acceptodds
Under review as a conference paper at ICLR 2027

Persistent Watermarking of Text-to-Image Models

Abstract

Text-to-image (T2I) generation is gaining increasing popularity with the general public, motivating the development of reliable mechanisms for copyrighting such models given their expensive training costs. An adversary may obtain and reuse a pretrained T2I model without authorization, and then serve a modified version through an API service. Such modifications may arise from ordinary downstream adaptation or deliberate attempts to erase ownership, including input-prompt preprocessing, model fine-tuning, and output post-processing. From the model owner's perspective, a key challenge is therefore to embed trigger data that remain persistent under such changes while preserving the model's normal image-generation capabilities. In this work, we propose a contrastive-style watermarking objective with a term that explicitly encourages the watermarked model to behave *differently* from the original model on trigger inputs. Experiments show substantially stronger trigger-data persistence than prior methods across a wide range of common downstream modifications and deliberate attempts to weaken the watermark, resulting in substantially higher detection rates, often approaching 100% TPR@FPR<.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Text-to-Image Model, Watermark, Copyright, Contrastive Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ak5TOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies black-box ownership verification for text-to-image models under downstream model and generation-pipeline modifications. Building on existing text-triggered semantic watermarking approaches, it constructs mappings between ordinary-word trigger phrases and deliberately unrelated target objects. It adds a negative loss term that encourages divergence from the original model on trigger data while preserving regular generation behavior. The resulting models are evaluated under a broad range of transformations, with reported improvements in watermark persistence over the evaluated baselines.

Critical strengths

Broad experimental coverage. The paper devotes substantial effort to evaluating 21 model-weight modifications, including fine-tuning, distillation, quantization, and pruning. Additional experiments cover different architectures and targets, prompt and output processing, and trigger-construction ablations. Reporting regular-generation quality alongside watermark detection is also useful. These experiments provide valuable empirical evidence under the tested transformations.

Critical weaknesses

W1. The trigger construction offers limited incremental contribution and leaves its security trade-off insufficiently assessed. Similar ingredients already exist in semantic watermarking: Watermarking Diffusion Model evaluates the ordinary-word trigger "chicken" [1], and WatermarkDM uses predefined photographic targets, not only QR codes [2]. Although the deliberately unrelated phrase–target pairing is a specific design choice here, its contribution appears to be a modest adaptation of existing components rather than a substantial methodological advance. Moreover, compared with rare identifiers or random strings, ordinary-word triggers may be more discoverable through active probing guided by natural-language priors. Avoiding input filtering therefore does not necessarily imply stronger trigger secrecy. The paper discusses the cost of exhaustive enumeration and the use of combinatorial phrases, but does not adequately assess this comparative discovery risk or the associated security trade-off.

W2. The connection between the negative term and persistence under specific modifications remains insufficiently explained. Encouraging trigger predictions to differ from the original model does not by itself establish why the resulting mapping should survive fine-tuning, transfer through distillation, or resist quantization. Appendix B.2 provides gradient-cosine observations, and B.3 provides residual-reconstruction evidence, but neither establishes how the negative term causes the relevant behavior. In particular, preserving an existing watermark during fine-tuning is different from transferring it to a clean student trained only on regular teacher signals. The reconstruction probe uses trigger residuals for fitting/selection, whereas the actual student does not access them. Appendix F offers a useful conditional result for a fixed-feature model under bounded common perturbations, but it does not isolate reduced modification sensitivity from improved initial fitting margins or explain the actual distillation and quantization processes. Controlled comparisons with matched initial continuous trigger error and regular utility would help establish the source of the persistence advantage.

W3. The attack-coverage claim is too strong, particularly given the missing subject-level adaptive evaluation. The statement "we consider the comprehensive set of transformations adversaries may use" overstates the scope of the evaluation, which does not establish coverage of the broader space of adaptive evasion strategies. Although Appendix C evaluates global prompt–image consistency filtering, the proposed watermark changes the depicted subject, while agreement with the background and other prompt details may obscure that mismatch in a global assessment. A subject-level consistency filter would therefore be more directly targeted at this design and could constitute a stronger adaptive attack. The paper neither discusses nor evaluates this specific strategy. Its effectiveness remains to be tested, and the robustness claims should be restricted to the evaluated transformations.

[1] Liu et al. Watermarking Diffusion Model.

[2] Zhao et al. A Recipe for Watermarking Diffusion Models.

Questions for authors
  1. Regarding W1: What substantial contribution does the trigger construction provide beyond existing ordinary-word triggers and predefined image targets? Compared with rare identifiers or random strings, how does it affect trigger discoverability under active probing?
  2. Regarding W2: Please explain the rationale for the specific negative term in greater detail. How do the mechanisms of fine-tuning, distillation, and quantization motivate encouraging trigger behavior to diverge from the original model? What analysis or targeted evidence links this design to watermark persistence after modification, beyond showing that adding the term improves empirical performance?
  3. Regarding W3: Can the authors evaluate subject-level consistency filtering as a design-specific adaptive attack?
Reviewer_x6DMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the protection of intellectual property (IP) in text-to-image (T2I) models through robust watermarking. Existing methods typically embed secret trigger–image associations into pretrained models, but these associations are vulnerable to model modifications, including fine-tuning, quantization, pruning, and distillation. To address this limitation, the authors propose a contrastive watermarking objective that encourages the watermarked model to produce distinct outputs from the original model for trigger inputs while preserving its behavior on benign inputs. Experiments across various model architectures and modification techniques demonstrate that the proposed method consistently outperforms existing baselines in watermark robustness.

Critical strengths
  1. The authors propose a simple yet effective method for protecting the IP of T2I models, demonstrating its robustness against model modifications. In addition, the proposed contrastive loss can be incorporated into existing watermarking methods to enhance robustness.

  2. The authors systematically evaluate the proposed method against 21 downstream model modifications, including fine-tuning, LoRA, quantization, pruning, and distillation. The method achieves at least 90% trigger detection in 19 of the 21 settings, substantially outperforming existing baselines.

  3. Beyond Stable Diffusion, the authors evaluate their approach on five additional T2I models spanning diffusion transformers, flow-based models, masked-token models, and autoregressive models. The consistent improvements over baselines suggest that the proposed objective is not limited to a particular model architecture

Critical weaknesses
  1. While the paper reports a promising 74% trigger detection rate after KD, several concerns limit the strength of this claim. First, the qualitative results in Figure 19 appear inconsistent with the reported quantitative performance. The proposed method's generated image does not clearly preserve the target object (Cubone), raising concerns about whether the automated detection metric reliably reflects watermark persistence. Second, the residual reconstruction analysis in Appendix B.3 does not convincingly explain why the proposed watermark survives distillation. In particular, RoMA+ achieves a relatively low reconstruction error (0.739 versus 0.654 for the proposed method) but completely loses its watermark after KD. This discrepancy raises questions about the relationship between residual predictability and actual watermark persistence.

  2. Under the trigger-aware removal setting (#22), the proposed method achieves only a 2% detection rate, indicating that the watermark can be effectively removed when the adversary has access to the trigger data. While this setting falls outside the stated threat model, the lack of stronger adaptive evaluations (such as trigger identification and removal methods used in backdoor attacks) limits the evidence for robustness against adversaries actively attempting to evade ownership verification.

  3. The proposed watermarking strategy deliberately constructs trigger phrases that are semantically unrelated to their target images, potentially introducing an exploitable weakness. An adversary could employ a vision-language model to assess the semantic consistency between input prompts and generated images, flagging outputs that deviate substantially from the requested content. The lack of an evaluation of such semantic-consistency-based filtering leaves the claimed robustness against output-level watermark evasion insufficiently substantiated.

  4. The introduction contains excessive technical details that obscure the paper's high-level motivation, key challenges, and main contributions. The authors should streamline this section to emphasize the problem setting, limitations of existing approaches, and intuition behind the proposed method, while deferring detailed technical discussions to Section 3 or subsequent methodology sections.

Questions for authors

None

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

This paper investigates how to embed hard-to-remove model watermarks into text-to-image models. The authors bind natural language trigger words to unrelated target objects and introduce a loss term that causes the watermark model to behave differently from the original model when presented with the trigger input. Experiments show that, even after various fine-tuning, compression, and distillation operations, this method retains the trigger behavior more effectively than existing methods while largely preserving normal generation capabilities.

Critical strengths
  • The problem formulation is practically relevant. The paper considers black-box verification of models accessible only through text prompts and incorporates input processing, model modifications, and output processing into the threat model. The use of semantic objectives for verification also helps resist output operations that compromise pixel-level watermarks.

  • The experiments are comprehensive, and the reported improvements are significant. Tables 3–4 include 21 types of model modifications, involving long-term fine-tuning, distillation, quantization, pruning, and module replacement. The method maintains a trigger detection rate of at least 90% in 19 of these scenarios; the distillation experiments, which initialize students from clean models, are also valuable for reference.

  • The baseline and supplementary experiments are well-designed. The authors added regularization terms to WatermarkDM and RoMA to preserve normal behavior, thereby reducing some comparison biases. The appendix also includes prompt rewriting, output regeneration, multi-objective injection, and dynamic training analysis, making the evaluation more comprehensive.

Critical weaknesses
  • The claim of a low false positive rate lacks sufficient statistical support. The abstract highlights a TPR@FPR close to 100%, but most experiments tested only 100 negative samples. Although Appendix A.1 supplements the analysis with a test of 10,000 images under the single target of Cubone, even when calculated as independent samples, the one-sided 95% upper bound on the false positive rate corresponding to zero false positives remains approximately (3×10^{-4}). Furthermore, false positives under generic prompts are not equivalent to the misattribution errors of non-derived models when faced with the same trigger query. The existing results are insufficient to support such strong claims regarding the reliability of ownership verification.

  • The mechanism and contribution bounds of the new loss are not sufficiently clear. Previous work, such as RoMa, has already addressed watermarking in models resistant to fine-tuning; the main methodological contribution of this paper is the negative separation term in Equation (2). Therefore, it is necessary to distinguish whether the benefit stems from stronger initial trigger separation or from an actual change in the sensitivity of subsequent modifications. The gradient correlation observations in Appendix B.2 are not yet sufficient to provide such attribution.

  • The scope of the theoretical guarantees is significantly narrower than stated in the main text. Section 4.2 emphasizes the existence of a valid (lambda2) for bounded model modifications, but Appendix F actually requires (C0>0), (D0>0), and (rho<C0/D0). This indicates that bounded modifications alone are not sufficient. Furthermore, this conclusion holds only for linear models with polynomial features and under specific parameter perturbations; it cannot be directly generalized to provide guarantees for practical fine-tuning, pruning, or cross-model distillation. The main text should explicitly state these limitations.

  • Conclusions regarding cross-scenario robustness still need to be refined. The coverage of the 21 weight modifications is sufficient, but the input rewriting and output regeneration tests in the appendix were primarily conducted on the unmodified (thetaw) and have not yet fully validated cross-layer combined processing. The other architectures in Table 5 were also tested only with modification #12, with detection rates of 0.22, 0.42, and 0.62 for aMUSEd, LlamaGen, and FLUX, respectively, indicating that the conclusion of near-complete persistence is not universal. Furthermore, both watermark training and attack training used only a single random seed, so the stability of the results remains unclear.

Questions for authors

The paper demonstrates solid experimental work and clearly outperforms the baseline across multiple settings—strengths that deserve recognition. However, model ownership verification requires credible error attribution controls, and the most prominent claim regarding low false positive rates has not been adequately supported. Additionally, there remains a lack of clear evidence as to whether the newly introduced loss introduces an independent persistence mechanism or primarily enhances the initial trigger separation.

If the authors can clarify the verification protocol, revise their statistical claims, and more accurately delineate the method’s benefits and theoretical guarantees, I will raise my evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.