acceptodds
Desk rejected at ICLR 2027

Beyond Single Watermarks: Scalable Adaptive Tree-Ring Watermark Coexistence for Sequential Editing

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Digital watermarks are essential for provenance tracking, ownership attribution, and misuse deterrence in AI-generated imagery. In realistic creative pipelines, images often undergo multiple editing rounds, where new watermarks may need to be embedded while earlier provenance signals remain detectable. Existing coexistence studies mainly focus on training-based encoder-decoder schemes, leaving training-free Tree-Ring-style watermarks underexplored despite their capacity, model compatibility, and deployment flexibility. In this paper, we study multi-watermark coexistence for continual image editing based on Tree-Ring-style watermarks. We introduce SeqEdit-BM, a multi-round editing benchmark for evaluating whether multiple watermarks can be repeatedly embedded, preserved, and independently detected across sequential edits. Using this benchmark, we show that Tree-Ring-style watermarks have strong intrinsic compatibility but suffer progressive attenuation during iterative diffusion reconstruction. We attribute this attenuation to trajectory mismatch between watermark injection and detection, and propose an adaptive watermarking method based on trajectory alignment to calibrate the watermark template before injection. We further formulate two allocation protocols for multi-watermark coexistence: single-slot multiplexing, where multiple marks share one spatial-spectral slot, and deterministic slot assignment, where marks are assigned to disjoint latent-frequency slots. Experimental results on verification and identification tasks demonstrate that our method substantially improves multi-watermark coexistence and robustness against sequential editing, sustaining a near-perfect detection rate after five consecutive rounds while preserving image quality with no perceptible artifacts. These results suggest that Tree-Ring-style watermarks offer a practical foundation for provenance preservation in continual image editing.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multi-watermark coexistence, sequential image editing, Tree-Ring watermarks

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_9cmKOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the coexistence of multiple watermarks under sequential image editing, where each editing round introduces a new watermark while earlier ones should remain detectable. It proposes an adaptive Tree-Ring-style method that combines image-dependent injection calibration with a key-derived, inversion-based detection reference, and evaluates different watermark allocation protocols.

Critical strengths

The paper addresses a practically relevant provenance problem that extends beyond conventional single-watermark detection. The proposed method achieves substantial improvements over the evaluated baselines.

Critical weaknesses
  1. In Table 1, the PSNR and SSIM after the fifth editing round are 23.584 and 0.850, respectively. These values are insufficient to support the claim of perceptual indistinguishability. Tree-Ring watermarking can itself introduce substantial generation artifacts, affect output diversity, and cause stylistic shifts. It remains unclear whether five rounds of watermark embedding further exacerbate these issues.
  2. The paper reports aggregate detection performance but does not provide a breakdown by the round in which each watermark was embedded and the number of subsequent edits it has undergone. In particular, the detection rates of early watermarks after successive editing rounds should be reported to establish how effectively the method preserves earlier provenance information.
  3. The ablation study does not isolate the contributions of injection calibration, multi-round watermark embedding, and the reference transformation used during detection. Relevant controlled experiments would include applying injection calibration with only a single embedding followed by multiple editing rounds, applying the proposed reference transformation directly to Tree-Ring, as well as other comparative experiments. Such experiments are needed to clarify how much each component contributes to the reported improvements.
  4. The comparisons focus primarily on the Tree-Ring family, with insufficient discussion of a broader range of watermarking methods designed for robustness against editing and adversarial attacks. One relevant example is Attack-Resistant Watermarking for AIGC Image Forensics via Diffusion-based Semantic Deflection (ICLR 2026). Discussing such methods and their applicability to the sequential editing setting would help contextualize the proposed approach
Questions for authors

See Weaknesses!

Reviewer_iRk4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript explores the multi-watermark coexistence for continual image editing based on Tree-Ring-style watermarks. Authors construct a dataset named as SeqEdit-BM, which is a multi-round editing benchmark for evaluating whether multiple watermarks can be repeatedly embedded, preserved, and independently detected across sequential image edits. Then, applying SeqEdit-BM benchmark, authors show that Tree-Ring-style watermarks suffer progressive attenuation during iterative diffusion reconstruction due to the trajectory mismatch between watermark injection and detection, and they propose an adaptive watermarking method based on trajectory alignment to calibrate the watermark template before injection. Authors further formulate two allocation protocols for multi-watermark coexistence, i.e., single-slot multiplexing and deterministic slot assignment. For single-slot multiplexing, multiple watermarks share one spatial-spectral slot, while for deterministic slot assignment, keys are mapped to a predefined partition of non-overlapping latent-frequency slots with collisions allowed under blind writing. Authors conduct extensive experiments and demonstrate that the proposed method substantially improves multi-watermark coexistence and robustness against sequential editing, maintaining strong detection performance after five consecutive rounds while preserving image quality with no perceptible artifacts.

Critical strengths
  1. This manuscript accurately identifies the practical and meaningful generative image application scenario about multiple editing rounds, and correctively points out the demand of each round’s participated editor to embed corresponding watermark and keep the watermark consistently detectable after the following editing rounds. This reflects the authors’ thorough consideration within the field of generative image watermarking, and could also further inspire more related future researches about this topic.

  2. This manuscript proposes a simple yet effective strategy for enhancing Tree-Ring-style watermarks, which effectively enhances the maintenance effect of multiple watermarks that are embedded through the multi-round image editing process, obtaining the accurate detection results.

  3. The authors conduct extensive workloads in this manuscript, executing a wide range of experiments, and providing a relatively comprehensive evaluation of their proposed method, which can effectively demonstrate the validity of the proposed method through experimental results.

Critical weaknesses

Considering the above-mentioned strengths of this manuscript, to encourage this manuscript’s accurate identification and exploration about the realistic demand to keep the embedded watermark in each edit round consistently detectable, and the promising experimental results achieved by the proposed method, a weak acceptance rating can be given. However, the following weaknesses must be well addressed to further polish the manuscript.

  1. In Sec. 2 of this manuscript, some previous works are inaccurately categorized. For example, Tang & Aoki, 1997; Wang et al., 1998; Al-Haj, 2007 are expressed as training-based watermarking methods, however, these works apply traditional watermark embedding and extraction technologies, and do not relying on model training. Authors should carefully examine the content of Sec. 2 and ensure a totally correct previous works’ introduction and categorization.

  2. The proposed adaptive watermark with a trajectory alignment strategy is an important component in the design of the proposed method. However, even if the calibrated pattern is reinserted into the original spectrum, the actually conducted image editing may still cause the trajectory alignment degradation. The authors lack sufficiently in-depth analysis and theoretical justification to support the rationality and effectiveness of the proposed adaptive watermark. Similarity, the explanation about the reason why prompt-free inversion is additionally exerted on the obtained watermark-bearing latent in Eq. (6) is also inadequate. Furthermore, specific ablation studies evaluating the individual contribution of the two above-mentioned components are also missing. The above-mentioned sufficiently in-depth analysis and specific ablation studies should all be supplemented.

  3. In the experiments, PSNR and SSIM are applied to measure image quality. However, as a work related to generative image watermarking, visual quality metrics especially applied for generative images, such as CLIP-T and PickScore, are not included, causing a limited evaluation about the influence of the proposed method’s introduced watermark. Authors should add the evaluation of visual quality metrics especially applied for generative images when conducting the comparison.

  4. The expression explaining the designed method’s watermarking embedding and extraction is not clear enough. The meaning of key notations is separated too far from the related sentences, increasing the reading inconvenience. Besides, the related sentences are also not fully well expressed and organized. Authors should further enhance the related expression and organization when explaining the designed method’s details, and clearly introduce the meaning of key notations near the corresponding sentences that they are actually used.

  5. Tab. 5 is mentioned in the main manuscript, while this table is placed in the appendix, reflecting that this manuscript’s arrangement of some involved important contents is not planned carefully enough. The authors could consider combining Tab.1 and Tab. 2, as well as their related sentences, to leave more space in the main manuscript to accommodate Tab. 5 and its corresponding analysis, facilitating the comprehensive exhibition of important evaluation results.

Questions for authors
  1. What is the theoretical basis of the proposed adaptive watermark and applying the additional inversion on the obtained watermark-bearing latent in the designed method?

  2. What is the individual contribution of the proposed adaptive watermark and applying the additional inversion on the obtained watermark-bearing latent on the performance of the proposed method?

  3. Apart from the achieved PSNR and SSIM values, how does the proposed method perform on visual quality metrics especially applied for generative images?

  4. Based on keeping involved watermarks consistently detectable after all editing rounds, is there any way to also indicate the order of watermark embedding? By doing so, the method can simultaneously reveal the chronological order of different involved editors conducting their corresponding operations, thus providing more detailed information for efficient provenance tracking and ownership attribution.

Reviewer_iHCiOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the problem of enabling the detection and attribution of prior watermarks while adding new watermarks during sequential image editing. The authors propose SeqEdit-BM alongside two slot allocation protocols, combining adaptive injection and inversion-based reference construction with a Wind-style detection scheme. The primary evaluation is conducted across five editing rounds using FLUX and FireRed, with cross-model experiments including InstructPix2Pix provided in the appendix.

Critical strengths
  1. The paper tackles a timely and practical problem. Setting the objective for multiple entities to maintain their respective provenance information as an image undergoes sequential edits is meaningful. The protocol distinguishing between blind/informed writing and the two slot allocation strategies is also useful.

  2. The paper reports substantial improvements over the direct baseline, Wind. In the 5th-round Clean setting on FLUX, the reported TPR improves from 0.439 to 0.841, and the All-Detection Rate increases from 0.011 to 0.561. Consistent improvements under corresponding conditions are also observed on FireRed.

Critical weaknesses
  1. The evidence supporting independent attribution and robust coexistence is insufficient. The All-Detection Rate does not penalize false positives of absent keys, and controlling for false attribution on images embedded with different keys is left to future work. Therefore, a high detection rate within a single slot alone is insufficient to confirm key-wise coexistence. Furthermore, the 1% FPR in TPR@1%FPR is a nominal value calibrated on a calibration set, while empirical test FPR figures across specific conditions are difficult to find. Under FLUX with blind writing and deterministic slot assignment, the 5th-round Blur TPR is 0.216 and the All-Detection Rate drops to 0; claims regarding robustness should accurately reflect these limitations.

  2. The methodological contributions and trade-offs among detection performance, visual quality, and computational cost are not adequately validated. There are no ablation studies disentangling the individual effects of adaptive injection, reference inversion, and Top3-Mean, making it difficult to discern the exact drivers of the reported improvements. While the paper acknowledges distributional shifts caused by discarding the imaginary component following an inverse Fourier transform, comparisons against subsequent methods that address this issue via Hermitian symmetry preservation are lacking. Visual quality evaluation relies primarily on PSNR/SSIM and limited qualitative examples, lacking grounded assessments of perceptual differences and output distribution shifts, as well as measured wall-clock overhead. Finally, concluding that coexistence outperforms training-based methods without controlled comparisons under identical settings extends beyond the empirical evidence provided.

Questions for authors
  1. Can you report the score distributions and empirical FPRs for absent keys on images embedded with different keys? Please include absent keys sharing the same slot as well as empirical test FPRs across each condition. In addition, please clarify why the proposed method achieves improved performance within a single slot.

  2. Can you provide an ablation study isolating the individual contributions of adaptive injection, reference inversion, and Top3-Mean? Please also provide separate detection results for earlier-round keys versus the most recent keys evaluated on the identical final image.

  3. Given the frequency-pattern distortions arising from discarding the imaginary component in Tree-Ring-style approaches, can you compare against HSTR and HSQR—which preserve Hermitian symmetry—or justify their omission? It would be helpful to examine the impact of the underlying pattern-preservation strategy on detection performance and image quality.

  4. What is the reference image used for calculating PSNR and SSIM? Can you provide LPIPS to evaluate perceptual differences against corresponding unwatermarked edits, as well as FID to evaluate distributional shifts in the generated outputs? Please also specify the reference targets for each metric and the sample size used for FID.

  5. How are model-specific image conditioning inputs handled during detection? Please also report the measured additional runtime per round incurred by adaptive injection.

  6. Please clarify the configuration of the averaging attack, including the number of images, key-sharing setups, the exact procedures for removal and forgery, and the precise positive/negative class definitions used for computing AUC.

At present, given the empirical and analytical gaps noted above, it is difficult to conclude that the core claims are sufficiently substantiated. However, should the authors convincingly address these concerns in their rebuttal with concrete supporting evidence, I am open to reconsidering my evaluation toward supporting acceptance.

Comments on the paper

1 comment
Area_Chair_mNBaOpenReview ↗

Participate in the discussion period

Dear reviewers,

The authors have replied to your comments and the discussion is close to the end. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you!

Best, Your AC

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

Related papers

Discussion (0)

Sign in to comment.