acceptodds
Under review as a conference paper at ICLR 2027

RESON: Persistent Watermarking Across Heterogeneous Generative Lineages

Abstract

AI-generated images are increasingly edited and re-synthesized by downstream generative models, yet existing watermarking evaluations largely focus on single-stage transformations rather than provenance across successive generations. We introduce RESON, a generation-time latent watermark for generative lineage persistence that embeds a keyed, message-bearing carrier once in the source noise and enables detection directly from pixels, without downstream reinsertion, generator modification, or diffusion inversion. We evaluate the method across heterogeneous multi-hop lineages, alternative lineage orderings, conventional perturbations, deliberate generative removal attacks, and varying payload capacities. The watermark remains detectable through repeated cross-model re-synthesis while retaining recoverable source-associated information at increasing lineage depths. Controlled carrier and strength ablations show that persistence is driven by structured carrier content rather than injection energy alone, while generation-quality analysis indicates limited degradation in semantic fidelity. Together, these findings establish generative lineage persistence as a distinct watermarking objective and demonstrate the feasibility of insert-once, detect-later provenance across heterogeneous generative transformations. The source code is available at https://airesearchfolio.github.io/reson.

Primary area:
generative models
Keywords:
Diffusion Watermarking, Image Provenance, Multi-Hop Regeneration, Heterogeneous Generative Models, Latent Watermarking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_srKDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RESON, a watermarking method designed to preserve source information across multiple generations of an image, even when different generative models are involved. RESON embeds a keyed, message-bearing watermark into the initial diffusion noise using a combination of structured and white-noise carriers. A lightweight decoder then detects the watermark and recovers its payload without requiring diffusion inversion or modifications to downstream models.

Critical strengths

The problem setting is interesting and relevant to the audience. Basically, the paper goes beyond conventional single-step watermark robustness and investigates whether source information can survive multiple heterogeneous generative transformations. I find this setting meaningful, particularly since downstream models do not need to cooperate or reinsert the watermark.

RESON maintains strong detection performance across several regeneration steps and generally outperforms the evaluated baselines at later depths.

Critical weaknesses

However, the novelty seems limited compared to existing methods. The method is closely related to SERUM (ICLR 2026), which also modifies initial diffusion noise and uses a lightweight detector without inversion. RESON mainly changes the carrier construction and introduces a multi-hop evaluation setting. The improvement over it seems relatively small too; it would be better for the paper to have a stronger evidence that the proposed carrier provides a meaningful methodological advantage rather than an incremental modification.

Generative transformations are of limited diversity. The multi-hop experiments use the original prompt at every generation, a fixed image-to-image strength of 0.5, and a relatively small collection of model families. In practice, images may undergo prompt changes, inpainting, object replacement, or stronger regeneration. It is unclear whether the reported persistence would hold under these less controlled settings.

Also, moving to the multi-bit setting, the multi-bit performance is considerably weaker. While the paper emphasizes 92.5% detection at G4, this corresponds to the one-bit setting. For four-bit payloads, detection TPR drops to 54.0%, and for eight-bit payloads it drops to 36.4%. Exact-message recovery also falls to 78.4% and 34.2%, respectively. These results substantially limit practical source attribution. The paper should report end-to-end verification accuracy, requiring both successful detection and correct message recovery.

Another concern is incomplete security evaluation. The paper does not evaluate adaptive watermark removal, multi-image carrier estimation, or forgery attacks. Since the carrier is deterministically generated from a secret key, these are important threats, especially when many images share the same watermark. Robustness against watermark-unaware regeneration alone is insufficient to establish reliable source provenance.

Missing literature: Several works need discussion/comparison; e.g., MOLM (ICLR 2026), GROW (CVPR 2026), SPDMark (CVPR 2026), Semantic Watermarking Reinvented (ICCV 2025), SEAL (ICCV 2025), Guidance Watermarking (ICLR 2026).

Controlled experiments for fair analysis would benefit the paper; e.g., matched watermark strength, payload, and image quality.

Questions for authors

Please refer to the weaknesses above.

10 comments
AuthorsOpenReview ↗

W1 - Novelty and Comparison with SERUM

We appreciate the reviewer’s concern about the distinction from SERUM. Both methods modify the initial diffusion noise and use inversion-free detection. RESON combines a keyed, bit-conditioned structured–stochastic carrier, joint presence-and-payload decoding, and an insert-once, detect-later evaluation of source-payload persistence through heterogeneous multi-hop regeneration, without downstream reinsertion or detector adaptation. Section 2 (Related Work, “Diffusion watermarking”) clarifies the methodological distinction (highlighted in blue in the revised manuscript), while Section 4.1 (“Multi-hop persistence and matched baselines,” Table 1) reports the cross-method lineage comparison. The Introduction additionally distinguishes the headline one-bit detection result from multi-bit verification (highlighted in red).

Descriptive comparison with SERUM. Table 1 of the manuscript compares RESON and SERUM under the heterogeneous G0–G4 lineage protocol. The results are reproduced below to make the comparison explicit.

GenerationSERUM TPR (%)RESON TPR (%)SERUM AUCRESON AUC
G0100.00100.001.00001.0000
G1100.00100.000.99990.9996
G297.6798.300.99680.9980
G392.0093.100.99410.9967
G489.3392.500.99160.9963

At G4, the observed TPR is 92.50% for RESON and 89.33% for SERUM, with AUCs of 0.9963 and 0.9916. These are descriptive differences of 3.17 percentage points in TPR and 0.0047 in AUC; we do not infer statistical superiority without a paired analysis of these exact predictions. More importantly, the evaluation tests whether source-embedded information survives repeated regeneration by different model families, rather than only a single generation or a conventional image perturbation.

Carrier-mechanism evidence. At fixed injection strength α=0.155\alpha=0.155, we ablate the RESON carrier composition with separate detectors and validation-frozen operating thresholds. This experiment examines carrier structure within RESON; it is not a direct substitution of SERUM’s carrier.

Carrierβ\betaG0 AUCG0 TPR (%)G4 TPR (%)G4 bit accuracy (%)
White-only0.000.791419.01.764.7
RESON mixed0.400.994093.364.398.3
Structured-only1.000.999798.381.799.7

The white-only carrier falls to 1.7% G4 TPR, compared with 64.3% for the mixed carrier and 81.7% for the structured-only carrier. The large gap between white-only and structured configurations indicates that the structured component is important for persistence. The structured-only variant achieves the strongest detection in this ablation; the mixed carrier is the canonical configuration; this ablation alone does not establish an optimal fidelity--robustness trade-off. These results appear in Table 5 (Controlled carrier-mechanism ablation), Section 4, “Carrier Design and Operating-Point Analysis.” The ablation isolates how structured and stochastic carrier components affect persistence within RESON.

AuthorsOpenReview ↗

W2 - Diversity of Generative Transformations

We agree that a broader range of transformations would make the evaluation more informative. Alongside the existing heterogeneous SD2.1 → FLUX.1-dev → RealVisXL → DreamShaperXL → FLUX.1-dev lineage and alternative generator orders, we evaluated prompt-modified FLUX regeneration, localized SDXL inpainting, and regeneration-strength variation.

Prompt modifications. Using 100 matched clean/watermarked pairs per condition, the frozen detector produced:

FLUX prompt conditionTPR (%)Realized FPR (%)
Original prompt96.00.0
Style-modified prompt97.01.0
Semantic-modified prompt95.01.0

Localized generative editing. SDXL inpainting yielded:

Masked image regionTPR (%)Realized FPR (%)
10%99.00.0
25%96.02.0
40%87.02.0

Regeneration strength. In the separate regeneration-severity sweep, TPR@1% FPR was 98.8%, 98.6%, 98.0%, and 94.4% for strengths 0.2, 0.4, 0.6, and 0.8, respectively. This sweep uses condition-specific operating thresholds and should not be interpreted as a single frozen-threshold comparison.

These experiments directly test whether a downstream generator can alter the prompt or a localized image region while the source watermark remains detectable. Even with 40% inpainting, detection remains at 87.0% TPR, and with semantic prompt changes it remains at 95.0% TPR. They complement, rather than replace, the heterogeneous multi-hop evaluation. The new prompt-change and inpainting results appear in Section 4, “Robustness to perturbations and deliberate removal,” and Appendix A.4.3, “Prompt-Modified Regeneration and Localized Generative Editing” (Table 10; label tab:prompt_inpaint). The transformation-scope clarification appears in Section 5, “Discussion and Limitations.” These additions are marked in red in the revised manuscript. The regeneration-strength sweep appears in Appendix A.4.2 (the regeneration-severity table, label tab:regen_severity).

AuthorsOpenReview ↗

W3 - Multi-bit Payload and End-to-End Verification

We appreciate the reviewer’s request for a stricter operational measure of source verification. One-bit presence detection and multi-bit message verification measure different outcomes. The headline 92.50% G4 TPR refers to the one-bit configuration. To evaluate operational multi-bit verification, we additionally require both a positive presence decision and exact recovery of all payload bits, using the original validation-frozen detection thresholds and 1,000 watermarked plus 1,000 clean held-out images at each lineage depth.

Depth4-bit joint success (%)4-bit presence FPR (%)8-bit joint success (%)8-bit presence FPR (%)
G091.91.558.61.8
G178.21.137.91.6
G260.01.125.40.8
G347.80.918.51.1
G448.00.817.81.0

At G4, joint verification succeeds for 48.0% of four-bit and 17.8% of eight-bit watermarked images, compared with presence-detection rates of 54.0% and 36.4%, respectively. The table's FPR values concern presence detection on clean images, not message-specific false acceptance.

These end-to-end results are added in Section 4, “Payload Capacity and Persistent Information” (Table 6, joint verification), and Section 5, “Discussion and Limitations,” under “End-to-end verification and payload capacity.” The additions are marked in red. The joint criterion is more stringent than either detection or payload decoding alone and makes the capacity–robustness trade-off explicit: four-bit verification remains measurable at G4, while eight-bit verification is considerably weaker. This distinguishes evidence of persistent watermark presence from reliable recovery of a complete source identifier.

Reviewer_fD8wOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work aims to handle a novel problem of generative lineage persistence in image watermarking. Existing watermarking evaluations largely focus on single-stage transformations rather than provenance across successive generations. Consequently, the authors introduce a novel persistent-provenance problem formulation: The authors define the insert-once, detect-later setting in which a source watermark must survive a heterogeneous generative lineage without downstream reinsertion, generator modification, intermediate-state access, or detector adaptation. To handle this problem, the authors introduce RESON, a generation-time latent watermark for generative lineage persistence. It embeds a keyed message-bearing carrier once in the source noise and supports pixel-level detection, requiring no downstream reinsertion, generator modification, or diffusion inversion. The authors evaluate the proposed method under a lineage-centered protocol, which contains: heterogeneous G0–G4 lineages, alternative generator orders, multi-bit payloads, conventional and generative attacks, carrier/strength ablations, and generation fidelity. On the canonical SD2.1 lineage, RESON retains 92.50% TPR at 1% FPR at G4, while four-bit recovery remains 94.03% per bit. Controlled ablations further attribute persistence to carrier structure rather than injection energy alone.

Critical strengths
  1. This work proposes a meaningful and interesting setting of generative lineage persistence in image watermarking, distinguished from the existing standard one-step watermark test.
  2. The experimental evaluation is comprehensive and contains solid analyses.
Critical weaknesses
  1. [Lineage Depth]. The current evaluation covers lineages from G0 to G4. Since the watermark detectability may gradually degrade after repeated regeneration, it is unclear whether RESON can reliably preserve provenance information beyond G4. Could the authors provide some experimental results on this issue? If not available, please provide a detailed discussion.
  2. [Compound Transformations]. This paper evaluates multi-hop regeneration attacks (Table 1 and 2) and conventional perturbations (Table 3) like JPEG compression, Gaussian noise, rotation, etc. These attacks are mostly in separate settings. However, realistic generative images may involve multi-hop generation and conventional perturbations like cropping or resizing between successive regeneration steps. I am curious about whether RESON can still preserve the watermark detection and payload recovery performance under these compound transformations. Could the authors provide some experimental results on this issue? If not available, please provide some detailed discussion.
Questions for authors
  1. [Lineage Depth].
  2. [Compound Transformations].

Please refer to the details in the critical weaknesses part.

2 comments
AuthorsOpenReview ↗

W1 - Lineage Depth

We thank the reviewer for recognizing the importance of evaluating lineage persistence beyond four downstream generations. We extended the heterogeneous regeneration chain to G8 (G5: RealVisXL; G6: DreamShaperXL; G7: FLUX.1-dev; G8: RealVisXL), evaluating 300 matched clean/watermarked image pairs at each new stage (G5–G8). This supplementary evaluation uses a matched β=0.4\beta=0.4 carrier/detector configuration, with the detector and its validation-selected decision threshold (0.9377391338348389) held fixed across all stages. There is no detector retraining or stage-specific recalibration. G5–G8 outputs are generated at 1024 × 1024 and resized to 512 × 512 using Lanczos interpolation before SD2.1 VAE-based detection.

GenerationDetection TPRFPRAUCPayload bit accuracy
G0 (supplementary matched setting)93.33%0.67%0.994099.67%
G556.67%0.00%0.949998.00%
G658.67%0.00%0.941698.00%
G750.00%0.00%0.932297.00%
G846.67%0.00%0.934396.00%

These results reveal an important distinction between persistence of watermark-related information and reliable fixed-threshold verification. At G8, ROC-AUC remains 0.9343, indicating measurable separation between watermarked and clean outputs, whereas fixed-threshold TPR decreases to 46.67%. Thus, the experiment supports residual watermark persistence beyond G4 but does not establish uniformly reliable provenance decisions at arbitrary lineage depths. The modest G5-to-G6 increase also shows that degradation is not strictly monotonic. Payload-bit accuracy of 96.00% at G8 is conditional on known-watermarked images and must not be interpreted as joint detection-and-message verification.

Manuscript changes (blue). In Appendix A.1, subsection “Extended Lineage Depth and Compound Transformations” (app:fd8w_supplementary), we added specifically the paragraph “Extended lineage (G5–G8)” and the supplementary lineage-depth table (tab:fd8w_depth). We also updated Section 5, “Discussion and Limitations,” under “Limitations,” to discuss the longer-lineage findings and the limitations of fixed-threshold verification. These additions are marked in blue.

AuthorsOpenReview ↗

W2 - Compound Transformations

We thank the reviewer for this suggestion. To test a more realistic mix of generative and conventional edits, we applied JPEG compression (quality 50) or Crop80 once to the G1 image, then regenerated through G2, G3, and G4. Crop80 is a centered crop that keeps 80% of the width and height and is resized back to the original size with Lanczos interpolation. Each condition is compared against a regeneration control with no edit, using 300 matched clean/watermarked pairs per stage. All scenarios use the same supplementary matched β=0.4 detector and the same frozen threshold (0.9377). We report detection at G2, G3, and G4, along with the G4 false-positive rate, AUC, and payload-bit accuracy on watermarked images.

ScenarioG2 TPRG3 TPRG4 TPRG4 FPRG4 AUCG4 payload bit accuracy
Regeneration only76.00%65.33%64.00%0.67%0.957697.67%
JPEG5075.00%64.67%62.33%0.67%0.950197.67%
Crop8064.00%54.33%52.00%0.00%0.924697.00%

The compound evaluation shows that the effect depends on the intervening edit. At G4, JPEG50 reduces detection TPR by 1.67 percentage points relative to the matched regeneration-only control, whereas Crop80 reduces it by 12.00 points. The G4 AUC values remain 0.9501 and 0.9246, respectively, indicating residual discrimination even when fixed-threshold detection weakens. Payload-bit accuracy remains near 97% on watermarked samples, but this conditional measure is distinct from joint verification. In addition to this , we evaluated a mixed-transformation condition at G2 (73.67% TPR, 0.67% FPR, 0.9707 AUC, and 98.33% payload-bit accuracy). We report this as a G2 diagnostic rather than extrapolating unmeasured mixed G3/G4 results. These experiments do not include a separate resizing-only condition. Together, the findings show that a conventional edit preceding a multi-hop continuation can preserve substantial residual watermark information while also revealing greater sensitivity to cropping than to JPEG compression.

Manuscript changes (blue). We added the paragraph “Compound transformations” and the edit-before-G2 table (tab:fd8w_compound) to Appendix A.1, “Extended Lineage Depth and Compound Transformations” (app:fd8w_supplementary). Section 5, “Discussion and Limitations,” under “Limitations,” now summarizes the observed sensitivity to interleaved cropping. These additions are marked in blue.

Evaluation scope. These supplementary experiments use a different matched detector/threshold configuration from the canonical G0–G4 benchmark. In particular, the supplementary G4 regeneration-only TPR of 64.00% (near the 64.3% G4 TPR in the separate mixed-carrier ablation, Table 5) is not directly comparable to the 92.50% G4 TPR in the original matched-lineage Table 1; the two results use distinct detector/evaluation configurations. We retain the canonical benchmark unchanged and use these experiments to characterize the effect of additional lineage depth and compound transformations.

Reviewer_i7gHOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to address a practical problem of watermark persistence: "how to design and implement watermarking under various generated image modifications". The authors propose a multi-hop persistent watermark, RESON, which adds a secret-key-derived pattern to the initial noise used by Stable Diffusion. The intuition of a sustainable watermark is that a low-capacity, structured source signal can remain decodable through several heterogeneous regeneration channels with a source-trained decoder.

Critical strengths
  1. The threat model is realistic; the authors combine multiple evaluation setups.

  2. The performance seems promising. RESON achieves 92.5% detection after four regeneration steps under the canonical setup, outperforming the tested baselines.

  3. The authors further explore what contributes to the persistence. Despite reporting ambiguities, the results provide evidence that spatial structure improves persistence compared with white-noise carriers.

Critical weaknesses
  1. The research gap and methodological novelty are unclear. Related Work does not sufficiently distinguish the missing evaluation of multi-hop regeneration from the limitations of existing watermarking methods. The technical differences from SERUM are not clearly established; most components follow existing approaches, with carrier construction appearing to be the main incremental change.

  2. The evaluation protocol is too narrowly prescribed. Experiments primarily follow author-selected regeneration chains. There is no systematic attacker action space specifying available operations, knowledge, budget, and content-preservation constraints. Randomly sampled and adversarially selected sequences would provide stronger evidence of robustness.

  3. Table 3 shows detection falling to approximately 70–72% under rotation. The paper should explain this weakness and make the rotation severity clear.

  4. The reliable payload capacity seems to be quite small. Figure 6 shows weak eight-bit recovery, while Table 8 reports only 33% exact-message recovery for eight bits even before regeneration. Above-chance bit accuracy does not establish reliable identifier recovery or scalable attribution.

  5. Multiple important terms and metrics are insufficiently explained. Terms such as “RINSE-style reconstruction” lack immediate operational explanations. Carrier mixing is mathematically defined but poorly motivated. “Prompt fidelity” is associated with CLIP similarity, but its exact calculation and evaluation configuration are unclear.

  6. Minor points. Model names, datasets, and some parameters are provided, but exact checkpoint identifiers, release links, prompt subsets, and seeds are insufficiently clarified.

Questions for authors
  1. Do you identify RSSON as a more complete evaluation or a novel watermark method? What method-level lessons should guide the design of persistent watermarking?

  2. The evaluation primarily follows author-defined pipelines. Why is there no explicit attacker action set, with defined capabilities, budgets, and content-preservation constraints? Why are regeneration sequences not randomly sampled from a defined action set or selected to maximize watermark removal? Why not consider a setting where the adversary can randomly perform both image regeneration and image perturbations?

  3. Figure 3 explores the trade-off between watermark robustness and image fidelity across two hyperparameters. The strongest robustness appears to come at the greatest fidelity cost. How noticeable is this degradation in practice?

  4. Table 3 shows substantial performance degradation under rotation. Why is RESON sensitive to rotation, and what rotation angles did you evaluate?

  5. Figure 6 shows poor performance for eight-bit payloads. What payload size can RESON reliably recover, and is that capacity sufficient for practical attribution?

7 comments

We thank the reviewer for the careful assessment and for recognizing the relevance of our threat model, the encouraging multi-hop results, and the role of carrier structure. We appreciate the opportunity to clarify our contributions and provide additional evidence. To make the revisions easier to identify, all manuscript changes addressing Reviewer i7gH are highlighted in purple, while revisions for the other reviewers retain their respective colors.

We address each weakness and question individually below.

W1 — Research gap, novelty, and distinction from SERUM

We agree that modifying diffusion initialization and using inversion-free detection are not, by themselves, new. RESON's contribution is the combination of a keyed, bit-conditioned structured–stochastic carrier, joint presence/payload decoding, and the insert-once, detect-later objective across heterogeneous regeneration without downstream reinsertion or detector adaptation. The evaluation asks whether source information remains recoverable after multiple independent generative transformations.

Matched G4 lineage comparisonRESONSERUM
Detection TPR92.50%89.33%
Detection AUC0.99630.9916

These are observed differences, not a claim of statistically established superiority. At fixed injection strength (α = 0.155), the RESON carrier ablation gives 1.7% / 64.3% / 81.7% G4 TPR for white-only / mixed / structured-only carriers, respectively. This shows that structure matters within RESON; it is not a direct substitution of SERUM's carrier. The mixed setting balances persistence with fidelity rather than maximizing TPR alone.

AuthorsOpenReview ↗

W2 - Prescribed regeneration chains and attacker action space

The canonical SD2.1 → FLUX.1-dev → RealVisXL → DreamShaperXL → FLUX.1-dev chain was selected to study successive changes of generator family under a reproducible protocol. We also tested alternative generator orders and additional editing conditions:

Additional conditionDetection TPR
FLUX regeneration: original / style-modified / semantic-modified prompt96% / 97% / 95%
SDXL inpainting: 10% / 25% / 40% mask99% / 96% / 87%
Regeneration strength 0.2 / 0.4 / 0.6 / 0.898.8% / 98.6% / 98.0% / 94.4%

Prompt and inpainting tests use 100 matched pairs per condition with the frozen detector. The strength sweep reports TPR@1% FPR with condition-specific thresholds. These tests broaden the evaluated transformations, but we agree that randomly sampled mixed action sequences or an adversarially optimized search with explicit quality and query budgets would be a different, stronger evaluation. Our claims concern watermark-unaware regeneration and the specified black-box transformations, not worst-case key-aware removal.

AuthorsOpenReview ↗

W3 - Rotation sensitivity

We thank the reviewer for highlighting the need to clarify rotation robustness. We agree that rotation is challenging for RESON because its spatially structured carrier is not inherently rotation-invariant.

Table 3 reports rotation TPRs of 70.30% and 72.47% for SD2.1 and SD2.0, respectively. In an additional evaluation involving 1,000 images, a 30° rotation resulted in 40.7% TPR using the frozen detector. This experiment follows a separate protocol and is therefore not directly comparable with Table 3. We have updated the Robustness to Perturbations subsection and Discussion and Limitations section to clarify these findings and discuss geometric sensitivity as a limitation of the current method.

Comments on the paper

1 comment
Area_Chair_44R9OpenReview ↗

Final feedback

Dear Reviewers,

As 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!

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.