates for every released image. This cost may be prohibitive in large-scale or latency-sensitive applications.

Recent work suggests that earlier candidate selection may be feasible for image diffusion models. Guo et al. [18] show that intermediate denoiser activations can predict final-image quality and allow unpromising trajectories to be terminated early. Habibian et al. [19] show that low-resolution diffusion features carry semantic layout and can be reused across denoising steps. These results do not establish that our final semantic code can be predicted before generation is complete, but they show that relevant structural and semantic information is available at intermediate stages.

A promising direction is to predict the final semantic code or keyed score from intermediate features, discard unlikely candidates, and complete only promising trajectories. Any such optimization must preserve the selection probabilities required by Theorem 3. Prediction errors may discard trajectories that would have received high final scores, weakening detection even when image quality is unchanged. A useful evaluation must therefore report expected completed denoising steps and wall-clock latency together with detection, image quality, and distribution shift. This would expose the full computation–detection tradeoff rather than treating early termination as a speedup in isolation.

Distribution Preservation under Fixed-Key Reuse

Theorem 3 establishes equality with the original generator distribution after averaging over the secret key and the sampler’s internal randomness. It does not establish distribution preservation separately for every fixed key. A fixed key deliberately favors particular semantic codes. Theorem 4 characterizes the resulting detection probability only after taking expectations over the key and random hyperplanes; it does not imply the same guarantee for every fixed deployed key.

The matched direct-FID experiment shows that our method produces the smallest measured distribution shift among the evaluated methods, while the same-prompt diversity experiment shows that it stays within 1.3% of the diversity of matched clean generations (Table 1). These experiments provide evidence that fixed-key reuse preserves quality and diversity in the evaluated setting, but they do not constitute a general per-key guarantee. Establishing fixed-key bounds in terms of the number of candidates, code occupancy, and encoder geometry remains an important theoretical direction.

MethodFID ↓Relative FID ↓DINOv2-base diversity change
Clean vs. clean4.641.000–
Ours (N,M)=(8,4)(N, M) = (8, 4)4.751.024−0.5%
Ours (N,M)=(16,8)(N, M) = (16, 8)5.041.086−1.2%
Ours (N,M)=(32,12)(N, M) = (32, 12)5.221.125−1.0%
Tree-Ring5.611.209+0.4%
Gaussian Shading5.421.168−9.1%
SEAL6.881.483−13.2%

Table 1. Fixed-key generation quality and diversity. FID between 10,000 matched watermarked and clean generations from 1,000 COCO-2017 prompts; Relative FID divides by the clean-vs-clean reference of 4.64. Diversity change is 100(Vendiwm/Vendiclean−1)100(V_{\mathrm{endiwm}}/V_{\mathrm{endiclean}} - 1) in DINOv2-base space over 60 prompts with 60 generations each.

Adaptive Removal and Forgery Attacks

Our robustness evaluation considers non-adaptive appearance, compression, geometric, and generative transformations. Figure 2 summarizes the resulting detection and retention rates, while Appendix § 20.§ 2 reports the individual settings, detector configurations, and achieved false-positive rates.

These experiments do not establish security against an attacker who deliberately targets the encoder or detector. Because our score is bound to the semantic representation of the final image rather than to diffusion-specific latent noise, the method may be less susceptible to reprompting attacks that transfer a watermark to semantically different content, as studied by Müller et al. [32] and addressed by SemBind (Zhang et al. (2026)). This possibility has not been evaluated directly. An adaptive attacker may instead attempt to cross encoder or hash boundaries while preserving apparent content, remove a high-scoring code, or produce altered content that retains the original score. Lemma 22 characterizes non-adversarial behavior under representation similarity and should not be interpreted as a security guarantee against such attacks.

Extension beyond Images

The framework treats the generator as a sampling oracle and is not tied to a particular architecture or output modality.

Each new domain requires a semantic encoder that separates independent outputs while remaining stable under the relevant content-preserving transformations. Future work should determine whether such encoders provide sufficient code occupancy for detection, whether their geometry supports robustness to realistic transformations, and whether sampling enough candidates remains computationally practical.

References

  1. [1]Anthropic. How Claude’s text watermark works, August 2026. URL https://www.anthropic.com/news/claude-text-watermark.
  2. [2]Kasra Arabi, R. Teal Witter, Chinmay Hegde, and Niv Cohen. SEAL: Semantic aware image watermarking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 16196–16205, 2025. URL https://openaccess.thecvf.com/content/ICCV2025/html/Arabi_SEAL_Semantic_Aware_Image_Watermarking_ICCV_2025_paper.html.arxiv.org/abs/2503.12172
  3. [3]Dara Bahri and John Wieting. A watermark for black-box language models. Transactions on Machine Learning Research, 2026. URL https://arxiv.org/abs/2410.02099.
  4. [4]Shuai Bai et al. Qwen3-VL technical report, 2025. URL https://arxiv.org/abs/2511.21631.
  5. [5]Black Forest Labs. Announcing black forest labs, 2024. URL https://bfl.ai/blog/24-08-01-bfl.
  6. [6]Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to follow image editing instructions. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 18392–18402. IEEE, 2023.arxiv.org/abs/2211.09800
  7. [7]Moses S. Charikar. Similarity estimation techniques from rounding algorithms. In Proceedings of the 34th Annual ACM Symposium on Theory of Computing, pp. 380–388, 2002. doi: 10.1145/509907.509965.DOI
  8. [8]Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. URL https://arxiv.org/abs/1504.00325.
  9. [9]Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2818–2829, 2023. URL https://openaccess.thecvf.com/content/CVPR2023/html/Cherti_Reproducible_Scaling_Laws_for_Contrastive_Language-Image_Learning_CVPR_2023_paper.html.DOI
  10. [11]Miranda Christ, Sam Gunn, and Or Zamir. Undetectable watermarks for language models. In Proceedings of the Thirty-Seventh Conference on Learning Theory, volume 247 of Proceedings of Machine Learning Research, pp. 1125–1139. PMLR, 2024. URL https://proceedings.mlr.press/v247/christ24a.html.arxiv.org/abs/2306.09194
  11. [12]Sumanth Dathathri, Abigail See, Sumedh Ghaisas, Po-Sen Huang, Rob McAdam, Johannes Welbl, Vandana Bachani, Alex Kaskasoli, Robert Stanforth, Tatiana Matejovicova, Jamie Hayes, Nidhi Vyas, Majd Al Merey, Jonah Brown-Cohen, Rudy Bunel, Borja Balle, Taylan Cemgil, Zahra Ahmed, Kitty Stackpoole, Ilia Shumailov, Ciprian Baetu, Sven Gowal, Demis Hassabis, and Pushmeet Kohli. Scalable watermarking for identifying large language model outputs. Nature, 634: 818–823, 2024. doi: 10.1038/s41586-024-08025-4.DOI
  12. [13]Liancheng Fang, Aiwei Liu, Henry Peng Zou, Yankai Chen, Hengrui Zhang, Zhongfen Deng, and Philip S. Yu. MUSE: Model-agnostic tabular watermarking via multi-sample selection. In International Conference on Learning Representations, 2026.arxiv.org/abs/2505.24267
  13. [14]Zheng Gao, Yifan Yang, Xiaoyu Li, Xiaoyan Feng, Haoran Fan, Yang Song, and Jiaojiao Jiang. SLICE: Semantic latent injection via compartmentalized embedding for image watermarking. arXiv preprint arXiv:2603.12749, 2026. URL https://arxiv.org/abs/2603.12749.
  14. [15]Oded Goldreich, Shafi Goldwasser, and Silvio Micali. How to construct random functions. Journal of the ACM, 33(4):792–807, 1986. doi: 10.1145/6490.6503.DOI
  15. [16]Nir Goren, Oren Katzir, Abhinav Nakarmi, Eyal Ronen, Mahmood Sharif, and Or Patashnik. Noiseprints: Distortion-free watermarks for authorship in private diffusion models. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=gWHCXvPDBd.arxiv.org/abs/2510.13793
  16. [17]Sam Gunn, Xuandong Zhao, and Dawn Song. An undetectable watermark for generative image models. In International Conference on Learning Representations, 2025. URL https://arxiv.org/abs/2410.07369.
  17. [18]Huanlei Guo, Hongxin Wei, and Bingyi Jing. Toward early quality assessment of text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 38410–38419, 2026. URL https://openaccess.thecvf.com/content/CVPR2026/html/Guo_Toward_Early_Quality_Assessment_of_Text-to-Image_Diffusion_Models_CVPR_2026_paper.html.arxiv.org/abs/2603.02829
  18. [19]Amirhossein Habibian, Amir Ghodrati, Noor Fathima, Guillaume Sautiere, Risheek Garrepalli, Fatih Porikli, and Jens Petersen. Clockwork diffusion: Efficient generation with model-step distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8352–8361, 2024. URL https://openaccess.thecvf.com/content/CVPR2024/html/Habibian_Clockwork_Diffusion_Efficient_Generation_With_Model-Step_Distillation_CVPR_2024_paper.html.arxiv.org/abs/2312.08128
  19. [20]R. W. Hamming. Error detecting and error correcting codes. Bell System Technical Journal, 29(2): 147–160, 1950. doi: 10.1002/j.1538-7305.1950.tb00463.x.DOI
  20. [21]Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems, volume 30, 2017.
  21. [24]Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=nZeVKeeFYf9.arxiv.org/abs/2106.09685
  22. [26]Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: Towards removing the curse of dimensionality. In Proceedings of the Thirtieth Annual ACM Symposium on Theory of Computing, pp. 604–613, 1998. doi: 10.1145/276698.276876.DOI
  23. [27]Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. AudioCaps: Generating captions for audios in the wild. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 119–132, 2019.DOI
  24. [28]John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, and Tom Goldstein. A watermark for large language models. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp. 17061–17084, 2023. URL https://proceedings.mlr.press/v202/kirchenbauer23a.html.arxiv.org/abs/2301.10226
  25. [29]Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto, and Percy Liang. Robust distortion-free watermarks for language models. Transactions on Machine Learning Research, 2024. URL https://arxiv.org/abs/2307.15593.
  26. [30]Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In European Conference on Computer Vision, pp. 740–755. Springer, 2014.arxiv.org/abs/1405.0312
  27. [31]Haohe Liu, Yi Yuan, Xubo Liu, Xinhao Mei, Qiuqiang Kong, Qiao Tian, Yuping Wang, Wenwu Wang, Yuxuan Wang, and Mark D. Plumbley. AudioLDM 2: Learning holistic audio generation with self-supervised pretraining. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:2871–2883, 2024.
  28. [32]Andreas Müller, Denis Lukovnikov, Jonas Thietke, Asja Fischer, and Erwin Quiring. Black-box forgery attacks on semantic watermarks for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20937–20946, 2025. URL https://arxiv.org/abs/2412.03283.
  29. [33]Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yu Huang, Shang-Wen Li, Isham Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOV2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. URL https://arxiv.org/abs/2304.07194.
  30. [34]Gaurav Parmar, Or Patashnik, Daniil Ostashev, Kuan-Chieh Wang, Kfir Aberman, Srinivasa Narasimhan, and Jun-Yan Zhu. Scaling group inference for diverse and high-quality generation. In International Conference on Learning Representations, 2026. URL https://proceedings.iclr.cc/paper_files/paper/2026/hash/9132fb7713ccc4860399d41695be6822-Abstract-Conference.html.arxiv.org/abs/2508.15773
  31. [35]Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In International Conference on Learning Representations, 2024. URL https://proceedings.iclr.cc/paper_files/paper/2024/hash/081b080868e4733ae3e7ad019fe8d172f-Abstract-Conference.html.arxiv.org/abs/2307.01952
  32. [36]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pp. 8748–8763. PMLR, 2021.arxiv.org/abs/2103.00020
  33. [37]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10684–10695, 2022. URL https://openaccess.thecvf.com/content/CVPR2022/html/Rombach_High-Resolution_Image_Synthesis_With_Latent_Diffusion_Models_CVPR_2022_paper.html.arxiv.org/abs/2112.10752
  34. [38]Oriane Siméoni, Huy V. Vo, Maximilian Seit­zer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julien Mairal, Hervé Jégou, Patrick Labatut, and Piotr Bojanowski. DINOv3. arXiv preprint arXiv:2508.10104, 2025. URL https://arxiv.org/abs/2508.10104.
  35. [39]Gowthami Somepalli, Anubh av Gupta, Kamal Gupta, Shramay Palta, Micah Goldblum, Jonas Geiping, Abhinav Shrivastava, and Tom Goldstein. Measuring style similarity in diffusion models. arXiv preprint arXiv:2404.01292, 2024.arxiv.org/abs/2404.01292
  36. [40]Stability AI. Stable diffusion v2.1 and dreamstudio updates, 2022. URL https://stability.ai/news-updates/stablediffusion-2-1-release-7-dec-2022.
  37. [41]Vinod Vaikuntanathan and Or Zamir. Undetectable conversations between AI agents via pseudorandom noise-resilient key exchange, 2026. URL https://arxiv.org/abs/2604.04757.
  38. [42]Sheng-Yu Wang, Yotam Nitzan, Aaron Hertzmann, Jun-Yan Zhu, Eli Shechtman, Alexei A. Efros, and Richard Zhang. The many senses of visual similarity: A text-prompted image perceptual metric. arXiv preprint arXiv:2607.18237, 2026. URL https://arxiv.org/abs/2607.18237.
  39. [43]Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. In Advances in Neural Information Processing Systems, volume 36, 2023.DOI

Paper details

vlnc:W_n4s4ajes
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Generative AI assistants

View statement
Looking for related papers…
Contents