acceptodds
Withdrawn from ICLR 2027

TrajRing: Taming Trajectory Drift for High-Fidelity Diffusion Watermarking

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Diffusion watermarking methods such as Tree-Ring enable provenance verification of diffusion-generated images by embedding watermark patterns in the initial latent noise. Despite their robustness to common image transformations, they inevitably compromise generation fidelity, causing watermarked outputs to deviate from their clean counterparts. Existing optimization-based approaches improve generation fidelity but incur substantial per-image computational overhead. To explore an efficient approach to preserving generation fidelity, we analyze paired clean and watermarked denoising trajectories. We identify watermark-induced trajectory drift: perturbations initially confined to the watermark region induce growing deviations in complementary frequencies during denoising, progressively driving the watermarked trajectory away from its clean counterpart. Motivated by this finding, we propose TrajRing, a training-free framework that suppresses this drift while preserving watermark detectability. At each denoising step, TrajRing performs trajectory-aware spectral projection, retaining watermark-bearing components while projecting complementary components onto a timestep-aligned clean reference trajectory. To complement this spectral constraint with spatial guidance, clean-reference query-key injection reuses queries and keys from the clean reference trajectory while retaining value representations from the watermarked branch. Across two benchmarks, TrajRing improves PSNR by up to 14.91 dB over Tree-Ring while maintaining average detection AUCs of at least 0.98 across clean and transformed images. Evaluations across multiple diffusion backbones further show consistent fidelity improvements. Without per-image optimization, TrajRing runs in 4.77 seconds per image, achieving 24 and 98 speedups over PT-Mark and ZoDiac, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Watermarking, Content Provenance, Diffusion Models, Trajectory Control, Image Fidelity

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UJiVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TrajRing, a training-free method for improving the fidelity of Tree-Ring watermarked images. The authors observe that perturbations introduced into the initial latent noise can propagate during denoising, causing the watermarked generation to deviate from its clean counterpart. To address this, they propose a combination of trajectory-aware spectral projection and clean-reference query-key injection to preserve the watermark while keeping the generated image close to the unwatermarked output. The paper reports improvements in paired fidelity while maintaining comparable watermark detection performance and avoiding per-image optimization.

Critical strengths

The paper does not require additional training or per-image optimization, making it computationally efficient.

Critical weaknesses
  • My main concern is why a watermarked generation needs to be highly similar to its clean counterpart in the first place (which is the main motivation of the paper). As long as the generated image is realistic, consistent with the prompt, and contains an imperceptible watermark, that is enough in most cases where the user has no idea what the clean image would look like. Building on this, I don't think the reported PSNR, SSIM, and LPIPS against paired clean generations are fair to Tree-Ring, since Tree-Ring produces a different but equally valid image. Additionally, given that the main benefit is preserving similarity to a particular clean generation, while watermark detection remains largely unchanged, I am not convinced that the paper establishes sufficiently important improvements over stronger fidelity-preserving baselines such as PT-Mark.

  • Several previous works have demonstrated targeted removal and forgery attacks against Tree-Ring. Since TrajRing retains the original Tree-Ring watermark construction and detector, it is important to understand whether these vulnerabilities are also inherited. The current evaluation does not establish robustness against such targeted attacks.

Questions for authors
  • Can you please provide a scenario in which the watermarked and clean images should look the same? This will raise the practical significance of the proposed method. More precisely, please provide your threat model.

  • How does TrajRing compare with simpler approaches? I am asking this because the novelty of the proposed method is limited. The observation that perturbations in the initial noise propagate and cause deviations during denoising seems expected given the nonlinear nature of diffusion models. Similarly, it combines spectral replacement with clean-reference attention guidance, both of which are conceptually straightforward and related to existing reference-guided generation techniques. Therefore, you need to establish why the proposed approach is preferable to simpler alternatives, such as one-time spectral correction or latent blending using the same clean reference.

Reviewer_bxg8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles the generation fidelity loss inherent to initial-noise diffusion watermarking methods such as Tree-Ring, where embedding watermark patterns distorts the final output relative to clean generation under matched settings.

Key contributions:

  1. It identifies and characterizes watermark-induced trajectory drift: initially localized watermark perturbations induce accumulating deviations in complementary frequency bands across denoising steps, progressively driving the watermarked latent trajectory away from its clean counterpart.
  2. It proposes TrajRing, a training-free, two-pass framework that suppresses this drift via two complementary mechanisms: trajectory-aware spectral projection that retains watermark-bearing frequency components while aligning complementary components to a timestep-aligned clean reference, and clean-reference query-key injection that guides spatial attention affinities while preserving watermarked feature values. The method requires no model fine-tuning or per-image optimization and is compatible with the original Tree-Ring inversion detector.

Validation:

Experiments on DiffusionDB and MS-COCO demonstrate 12.69–14.91 dB PSNR improvements over Tree-Ring while maintaining average detection AUCs of 0.98–0.99. Consistent fidelity gains are verified across four diffusion backbones. Running at 4.77 seconds per image, TrajRing achieves approximately 24× and 98× speedups over per-image optimization methods PT-Mark and ZoDiac, respectively. Component ablations and trajectory-level analyses confirm the complementary roles of spectral projection and attention guidance.

Critical strengths
  1. Mechanistic insight paired with an elegant, practical training-free solution The paper first identifies and characterizes watermark-induced trajectory drift, showing that initially localized watermark perturbations cause accumulating deviations in complementary frequency bands during denoising, which is the root cause of fidelity degradation in initial-noise diffusion watermarking. Building on this insight, TrajRing combines spectral projection and clean-reference attention injection to suppress drift without model fine-tuning or per-image optimization, while remaining fully compatible with the original Tree-Ring inversion detector. This is consequential because it delivers both a principled explanation of a known problem and a low-cost, easily deployable solution, rather than a heuristic or optimization-heavy workaround.
  2. Comprehensive, multi-faceted empirical evaluation with clear tradeoff characterization The evaluation is unusually thorough: it benchmarks against 10 competing methods across two datasets, verifies consistent gains across four diffusion backbones, reports both fidelity metrics and detection robustness under common and reconstruction-based attacks, measures computational efficiency, and includes component ablations, coverage sweeps, and trajectory-level diagnostics to validate the proposed mechanism and design choices. It also explicitly characterizes the fidelity-robustness tradeoff across design variants rather than only reporting best-case numbers. This breadth and rigor strengthen confidence in the results and make the work useful both as a reference point and as a basis for follow-up research.
Critical weaknesses
  1. Double-pass architecture doubles inference cost, with limited discussion of deployment tradeoffs The fidelity improvements of TrajRing fundamentally rely on an additional clean-reference diffusion pass, effectively doubling inference computation compared with single-pass methods like Tree-Ring. While the paper emphasizes speedups over per-image optimization baselines, it does not squarely address that the method doubles compute overhead relative to the baseline initial-noise watermarking scheme, nor does it explore lighter single-pass approximations. For real-world deployment where generation throughput is a primary concern, doubling inference cost substantially limits practical applicability, weakening the practical impact of the fidelity gains.

  2. Watermark robustness and capacity tradeoffs are insufficiently characterized First, detection AUC drops under certain transformations (e.g., rotation) relative to Tree-Ring, and the paper does not systematically evaluate core watermark metrics such as information capacity or robustness against strong adversarial attacks. Second, by only retaining watermark components within the protected region, the effective payload space of the watermark is constrained. The paper does not compare fidelity under matched watermark capacity across methods, making it unclear whether the fidelity gains are inherently due to the drift-suppression mechanism or simply due to reduced effective watermark payload. For a watermarking work, this is a material limitation, since the primary purpose of a watermark is verifiability, and fidelity improvements should not come at the cost of eroding core watermark functionality.

Questions for authors

TrajRing relies on an additional clean-reference diffusion pass, which effectively doubles inference computation compared with single-pass initial-noise watermarking baselines such as Tree-Ring. Could the authors clarify whether they have evaluated lighter single-pass approximations (e.g., injecting reference guidance at only a subset of denoising steps rather than a full second pass), and how they assess the practical applicability of the double-pass architecture in deployment scenarios where generation throughput is a primary concern?

The fidelity comparisons use a fixed watermark configuration, but the spectral projection design, which retains only watermark-region components, may constrain the effective watermark payload. Could the authors clarify whether fidelity improvements have been measured under matched effective watermark information capacity across methods? Additionally, have core watermark properties such as information capacity and robustness against strong adversarial attacks been evaluated, beyond common image transformations?

Reviewer_S16tOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the fidelity loss associated with initial-noise diffusion watermarking methods (e.g., Tree-Ring). The authors identify and formalize watermark-induced trajectory drift, where localized watermark perturbations in initial noise cause compounding deviations in complementary frequency bands and spatial attention structures during iterative denoising. To tackle this without costly per-image optimization or model retraining, the authors propose TrajRing, a training-free two-pass sampling framework. TrajRing runs a clean reference sampling pass to cache latent trajectories and self-attention queries/keys. During the second watermarked pass, it applies: (1) Trajectory-aware Spectral Projection to constrain non-watermarked latent frequencies to the clean reference, and (2) Clean-reference Query-Key Injection to preserve spatial layout while retaining watermarked value representations.

Critical strengths

a. Insightful Analysis and High-Fidelity Preservation: The identification of watermark-induced trajectory drift is a solid conceptual contribution. TrajRing achieves remarkable fidelity gains over Tree-Ring (PSNR improves by 12.69–14.91 dB and LPIPS drops to 0.02–0.03) without altering the underlying Tree-Ring inversion/verification protocol or compromising clean detection AUC. b. Training-Free & Efficiency Advantage over Optimization Baselines: Compared to per-image optimization baselines like PT-Mark (~114s) and ZoDiac (~466s), TrajRing delivers superior visual quality in just 4.77 seconds per image (a 24x–98× speedup) by substituting iterative optimization with cached clean-pass guidance.

Critical weaknesses

a. Insufficient Robustness Benchmarking and Metric Discrimination: Missing Standardized Benchmarks & Image Fidelity Metrics: Although basic transformations are evaluated, the paper omits community-wide robust watermarking benchmark suites such as WAVES (ICML'24), WaterBench (ICLR'25), MarkDiffusion (JMLR'26), WARP (ACM MM'26), or Stirmark 3.0/4.0. Additionally, Table 5 (Ablation Study) omits FID. Metric Saturation & Need for Strict FPR Thresholds: Standard AUC metrics approach saturation (near 100% for baseline methods under light attacks), making it difficult to reveal genuine performance bottlenecks. To substantiate practical usability and expose true robustness limits, the authors must report detection performance using stricter operating metrics such as TPR at 0.01% FPR. b. Phenomenological Analysis without Mechanistic Causality for Trajectory Drift and Component Design: Lack of Causal Drift Impact Analysis: Although the authors identify watermark-induced trajectory drift, the discussion remains largely phenomenological. The paper provides little rigorous analysis or evidence explaining how this drift quantitatively translates into performance trade-offs across specific metrics (e.g., how it selectively impacts AUC, robustness, versus PSNR). Ambiguous Component-wise Mechanics and Empirical Trial-and-Error: The specific roles of individual components (spectral projection vs. Q/K attention guidance) are treated empirically through basic ablation without a closed-loop causal proof mapping them to specific performance dimensions. Furthermore, the sharp drop in detection TPR caused by Value (VV) injection is merely reported as an empirical observation without deep physical or mathematical justification (e.g., how VV replacement washes away latent watermark traces during DDIM inversion).

Questions for authors

a. Benchmark Validation & Strict Metrics: Can you provide results under standard benchmark suites (e.g., WAVES, WaterBench) reporting TPR at 1% / 0.01% FPR and include FID in Table 5 to complement the fidelity ablation? b. Mechanistic Analysis of Trajectory Drift & Component Roles: Can you elaborate on the causal link between trajectory drift and specific performance metrics (AUC vs. robustness vs. PSNR)? Which specific components of TrajRing directly suppress drift versus those dedicated to safeguarding watermark detectability? c. Physical Justification of VV-Injection: Why does injecting Value (VV) severely hurt watermark detectability during DDIM inversion? Can you provide a physical or mathematical explanation for this phenomenon rather than just empirical trial-and-error?

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

Related papers

Discussion (0)

Sign in to comment.