acceptodds
Under review as a conference paper at ICLR 2027

DiffStamp: Differentiably Rendered Proactive Watermark for Visual Disclosure

Abstract

The rapid proliferation of high-fidelity AI-generated content (AIGC) has made in-content provenance especially important for public disclosure of realistic synthetic media, and potentially also for creator-side intellectual-property (IP) claims. We propose DiffStamp, a proactive visible watermarking framework for addressing them. By integrating a differentiable renderer into diffusion models' inference, DiffStamp jointly optimizes stamp placement and image content through four stages: locating, blending, filtering, and embossing. Experiments on FLUX.1-dev, Qwen-Image-Edit, and HunyuanVideo show model-agnostic applicability, modest overhead, and persistence against automated removal and regeneration attacks. A 108-participant visual-search study further shows that it harmonizes with image content; its stamps are less noticeable than naïve overlays yet discoverable under deliberate search. We further present a concept-aware attribution extension in which designated concepts are detected from the prompt, their cross-attention maps are aggregated, and concept-specific stamps are localized near the corresponding visual regions, illustrating a path from global disclosure to localized content/IP attribution without retraining the diffusion model.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
watermark, diffusion-generated images and videos, AI-content disclosure

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_E4jGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores a method for embedding visible watermarks during diffusion model generation. The approach is primarily based on a differentiable renderer and consists of four stages (Locating, Blending, Filtering, and Embossing) that select appropriate positions for watermark embedding while minimizing the impact on image aesthetics. The proposed method is adaptable to various image and video generation models and achieves better perceptual quality.

Critical strengths
  1. The problem addressed in this paper is highly important: how to add visible watermarks to images while preserving their visual quality.

  2. The idea of using a differentiable renderer to embed visible watermarks is somewhat novel.

Critical weaknesses
  1. The practical utility of the proposed method remains questionable. First, the method requires intervention in the diffusion model's generation process and incurs non-negligible computational overhead. The authors need to clarify what specific advantages the proposed method offers compared to post-processing watermarking algorithms, and whether these advantages are sufficient to offset the inconvenience of embedding the watermarking process into the generation pipeline. Additionally, the algorithmic pipeline designed by the authors lacks a mechanism for seamlessly integrating watermarks with the semantic content of the image, which leads readers to question whether it is even necessary to embed the watermarking process into the diffusion model's generation pipeline.

  2. There is a clear lack of necessary experiments. The community already has numerous visible watermarking methods. Examples include DVW: Diffusion Visible Watermark (ACM MM 2025), Beyond Invisibility: Learning Robust Visible Watermarks for Stronger Copyright Protection (UAI 2025), and Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal (ECCV 2022). Yet the authors compare against only one baseline (Ran et al., 2025) in the main paper (Table 2), which is clearly insufficient. In addition, the ablation study provides only qualitative results in the appendix, with no quantitative results at all. Furthermore, many important experimental results that belong in the main paper are relegated to the appendix. The authors should condense unnecessary sections (e.g., the full-page Figure 2 and the overly detailed preliminaries) and allocate the freed space to these critical experiments.

Questions for authors
  1. Compared to post-processing watermarking algorithms, what significant advantages does the proposed method offer?
  2. Why does the main paper compare against only one baseline? Please explain the rationale for omitting other baselines.
  3. Why is the ablation study entirely qualitative?
4 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer E4jG [1/2]

Thank you for emphasizing the importance of visible watermarking and recognizing the novelty of using a differentiable renderer.

[W1] “… the method requires intervention in the diffusion model’s generation process and incurs non-negligible computational overhead. The authors need to clarify what specific advantages the proposed method offers compared to post-processing watermarking algorithms … the algorithmic pipeline … lacks a mechanism for seamlessly integrating watermarks with the semantic content of the image …”

[Q1] “Compared to post-processing watermarking algorithms, what significant advantages does the proposed method offer?”

Response. Three advantages post-processing cannot offer, at +3.9–14.2% runtime and with no training (Table 4, p. 9, lines 468–478):

vs. Ran et al. (ACM MM 2025)Post-processingDiffStamp
Semantic integrationPicks where to paste on a finished imageHarmonizes nearby content during denoising (Sec. 4.2)
NIQE / ILNIQE / BRISQUE ↓4.81 / 33.8 / 30.24.63 / 30.5 / 26.2
Aesthetic / less-noticeable preference46.1% / 27.9% (1 mark)53.9% / 72.1% (6 marks)
Concept attribution✗✓ 69%/68% vs. 50%/33% chance (K=2/3)

(Table 2, p. 8, lines 391–399; Fig. 4, p. 9, lines 432–439.)

We respectfully disagree that a semantic-integration mechanism is missing. Blending (Sec. 4.2, p. 5, lines 255–269) injects the stamp into the predicted clean image so that the remaining denoising steps harmonize the surrounding content with it. Changing content around a mark, and reading the generator’s cross-attention for attribution, are possible only inside generation (p. 1, lines 047–053). The cost is small, e.g., 23.3 s → 25.8 s per FLUX image.

AuthorsOpenReview ↗

Rebuttal for Reviewer E4jG [2/2]

[W2a] “There is a clear lack of necessary experiments. The community already has numerous visible watermarking methods. Examples include DVW: Diffusion Visible Watermark (ACM MM 2025), Beyond Invisibility: Learning Robust Visible Watermarks for Stronger Copyright Protection (UAI 2025), and Watermark Vaccine: Adversarial Attacks to Prevent Watermark Removal (ECCV 2022). Yet the authors compare against only one baseline (Ran et al., 2025) in the main paper (Table 2), which is clearly insufficient.”

[Q2] “Why does the main paper compare against only one baseline? Please explain the rationale for omitting other baselines.”

Response. Of the three, only DVW produces a comparable output, and we now compare against it; the other two address different problems.

DVW (ACM MM 2025). No code is released, so we re-implemented it on FLUX.1-dev from the paper (same 200 prompts, logos, and six marks per image; new Table A9, App. C.4):

NIQE↓ILNIQE↓BRISQUE↓Decides placementAttribution
DVW4.5831.128.6✗ (user-given mask)✗
Ran et al.4.8133.830.2✓✗
DiffStamp4.6330.526.2✓✓

DiffStamp is best on ILNIQE and BRISQUE and on par on NIQE. More importantly, DVW only fuses a mark into a mask the user must supply; where a mark should go, which is our central contribution, is outside its scope.

Watermark Vaccine (ECCV 2022) is not a watermarking method. It adds an imperceptible adversarial perturbation to an already-watermarked image so that removal networks fail; it neither designs nor places a mark, so it has no output to compare in Table 2. It is complementary and can be applied on top of any visible mark, including ours. On its own axis, DiffStamp already withstands all five blind removal networks we tested, with no vaccine (Table A10).

Beyond Invisibility (UAI 2025) learns the watermark pattern itself for removal robustness and overlays it on existing images. Disclosure requires a prescribed emblem (a logo or standardized mark), which a learned pattern cannot carry.

Ran et al. was chosen because it is the closest prior work: an aesthetic method that decides where a given logo goes.

[W2b] “In addition, the ablation study provides only qualitative results in the appendix, with no quantitative results at all. Furthermore, many important experimental results that belong in the main paper are relegated to the appendix. The authors should condense unnecessary sections (e.g., the full-page Figure 2 and the overly detailed preliminaries) and allocate the freed space to these critical experiments.”

[Q3] “Why is the ablation study entirely qualitative?”

Response. It no longer is. The revision adds a quantitative per-stage ablation (Table A2; one stage removed at a time, 200 images each). Detection = a GPT-5.2 judge finds ≥1 mark, before / after the strongest regeneration attack; stamp-free reference 0.43 / 0.49.

NIQE↓BRISQUE↓Det. (0)Det. (180)
Full4.6326.20.9500.870
w/o Locating6.6946.40.9950.935
w/o Blending4.6226.40.9550.765
w/o Filtering5.0130.90.9750.825
w/o Embossing5.2935.10.5050.535

Each stage has a distinct, measurable role. Locating → quality: without it, stamps sit conspicuously on the content and quality collapses. Blending → robustness: +10.5 points (p=0.009) at no quality cost. Filtering → both. Embossing → visibility: without it, detection falls to the stamp-free level. The main paper now also carries the new quantitative robustness evidence: a stamp-free control with significance tests and five blind watermark removers (Table 3, Sec. 5.5).

Thank you for your detailed response. I think most of my questions have been satisfactorily addressed, so I’m willing to raise my score to 3 (Weak Accept).

Thank you for raising your score. Your review made the paper better: the DVW comparison (Table A9) clarified our scope, and the quantitative ablation (Table A2) now shows what each stage contributes.

Reviewer_ZX4XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DiffStamp inserts visible stamps (logos) into images and videos during diffusion sampling. The target use is manifest disclosure of AI-generated content, motivated by EU AI Act Art. 50; an extension adds prompt-triggered concept/IP attribution.

Pipeline.

  • Locating. For 5 early steps, stamp position, rotation and scale are optimized with a differentiable bitmap renderer to maximize grayscale cosine similarity with the predicted clean image, under boundary, repulsion and center penalties. For attribution, placement is gated by the concept token's cross-attention map.
  • Blending. At one step, a darkening residual inside the stamp mask is injected into the x0x_0-prediction through a LookingGlass-style latent correction, so that subsequent denoising harmonizes the content.
  • Filtering. The darkest candidate locations are kept.
  • Embossing. A luminance relief is applied at the final step.

Evaluation covers FLUX.1-dev, Qwen-Image-Edit and HunyuanVideo:

  • A visual-search study with 108 participants. DiffStamp stamps are found later than naive overlays, but 53–80% are found within 20 s.
  • No-reference IQA (NIQE, ILNIQE, BRISQUE) and a pairwise preference study (N=84) against a re-implementation of Ran et al. (2025).
  • Attribution on 165 multi-concept scenes: 69% / 68% correct concept assignment for K=2 / 3, vs 50% / 33% chance.
  • A regeneration attack scored by a GPT-5.2 judge.
  • Runtime overhead of 4–15%.
Critical strengths
  • Coupling a visible mark to the generation process is model-agnostic. Placement is optimized on x0x_0-predictions with a differentiable renderer, followed by one-step content harmonization.
    • It is demonstrated on a text-to-image, an image-editing and a video model, with modest runtime overhead (Table 4).
    • The within-subject visual-search study (108 participants; cumulative detection over time, Fig. 3) is a sensible way to quantify the noticeability trade-off.
    • The paper is clearly written and candid about limitations, including manual inpainting removal.
Critical weaknesses

1. Whether the stamps work as disclosure is not tested, and the design works against it.

  • The pipeline is built to make stamps blend in. Placement matches local content, filtering keeps dark backgrounds, and pop-out is deliberately suppressed (Sec. 4.1–4.3, 5.3). Visibility comes mainly from the final, post-hoc embossing (Sec. 4.4).
  • This conflicts with the regime the paper invokes (Art. 50(4)).
    • The EU Code of Practice asks that labels ensure "immediate recognition … without requiring … sustained attention" and "remain visible against any background".
    • DiffStamp aims to be "less noticeable … yet discoverable under deliberate search".
  • The visual-search study measures instructed search ("find and click the six hidden stamps", App. C.7). It does not test whether uninstructed viewers notice the mark or understand that the content is AI-generated. The preference study even rewards "the less noticeable watermark".
  • No disclosure mark is shown. App. A.1 calls for "a unified provenance emblem or platform-level disclosure mark", but every example uses brand logos (Nike, Disney, PayPal…).

2. The robustness, aesthetics and attribution evidence is weak.

  • Robustness (Table 3). The GPT-5.2 judge is told "there were exactly 6 embedded marks" and counts degraded remnants as a success. There is no stamp-free control, and the differences between methods are not significant at n=200.
  • Removal.
    • The abstract claims "persistence against automated removal", but learned removers are dismissed as needing "per-image localization" (Sec. 5.5). That is false for the cited WDNet, DENet and PatchWiper, and the uncited SplitNet and SLBR are blind too.
    • The only attack tested at scale, regeneration (Zhao et al.), targets invisible watermarks and is mild here: about 78% of the signal power remains at noise_step 180.
  • Aesthetics (Table 2). The 53.9% preference is not significant (binomial p≈0.16). The baseline is described as one watermark per image, although Fig. A7 shows about six.
  • Attribution. Concept assignment is 68–69% correct and is triggered by prompt keywords rather than detection. Its fallback fired on 0 of 396 maps, even with the concept absent (App. C.5).
  • Ablations are qualitative, on a single image (Fig. A1).
Questions for authors
  1. Address weaknesses.
  2. Show DiffStamp, a naive overlay and a standard visible AI label without mentioning marks. Afterwards, ask whether the content was labeled or AI-generated, and report noticing and comprehension rates.
  3. Run the GPT-5.2 judge on stamp-free images with the identical prompt and report its false-positive rate.
  4. Please make an effort to include removers in your evaluation, u can either use automatic methods, vlm-guided or instruction based guided, or even fine tune of diffStamp outputs.
  5. How does the design goal ("less noticeable") square with the EU Code of Practice's requirement that labels ensure "immediate recognition"?
  6. Quantitative ablations. Measure locating, blending, filtering and embossing separately, on noticeability (search time) and robustness.

I am open to changing my score given proper evaluation of the method.

3 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer ZX4X [1/2]

Thank you for the constructive review and for being open to revising the score.

[W1, Q1, Q4] “Whether the stamps work as disclosure is not tested, and the design works against it. … The EU Code of Practice asks that labels ensure ‘immediate recognition’ … The visual-search study measures instructed search …” (W1) / “Show DiffStamp, a naive overlay and a standard visible AI label without mentioning marks … report noticing and comprehension rates.” (Q1) / “How does the design goal (‘less noticeable’) square with … ‘immediate recognition’?” (Q4)

Response. “Less noticeable” is one operating point, not a fixed goal. Embossing strength γ and stamp count N_final tune how immediately the mark is seen (Fig. A2, p. 16, lines 831–834), so a deployment requiring immediate recognition can raise both and use a standardized emblem (p. 1, lines 036–037). We evaluate the subtle end because creators will not adopt a mark that hurts visual appeal (p. 4, lines 196–200). As for the uninstructed study in Q1: spontaneous recognition presupposes that the mark is already standardized and viewers know its convention (e.g., Gemini’s AI-generated mark is understood only because people already know it), so a present-day study would measure familiarity with an established label rather than the mark itself (Limitations, p. 10, lines 491–493).

[Ethics] “No disclosure mark is shown. App. A.1 calls for ‘a unified provenance emblem or platform-level disclosure mark’, but every example uses brand logos (Nike, Disney, PayPal…).”

[Ethics] “Real brand trademarks are used as watermark stamps on unrelated generated images (Nike, Adidas, Disney, Bosch, VISA, PayPal…). The paper’s own broader-impact section acknowledges the risk of false provenance claims. Generated images also include copyrighted characters (Batman, Captain America, Fortnite figures).”

Response. The emblem is an input bitmap, not part of DiffStamp: nothing in locating, blending, filtering or embossing depends on which mark is used. We drew stamps from public logo datasets only to test feasibility on diverse real-world emblem shapes, and the same pipeline served both disclosure and concept attribution, so a platform or rights holder can deploy whatever mark it chooses. To remove any trademark or IP concern, we regenerated the pipeline, qualitative and ablation figures, including the HunyuanVideo examples, with brand-free stamps, prompts and reference images (Figs. 1 and 2 and Figs. A1–A10). The pipeline figure, the ablation and controllability studies and the disclosure examples now carry a single “AI” disclosure badge (Figs. 1, A1–A3 and A5). Concept examples stamp generic icons that depict their objects, such as a paw on the teddy bear and a bag icon on the handbag. No prompt, edit instruction or reference image names or shows a brand or character (Table A8, App. C.3, p. 21). Broader Impacts now states that the logos used in our evaluations imply no association with their owners and that a deployment must stamp only marks it is authorized to use (App. E).

[W2a] “Robustness (Table 3). The GPT-5.2 judge is told ‘there were exactly 6 embedded marks’ and counts degraded remnants as a success. There is no stamp-free control, and the differences between methods are not significant at n=200.”

[Q2] “Run the GPT-5.2 judge on stamp-free images with the identical prompt and report its false-positive rate.”

Response. Done; the control strengthens the result. On stamp-free images (same prompts, seeds, judge prompt, and attack), the judge reports ≥1 “mark” in 43–50%, mostly text and signatures FLUX itself draws, so we now score its own count (≥3 of 6 found):

noise_stepNo stampNaïveNo BlendingDiffStamp
00.0900.8250.7900.815
600.1500.7550.7300.780
1200.1400.5700.6350.675
1800.1700.4900.4600.610

The differences are significant at n=200: at 180, DiffStamp beats Naïve by 12 points (p=0.02) and No Blending by 15 (p=0.004, Fisher’s exact). Removing the “exactly 6 marks” premise changes nothing: 0.850 for DiffStamp vs. 0.175 stamp-free. (Table 3, App. C.8.)

AuthorsOpenReview ↗

Rebuttal for Reviewer ZX4X [2/2]

[W2b] “The abstract claims ‘persistence against automated removal’, but learned removers are dismissed as needing ‘per-image localization’ (Sec. 5.5). That is false for the cited WDNet, DENet and PatchWiper, and the uncited SplitNet and SLBR are blind too. The only attack tested at scale, regeneration (Zhao et al.), targets invisible watermarks and is mild here: about 78% of the signal power remains at noise_step 180.”

[Q3] “Please make an effort to include removers in your evaluation. You can either use automatic methods, VLM-guided or instruction-based guided, or even fine-tune on DiffStamp outputs.”

Response. We ran all five blind removers the reviewer names (WDNet, DENet, PatchWiper, SplitNet, SLBR) with official code and checkpoints, fully automatically, on 200 images per method. None removes the stamps (Table A10):

Detection rateBeforeSLBRSplitNetWDNetDENetPatchWiper
Naïve0.9600.9350.9550.9800.9550.930
No Blending0.9500.9450.9550.9850.9750.935
DiffStamp0.9500.9450.9650.9800.9600.935

Tiling inputs at the removers’ 256-px training resolution, which favors the attacker, gives nearly identical results. Sec. 5.5 now reports this evaluation in place of the localization argument, and the abstract states the claim precisely (“blind watermark removers and regeneration attacks”). As for regeneration, noise_step 180 is 3× the attack’s default, and DiffStamp’s lead is largest exactly there.

[W2c] “Aesthetics (Table 2). The 53.9% preference is not significant (binomial p≈0.16). The baseline is described as one watermark per image, although Fig. A7 shows about six.”

Response. This result supports our aesthetics. Ran et al. (ACM MM 2025) is an expert-guided aesthetic method that places one mark; DiffStamp placed six (App. C.4, pp. 23–24, lines 1241–1244). Six times the marks should cost preference, yet DiffStamp won 53.9%. Under the reviewer’s binomial model (336 judgments), the 95% CI is 48–59%, excluding any aesthetic penalty beyond two points. Fig. A7 is illustrative, not a study stimulus (caption, p. 25, lines 1335–1337).

Table 2 (p. 8)Ran et al.DiffStamp
Marks per image16
Aesthetic preference46.1%53.9%
NIQE / ILNIQE / BRISQUE ↓4.81 / 33.8 / 30.24.63 / 30.5 / 26.2
AppliedAfter generationDuring sampling, adapting nearby content
Concept attribution✗✓ (Fig. 4)

[W2d] “Attribution. Concept assignment is 68–69% correct and is triggered by prompt keywords rather than detection. Its fallback fired on 0 of 396 maps, even with the concept absent (App. C.5).”

Response. Prompt-triggered by design: the operator knows which concept was requested; the task is where it appears, solved from the generator’s own cross-attention (p. 9, lines 445–446) well above chance (69%/68% vs. 50%/33% for K=2/3; Fig. 4). The fallback guards against degenerate maps, not absent concepts; 0/396 firings means none was degenerate (App. C.5, p. 24, lines 1266–1272). An unrendered concept was still requested, so its stamp remains truthful.

[W2e] “Ablations are qualitative, on a single image (Fig. A1).”

[Q5] “Quantitative ablations. Measure locating, blending, filtering and embossing separately, on noticeability (search time) and robustness.”

Response. Added (Table A2; one stage removed at a time, 200 images each). Detection = judge finds ≥1 mark, before / after the strongest attack; stamp-free reference 0.43 / 0.49.

NIQE↓BRISQUE↓Det. (0)Det. (180)
Full4.6326.20.9500.870
w/o Locating6.6946.40.9950.935
w/o Blending4.6226.40.9550.765
w/o Filtering5.0130.90.9750.825
w/o Embossing5.2935.10.5050.535

Each stage has a distinct role. Locating → quality: without it, stamps sit conspicuously on the content and quality collapses. Blending → robustness: +10.5 points (p=0.009) at no quality cost. Filtering → both. Embossing → visibility: without it, detection falls to the stamp-free level. Pre-attack detection serves as the noticeability measure; search time requires a human study per variant.

[Ethics] “Two user studies (108 and 84 participants) were run without IRB or equivalent review … There is no information on recruitment, informed consent or compensation.”

Response. Correct; we did not obtain IRB or equivalent review. Participants were recruited via the authors’ networks (many did not know the authors), read a written study description and consented, and ran the task on their own monitors. Most were unpaid; a few received KRW 5,000. These details are now in App. C.7 (pp. 24–25, lines 1295–1296 and 1343–1344).

Dear Reviewer ZX4X,

Thank you again for your detailed review. This is a gentle reminder that the discussion phase will end soon. For your convenience, we briefly summarize how our response addresses your comments:

  • Disclosure (W1, Q1, Q4). We clarify that the visual-search study measures discovery under instructed search, and we state this scope in Limitations.

  • Robustness (W2a, Q2). We add a stamp-free control for the judge, with significance tests (Table 3).

  • Removal (W2b, Q3). We evaluate five blind learned removers, none of which removes the stamps (Table A10).

  • Ablation (W2e, Q5). We add a quantitative per-stage ablation (Table A2).

  • Ethics. We add information on the user studies and replace brand logos with brand-free marks in the revised figures.

We would be happy to clarify any of these points further.

Sincerely,

Authors

Comments on the paper

4 comments
Area_Chair_V7dLOpenReview ↗

Review of DiffStamp

Paper summary

The paper introduces a visible watermarking method that aims to make watermark stamps less noticeable while preserving visual aesthetics and maintaining detectability. The proposed method embeds watermark stamps into generated images without modifying the generator's weights. It achieves this by optimizing various stamp properties, including location, shape, orientation, and luminance, and blending the stamps with the original image through a forward diffusion process. The method is evaluated on both text to image and text to video generation models. Experimental results demonstrate improvements in perceptual quality and human preference while maintaining reasonable computational overhead.

Critical Strengths

  • The proposed optimization framework has good motivation and is well-designed with clearly formulated loss functions.

  • This paper demonstrate improvement in perceptual quality and human preference.

  • The method does not require changing the parameter of the generator, allowing it to apply to different generative models.

Critical Weaknesses

  • The necessity of using attention maps at multiple selected layers and heads isn't well justified. A natural baseline would be to use an available segmentation model, such as SAM, to identify target concepts from the original generated image based on the text prompt and then optimize watermark placement within the resulting masks. This such comparison would help to show the benefits of the proposed attention guided approach.

  • Table 2 should include the perceptual quality scores of the original generated images without stamps. These reference values are important for assessing the actual quality degradation caused by the watermark and understanding the relative improvements achieved by the proposed method.

  • The paper does not evaluate whether all the embedded watermarks can be fully detected or not. Since the optimization process aims to make visible watermark stamps less noticeable, it may make stamps that are too small or blend too closely with the image content, making it potentially disappear, and making only the most visible big stamp can be detected.

Questions For Authors

  • Can the author provide more detail on how to prevent the watermark stamp from having identical pixel value with the background?

  • The experiment in the C8 ROBUSTNESS EXPERIMENT DETAILS check if the watermark stamp can be detect by the GPT model. However the guided prompt already mentioned that there are 6 stamps in the images, it may bias the answer from GPT. If it’s possible, can the author provide the watermark detection rate by a blind detector on both watermarked image and non watermark image to check False positive rate (FPR) and detection rate of hard watermark stamp? And also the True positive rate at certain False positive rate?

Rating: 2: Weak rejection

AuthorsOpenReview ↗

Rebuttal for Area Chair V7dL [1/2]

Thank you for the careful review. We are glad that you recognized the human-preference gains shown by our user studies and that DiffStamp applies to different generators without changing their parameters. We answer each point below, and all changes are reflected in the revision.

$ $

[W1] “The necessity of using attention maps at multiple selected layers and heads isn't well justified. A natural baseline would be to use an available segmentation model, such as SAM, to identify target concepts from the original generated image based on the text prompt and then optimize watermark placement within the resulting masks.”

Response. We ran the suggested segmentation baseline using SAM 3 and found a clear runtime advantage for native attention. The SAM 3 support masks are obtained from a completed stamp-free generation and used in a second sampling pass to retain the same in-generation stamping pipeline. DiffStamp reuses the generator's native attention during sampling, with an estimated aggregate end-to-end overhead of 5.4%. As the table shows, SAM 3 masks raise the hit rate mainly on small concepts, but at a large runtime cost.

ScenesGD: AllGD: >5%GD: ≤5%GD: No maskSAM 3 hitTime vs. no stamp
FLUX.1-dev (100)86 / 9496 / 9871 / 9176 / 7871 / 92+9.9 / +111
Qwen-Edit (100)76 / 9083 / 9456 / 9378 / 5768 / 92+4.5 / +105
All (200)81 / 9289 / 9664 / 9278 / 6669 / 92+5.4 / +106

Each cell reports Attention (ours) / SAM 3 masks. Hit rates and runtime overheads are reported in percent. Runtime overhead is relative to stamp-free generation. GD denotes Grounding DINO regions. A concept is hit if at least 25% of at least one of its stamp boxes lies inside its corresponding Grounding DINO region. The size split (>5% or ≤5%) refers to SAM 3 mask area relative to image area. “No mask” denotes concepts for which SAM 3 returns no support mask. In these cases, the SAM 3 baseline falls back to global placement.

This table makes the runtime advantage of native attention much clearer. The rationale for averaging across blocks and steps, failure examples of both maps, and timing details are provided in the new App. C.9 (pp. 28–30, Tables A12–A13, Fig. A11).

$ $

[W2] “Table 2 should include the perceptual quality scores of the original generated images without stamps. These reference values are important for assessing the actual quality degradation caused by the watermark and understanding the relative improvements achieved by the proposed method.”

Response. With the stamp-free reference added, DiffStamp shows no measurable quality degradation: its scores are at or below the reference on all three metrics. The reference images use the same prompts and seeds with the watermark disabled.

MethodNIQE ↓ILNIQE ↓BRISQUE ↓
No watermark4.87 ± 0.2234.3 ± 2.030.3 ± 2.7
Naïve Overlay5.04 ± 0.2132.5 ± 1.631.0 ± 2.5
DiffStamp (Ours)4.63 ± 0.1730.5 ± 1.326.2 ± 2.4

± is the 95% confidence interval over the 200 images.

$ $

[W3] “The paper does not evaluate whether all the embedded watermarks can be fully detected or not. Since the optimization process aims to make visible watermark stamps less noticeable, it may make stamps that are too small or blend too closely with the image content, making it potentially disappear, and making only the most visible big stamp can be detected.”

Response. The stamps do not disappear, and detection is not limited to one conspicuous stamp. In our visual-search study (108 participants), viewers found at least one stamp in 97.5% of images and at least half of the stamps in 84.7%.

  • Stamps are found, and quickly. At least one is found in 97.5% of trials, and in 89.2% within the first 5 s of the 20 s limit.
  • Not only one stamp. On average, 72% of the stamps in an image are found, and all of them in 29.0% of trials.
  • Individual stamps differ. The median stamp is found by 82% of participants, and the hardest third by 38% on average.

This variation is why we place several stamps. They are nearly the same size (93–99% of the unoptimized size), so how visible each one is depends on the content at its location, which is not known before generation. With several stamps, an image remains recognizably marked even when some of them are hard to find.

AuthorsOpenReview ↗

Rebuttal for Area Chair V7dL [2/2]

[Q1] “Can the author provide more detail on how to prevent the watermark stamp from having identical pixel value with the background?”

Response. A stamp is not pasted onto the image as fixed values. Embossing adds a signed relief to the luminance already there, Y′=Y+γ eY' = Y + \gamma\,\mathbf{e} (Eq. A15), where e\mathbf{e} is the gradient of the smoothed logo l~\tilde{\mathbf{l}} (Eq. A13). Whatever the background value, one side of each contour therefore becomes brighter than its surroundings and the other darker, so the stamp cannot take the same pixel values as the background.

$ $

[Q2] “The experiment in the C8 ROBUSTNESS EXPERIMENT DETAILS check if the watermark stamp can be detect by the GPT model. However the guided prompt already mentioned that there are 6 stamps in the images, it may bias the answer from GPT. If it’s possible, can the author provide the watermark detection rate by a blind detector on both watermarked image and non watermark image to check False positive rate (FPR) and detection rate of hard watermark stamp? And also the True positive rate at certain False positive rate?”

Response. We ran the requested blind detection. GPT-5.2 is given one image at a time without the stamp count, on the Table 3 DiffStamp images and on stamp-free images from the same prompts and seeds (200 each).

noise_stepFPR (no stamp)Detection rate (DiffStamp)TPR @ FPR ≤ 1%TPR @ FPR ≤ 10%
0 (no attack)0.1750.8500.4600.460
600.1600.7200.4150.535
1200.1550.6250.3450.435
1800.1600.5200.3300.385
  • FPR and detection rate. Without the count, the detector flags 16–18% of stamp-free images and 85.0% of DiffStamp images (52.0% after the strongest attack).
  • TPR at a fixed FPR. At 1% FPR, the TPR is 46.0% before the attack and 33.0% after the strongest.
  • Hard stamps. In the user study (W3), the median stamp is found by 82% of participants and the hardest third of the stamps by 38% on average.

The count in the original prompt did inflate the false-positive rate (43–50%, against 16–18% blind), and the blind protocol corrects this. We had stated it because the regeneration attack distorts fine detail across the whole image, so degraded stamp remnants can be mistaken for attack artifacts.

AuthorsOpenReview ↗

Summary of Rebuttal

We thank the AC and all reviewers for their thoughtful feedback and discussion. Reviewers found the problem important (E4jG), the differentiable-renderer approach novel and well-motivated (E4jG, V7dL), the method applicable to different generators without retraining (V7dL, ZX4X), and the visual-search study a sensible measure of noticeability (ZX4X). The revised paper now provides more direct evidence for DiffStamp's design choices. We summarize the main concerns and our responses below.

1. Baselines and quality reference (E4jG, V7dL)

  • Concern: Comparisons with post-processing and other visible watermarks, and a stamp-free quality reference, would help.
  • Response: DiffStamp harmonizes the mark with nearby content during sampling, which post-processing cannot do, at modest runtime cost (Table 4). It is on par with or better than DVW (Table A9) and scores better than the no-watermark reference on all three quality metrics (Table 2).

2. Quantitative ablation (E4jG, ZX4X)

  • Concern: The stage-wise ablation would be stronger with quantitative results.
  • Response: Each stage has a distinct, measurable role (Table A2). Locating preserves quality, blending adds robustness (p=0.009), filtering helps both, and embossing provides visibility.

3. Attention vs. segmentation (V7dL)

  • Concern: A segmentation model such as SAM could be an alternative to attention maps.
  • Response: Attention is far cheaper (Table A12). SAM 3 improves region hits on small concepts but needs a second sampling pass, raising overhead from +5.4% to +106%.

4. Detection (V7dL, ZX4X)

  • Concern: Small stamps may be missed, and the VLM judge is told the stamp count.
  • Response: Viewers find at least one stamp in 97.5% of images (App. C.7). Without the stamp count, a blind VLM detects 85.0% of DiffStamp images against 16–18% of stamp-free images (App. C.8, Table A11).

5. Robustness (ZX4X)

  • Concern: A stamp-free control and learned removers would strengthen the robustness evaluation.
  • Response: Under the strongest attack, DiffStamp is significantly more robust than both baselines (Table 3), and none of five blind learned removers removes the stamps (Table A10).

6. Disclosure settings (ZX4X)

  • Concern: Disclosure may call for more immediate recognition, and the examples use brand logos.
  • Response: Noticeability is tunable through embossing strength and stamp count (Fig. A2). All figures now use brand-free stamps.

We thank the AC and reviewers again. Our individual replies discuss these results in detail, and we hope the revision makes DiffStamp's contribution easier to assess.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.