acceptodds
Under review as a conference paper at ICLR 2027

C-TQIM: Cost-Aware Terminal Quantization for Inversion-Free Diffusion Watermarking

Abstract

Image watermarks must survive image processing while preserving the intended output. We study this trade-off with frozen generator and variational autoencoder (VAE) weights, no additional neural watermark decoder, and detection through one VAE encoding without diffusion inversion. We propose C-TQIM, a cost-aware terminal quantization index modulation (QIM) method that embeds a fixed keyed pilot between sampling and VAE decoding. Offline calibration estimates the RGB distortion cost and bias-preserving mean squared error of each candidate discrete cosine transform (DCT) direction under the fixed Identity-calibrated reader used at deployment. A joint criterion selects 144 carriers for nearest-lattice embedding, and an analytic detector aggregates their affine-corrected, pilot-aligned scores. On Stable Diffusion 3.5 Medium, with 1,000 threshold-calibration hosts and 4,000 matched test hosts, C-TQIM achieves a macro AUC of 0.9946 and macro true-positive rate (TPR) of 95.40% across 13 conditions. The maximum observed false-positive rate (FPR) is 0.775%, with paired PSNR of 33.36 dB and LPIPS of 0.0212. In a separate 800-host comparison matched by development PSNR, joint selection improves macro TPR over cost-only selection by 4.45 percentage points (95% paired interval: 3.85–5.04). C-TQIM provides a directly readable fixed-key presence channel without changing the sampling process or training a watermark-specific neural decoder.

Primary area:
generative models
Keywords:
Diffusion models, Image watermarking, Quantization index modulation, Inversion-free detection, Cost-aware carrier selection.

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_6uDCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents C-TQIM, a cost-aware diffusion model image watermarking technique that leverages the discrete cosine transform (DCT-II) and quantization index modulation (QIM) to embed a watermark. The diffusion model's terminal latent is used to embed the watermark to avoid diffusion inversion. Additionally, they rank candidate DCT directions using an MSE-based cost function and select the lowest-cost candidates to mitigate visual distortion.

Critical strengths
  1. The watermarking idea is interesting, and once understood, fairly elegant.
  2. The experimentation is thorough and extensive.
Critical weaknesses
  1. The writing needs significant improvements to ease the readability of the paper. The paper is dense with jargon that it does not explain or define, and the sentences are dense and hard to parse throughout. A preliminary section would help define core concepts and terminology.
  2. The mathematical notation is lacking. Many symbols are used but not defined. You cannot rely on a Figure to define terms, especially if that is not the figure's explicit purpose.  rr , ArA_r. Also, sometimes A is written as AA or A\mathcal{A}, seemingly interchangeably, which can be confusing. HwH_w and zTz_T are only defined within Figure 2. σjsrc\sigma_j^{src} and cc are not defined. 
  3. Significant digits are sometimes mishandled, showing a lack of mathematical rigor, like in Table 2, where results use up to 6 digits of precision, when only 800 samples are used. Other offending tables are Tables 10 and 11.
  4. Only a single dataset and a single model means it is hard to argue that your results would generalize to other datasets/models.
Questions for authors
  1. Adaptive attacks are not considered, and the only transformations considered are JPEG, Resize, Blur, and Noise. Why did you not consider adaptive attacks or attacks tailored to QIM-based watermarking schemes, such as [1]?
  2. In your cost-aware carrier selection, you use MSE to compute the RGB distortion cost; however, MSE is not necessarily a perceptual metric. Why did you choose MSE over other metrics to define cost? Two perturbations can have the same MSE but drastically different visual impacts.
  3. It is not clear to me what the key actually is. It seems to combine the dither and the pilot, but an explicit definition of your algorithm would be helpful.
  4. You focus exclusively on watermarking diffusion-model-generated images, but wouldn't your method generalize to any image using a VAE auto-encoder to extract a terminal latent?

[1] Pérez-González, Fernando, et al. "Rational dither modulation: A high-rate data-hiding method invariant to gain attacks." IEEE Transactions on Signal Processing 53.10 (2005): 3960-3975.

7 comments
AuthorsOpenReview ↗

Readability, terminology, and numerical reporting

Weakness 1: Readability and undefined terminology

We agree that the original manuscript introduced several concepts before providing sufficient definitions and motivation. We have reorganized the presentation and added a dedicated Preliminaries and Problem Formulation section before the method.

The new section introduces the frozen latent-diffusion pipeline, the terminal latent, the VAE encoder/decoder, DCT carrier directions, the fixed affine reader, QIM lattices, the pilot-presence detection task, and the threat model. We also reorganized the method in the order of offline calibration, carrier ranking, keyed terminal embedding, and one-encoding detection. This makes the relationship between the terminal-latent interface, carrier calibration, QIM embedding, and analytic detection explicit.

The qualitative comparison is retained in full as Figure 4 (p. 8), and the revised text introduces and interprets each main figure and table rather than relying on figures to define technical terms.

Changes in the manuscript

  • Section 3, “Preliminaries and Problem Formulation” (p. 3): introduces the frozen latent-diffusion pipeline, terminal latent, VAE round trip, DCT carriers, fixed reader, QIM lattices, and fixed-key presence-detection task.
  • Section 5, “C-TQIM Method” (pp. 5–6): presents calibration, carrier ranking, keyed embedding, and one-encoding detection in execution order.
  • Figures 1–4 and the experimental tables (pp. 4–10): are introduced and interpreted in the surrounding text; Figure 7 is retained in full as Figure 4.

Weakness 3: Significant digits and numerical rigor

We agree that the displayed precision should reflect both the sample size and the statistic being reported. In the revised main text, TP/FP counts are shown as integers, and count-derived rates are rounded to a precision appropriate to their denominators. AUC is treated separately as a ranking statistic rather than a direct sample proportion. Additional digits are retained only for frozen thresholds, configuration values, continuous AUC summaries, and historical aggregate values that must remain traceable to the submitted record.

For the new step-aware comparison, the three-budget results are recomputed on the same 200-host subset, and paired bootstrap differences are reported in percentage points. The manuscript specifies the host-level resampling unit and retains all 13 conditions within each resampled host.

Changes in the manuscript

  • Section 6.1, “Protocol and statistical conventions” (p. 6), Table 2 (p. 7), and Table 3 (p. 8): distinguish integer TP/FP counts, count-derived rates, and continuous AUC values; rates are displayed at precision appropriate to their denominators.
  • Supplementary Material, Appendix D.2–D.4 (pp. 20–21): presents count-based uncertainty and the separate Stable Signature results. Higher precision is retained for thresholds, configuration values, and historical aggregate values where it is needed for traceability; it is not used to imply greater sampling precision.
AuthorsOpenReview ↗

Mathematical notation and symbol definitions

Weakness 2: Mathematical notation and symbol definitions

We agree that the notation should be defined in the text and used consistently. The revised preliminary section now defines the text condition cc, initial noise, terminal latent HH, marked terminal latent HwH_w, source projection aija_{ij}, source scale σjsrc\sigma_j^{\mathrm{src}}, DCT directions uju_j, and the frozen VAE encoder and decoder. The embedding and detection algorithms also explicitly define their inputs and outputs.

We have now made the notation explicit at first use: ArA_r denotes the image-space operator associated with transformation rr, whereas Acal\mathcal A_{\mathrm{cal}} denotes the set of calibration transformations. We also state that zTz_T in Algorithm 2 denotes the sampler's initial-noise input zinitz_{\mathrm{init}}, while H=Gθ(c,zT)H=G_\theta(c,z_T) is the resulting terminal latent. These definitions are provided in the revised source rather than left to the figure.

Changes in the manuscript

  • Section 3.1, “Frozen latent-diffusion pipeline” (p. 3): defines cc, zinitz_{\mathrm{init}}, HH, and HwH_w.
  • Section 3.2, “DCT carriers and fixed-reader calibration” (p. 3): explicitly defines ArA_r as the image-space transformation operator indexed by rr, and distinguishes an individual operator ArA_r from the calibration set Acal\mathcal A_{\mathrm{cal}}; it also defines uju_j, aija_{ij}, and σjsrc\sigma_j^{\mathrm{src}}.
  • Supplementary Material, Appendix A.2, “Calibration and Deployment Algorithms” (p. 13): states that algorithmic zTz_T is the sampler's initial-noise input zinitz_{\mathrm{init}}, while H=Gθ(c,zT)H=G_\theta(c,z_T) is the terminal latent.
AuthorsOpenReview ↗

Cross-model generalization and adaptive/QIM-related attacks

Weakness 4: Generalization beyond one dataset and one model

We agree that the original SD3.5-only evaluation was insufficient to characterize generalization. The revised main text also places the method in a broader baseline context: seven adapted baselines are evaluated on the original shared 4,000-host cohort, while Stable Signature is reported separately because its later split includes development-exposed source IDs. A direct GROW comparison uses a separate 800-host shared test cohort with development PSNR matched within 0.2 dB. We distinguish these cohorts rather than treating them as one fully paired ranking.

We have also added an SD1.5 evaluation with model/VAE-specific recalibration, an SD3.5 abstract-prompt domain evaluation with both frozen and recalibrated settings, a native-512 evaluation, and external negative-source tests.

On SD1.5, the model-specific configuration obtains a macro TPR of 90.20% and a maximum observed FPR of 0.875% on 800 test hosts. On the SD3.5 abstract-prompt domain, the original frozen configuration has a maximum observed FPR of 3.75%; a separately recalibrated configuration obtains 1.50% on 200 test hosts. The native-512 configuration obtains a macro TPR of 92.96% and a maximum observed FPR of 1.50% on 200 hosts.

External negative-source tests include COCO real photographs, unmarked SD1.5 generations, Oxford-IIIT Pet images, and Z-Image-Turbo images. These tests characterize false-positive behavior on the corresponding negative sources; they do not constitute positive watermarking evaluations on every generator.

The revised manuscript therefore provides evidence of applicability across the evaluated SD1.5 and SD3.5 configurations after the required calibration, while distinguishing model-specific recalibration from zero-shot transfer.

Changes in the manuscript

  • Section 6.3, “Comparison with Existing Watermarking Methods” (pp. 7–8; Table 3, p. 8): adds seven shared-cohort SD3.5 adaptations and separately qualifies Stable Signature.
  • Section 6.4, “Direct Comparison with GROW” (p. 9): adds the 800-host development-PSNR-matched comparison.
  • Section 6.6, “Cross-model, domain, resolution, and negative-domain tests” (p. 10): reports SD1.5, domain, resolution, and external-negative evaluations.
  • Supplementary Material, Appendix C (pp. 18–19) and Appendix K.3 (pp. 37–38): provides the method adaptations and cohort-specific results.

Question 1: Adaptive attacks and QIM-related attacks

We agree that the original evaluation did not include adaptive removal or geometric transformations. We have added four attack settings that distinguish public-model access, detector feedback, and key disclosure.

With public VAE access but no key or detector oracle, a latent-drift attack under an RGB ℓ∞\ell_\infty budget of 4/255 removes detection from 175/200 initially detected marked hosts (87.5%). Under an unknown key, the tested continuous-score oracle attack achieves 0/200 removals at 100 and 500 queries; the binary accept/reject oracle attack achieves 9/200 and 13/200 removals at the same budgets. With full key and detector access, the white-box attack removes 200/200 marked inputs and causes 200/200 clean inputs to be accepted after optimization.

We also add gain and contrast transformations, alongside crop, rotation, translation, VAE reconstruction, diffusion regeneration, and composed transformations. The cited Rational Dither Modulation (RDM) work is relevant to gain-invariant data hiding and motivates examining gain changes. We include gain transformations as stress tests, but do not present RDM as an adaptive attack or claim to have reproduced a dedicated RDM attack implementation. The new results are reported under their corresponding fixed-threshold or independently calibrated protocols.

Changes in the manuscript

  • Section 6.7, “Geometric, regeneration, and external-operator evaluation” (p. 10): reports the three official WAVES operators and the separate T09-E evaluation.
  • Section 6.8, “Adaptive attacks and cross-key behavior” (p. 10): reports the public-VAE, query-based, and full-key attacks, together with cross-key acceptance.
  • Supplementary Material, Appendix K.4–K.5 (pp. 38–40): provides the corresponding transformation and attack protocols.
Reviewer_vgbjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Cost-Aware Terminal Quantization Index Modulation (C-TQIM), an inversion-free watermarking framework for latent diffusion models that keeps the generator and VAE weights frozen and requires no watermark-specific neural decoder. Through a six-position observability study, the authors motivate embedding at the terminal latent interface, after sampling and before VAE decoding, to enable direct detection via a single VAE encoding. C-TQIM selects latent DCT carriers using a joint criterion based on RGB distortion cost and bias-preserving fixed-reader mean squared error (MSE), then applies keyed quantization index modulation (QIM) and analytic fixed-key presence detection. On Stable Diffusion 3.5 Medium, the method achieves a macro AUC of 0.9946, a macro true-positive rate (TPR) of 95.40% across 13 evaluation conditions, a maximum observed false-positive rate (FPR) of 0.775%, and a paired PSNR of 33.36 dB. The evaluation nevertheless leaves geometric attacks, regeneration-based removal, adaptive attacks, and generalization to other VAEs and generators untested.

Critical strengths

Inversion-Free Detection Efficiency: C-TQIM eliminates the need for diffusion inversion at detection time. Its detector uses a single VAE encoding followed by affine correction and analytic QIM scoring, without reconstructing the initial noise or evaluating the denoiser. Systematic Latent Observability Study: The authors conduct a six-position observability study across the SD3.5 sampling trajectory. The results show improved detection and paired image preservation for the evaluated terminal configuration, providing empirical motivation for post-sampling embedding and single-encoding detection.

Critical weaknesses

a. Limited Evaluation Against External Robustness Benchmarks: C-TQIM is evaluated using a custom suite of 13 non-geometric conditions covering JPEG compression, resizing, Gaussian blur, and Gaussian noise. Although these experiments provide useful evidence of robustness to common image-processing operations, the paper does not report results under an established external benchmark such as WAVES (ICML'24), WaterBench (ICLR'25), MarkDiffusion (JMLR'26), WARP (ACM MM'26), or classical Stirmark. The evaluation also excludes geometric transformations and regeneration-based removal, limiting the strength of its practical robustness claims. Moreover, all main experiments use Stable Diffusion 3.5 Medium, leaving generalization to other generators, resolutions, and data domains unestablished. b. Limited Security Analysis Under a Public-Model Threat Model: C-TQIM relies on a fixed VAE encoder, calibrated carrier directions, and an analytic detector. The paper does not sufficiently characterize the security of this design when the VAE architecture and weights are public and an adversary can query the detector or adaptively modify images. The reported cross-key acceptance results also raise concerns about reliable multi-key attribution. The authors should clarify the intended threat model and evaluate adaptive removal, detector-query attacks, and cross-key false acceptance before making broader security claims. The absence of such experiments does not itself establish that the method is vulnerable, but it limits the security conclusions that can be drawn.

Questions for authors

a. Experimental Reproducibility and Key Management: Section 4.4 states that seeds 337 and 1201 generate the experimental dither fractions and pilot, respectively, and explicitly notes that these seeds do not define a cryptographic key-derivation protocol. Could the authors clarify how keys would be generated, managed, and protected in deployment? What security guarantees, if any, are provided when the key-generation procedure and detector configuration are known to an adversary? b. Generalization Beyond SD3.5 Medium: All main experiments are conducted on Stable Diffusion 3.5 Medium. Is the method applicable to other latent diffusion or rectified-flow generators and their corresponding VAEs? If so, which parts of the carrier calibration and reader configuration must be recomputed, and what evidence supports generalization beyond the evaluated model, resolution, and caption domain? c. External Robustness Benchmark Evaluation: The evaluation covers 13 non-geometric image-processing conditions but does not report results on an external benchmark such as WAVES. Could the authors evaluate C-TQIM under WAVES, or justify why the benchmark's threat models and evaluation protocol are not directly applicable? How would the conclusions change under a broader and standardized attack evaluation? d. Robustness to Geometric and Regeneration-Based Attacks: The paper evaluates JPEG compression, resizing, blur, and Gaussian noise, but does not test geometric transformations or regeneration-based removal. How does C-TQIM perform under mild cropping, rotation, or generative purification attacks? Can the authors provide quantitative results using clearly specified attack parameters and a fixed detection threshold? e. Security Against Black-Box and Adaptive Attacks: Given that the detector uses a fixed public-model VAE and analytic scoring, what security guarantees can the authors provide against black-box detector queries, adaptive watermark removal, or cross-key attacks? If formal guarantees are outside the scope of the current method, can the authors characterize these limitations experimentally and clarify the intended security threat model?

5 comments
AuthorsOpenReview ↗

External robustness evaluation and benchmark coverage

Weakness a: Limited evaluation against external robustness benchmarks

We agree that the original 13-condition suite did not adequately cover geometric and regeneration transformations or establish behavior under model and domain shifts. We therefore added an official WAVES-operator subset, a separate 33-condition T09-E evaluation, cross-model/domain/resolution tests, and external negative-source tests. We report each study with its own cohort and calibration protocol.

First, we evaluate three official RGB operators from the WAVES repository: resized-crop area 0.95, resized-crop area 0.90, and rotation by +3 degrees. Under the frozen D threshold on 800 positive and 800 negative images, these yield TPRs of 97.75%, 87.50%, and 88.38%, respectively. We identify this evaluation precisely as a subset of official WAVES operators.

Second, we add custom geometric and regeneration transformations, including reverse rotation, translation, VAE reconstruction, diffusion regeneration, gain/contrast changes, and composed JPEG-plus-blur transformations. The separate T09-E evaluation covers 33 conditions on 200 positive and 200 negative hosts per condition, with an independently calibrated frozen threshold of 0.08657886798027904. Its macro TPR is 79.9394%; each condition has 0/200 observed false positives. The corresponding two-sided 95% Wilson upper bound is approximately 1.885%, so the observed zero count is not interpreted as a zero population FPR. The recovered protocol manifest now identifies all 33 conditions: 12 reused condition IDs and 21 newly added operator/parameter pairs, including resized-crop, rotations, translations, VAE reconstruction, gain/contrast, erasing, brightness, blur, noise, JPEG, composed crop-plus-JPEG, and regeneration (strength 0.3). The per-condition TP counts are now reported in Appendix K.4. They range from 12/200 for translate_right005 and 27/200 for rotate_m5 to 200/200 for several VAE, gain, and regeneration conditions; all conditions have 0/200 observed false positives. The condition-level results make the transformation-specific detection variation explicit.

Third, we extend the model/domain evaluation. SD1.5 is evaluated with model/VAE-specific recalibration (800 test hosts; macro TPR 90.20%; maximum observed FPR 0.875%). For an SD3.5 abstract-prompt domain, the original frozen configuration has a maximum observed FPR of 3.75%; a separate domain-recalibrated evaluation reduces it to 1.50% on 200 test hosts. We also evaluate an SD3.5 native-512 configuration on 200 hosts. External negative-source tests include COCO real photographs, unmarked SD1.5 generations, Oxford-IIIT Pet, and Z-Image-Turbo images.

Together, these evaluations add evidence beyond the original JPEG/resize/blur/noise suite. They also show that performance is transformation- and domain-dependent: frozen transfer can raise FPR, and stronger geometric or regeneration transformations reduce detection. We therefore state the measured operating conditions rather than generalize to untested attacks or generators.

Changes in the manuscript

  • Section 6.7, “Geometric, regeneration, and external-operator evaluation” (p. 10): reports the three official WAVES RGB operators under the frozen D threshold and distinguishes them from custom transformations and the independent T09-E study.
  • Section 6.6, “Cross-model, domain, resolution, and negative-domain tests” (p. 10): adds SD1.5, abstract-prompt, native-512, and external-negative evaluations.
  • Supplementary Material, Appendix K.3–K.4 (pp. 37–39): provides the separate cohort definitions, operator results, and T09-E aggregate protocol.
AuthorsOpenReview ↗

Security under public-model and adaptive-access threat models

Weakness b: Limited security analysis under a public-model threat model

We agree that the security discussion should distinguish attacker capabilities and should not infer security from the use of a fixed analytic detector. We have added a threat-model taxonomy and evaluated removal and forgery under four access settings.

  • A1: Public VAE, unknown key, no detector oracle. A public latent-drift attack with an RGB ℓ∞\ell_\infty budget of 4/255, 20 steps, and three restarts removes detection from 175/200 initially detected marked hosts (87.5%).
  • A2: Unknown key, continuous score oracle. The tested SPSA procedure achieves 0/200 removals at both 100 and 500 queries.
  • A4: Unknown key, binary accept/reject oracle. The tested search removes 9/200 (4.5%) at 100 queries and 13/200 (6.5%) at 500 queries.
  • A3: Full key and detector white-box access. The white-box attack removes the watermark from 200/200 marked inputs. The corresponding clean-image forgery test causes all 200 clean inputs to be accepted after optimization.

These are empirical outcomes for the specified attacks, budgets, and cohort; they do not constitute formal security guarantees. In particular, the A1 result shows that access to a public VAE can enable substantial removal even without detector feedback, while A3 shows that key disclosure permits both removal and forgery.

We also re-examined cross-key acceptance. The fixed wrong-key pair key001 →\rightarrow key043 accepts 219/800 marked images in the discovery cohort and 204/800 (25.5%) in an independent fixed-pair confirmation; the confirmation clean false-accept count is 3/800. A separate 105-decoder maximum-threshold search yields zero wrong-key OR accepts among 1,000 owner-host records, but these records reuse only 200 unique marked hosts and use a different global threshold. It therefore does not resolve the fixed-pair false acceptance. We retain both results and explicitly limit the supported task to fixed-key presence detection rather than reliable multi-key attribution.

Changes in the manuscript

  • Section 3.4, “Threat model” (p. 3): separates public-model access, score-oracle access, binary feedback, and key disclosure.
  • Section 6.8, “Adaptive attacks and cross-key behavior” (p. 10): reports A1–A4, the clean-image forgery result, and the fixed-pair cross-key confirmation.
  • Section 7, “Limitations and Conclusion” (p. 10): states the resulting security and multi-key attribution boundaries.
  • Supplementary Material, Appendix K.5 (p. 40): gives attack protocols and counting units.
AuthorsOpenReview ↗

Key generation, management, and model generalization

Question a: Experimental reproducibility and key management

The reviewer is correct that the experimental seeds are reproducibility parameters, not a secure key-management mechanism. In the experiments, seed 337 generates the dither fractions and seed 1201 generates the pilot. These values are disclosed for reproducibility; consequently, the current experimental implementation does not provide cryptographic key secrecy against an adversary who can reconstruct the released key arrays.

We have also clarified the effective QIM phase: ϕj=(djΔj+pj+14) mod 1.\phi_j=\left(\frac{d_j}{\Delta_j}+\frac{p_j+1}{4}\right)\bmod 1. This makes explicit that pilot and dither are not independent sources of key entropy: different pilot/dither pairs can represent the same target lattice. We therefore do not claim a key-entropy level from their nominal lengths.

A practical deployment would require a high-entropy secret master key, a cryptographic KDF/PRF to derive domain-separated pilot and dither streams, secure key storage, access control, rotation, and revocation. These components are deployment requirements and are not implemented or evaluated in the present work. Accordingly, we make no formal cryptographic security claim. The unknown-key query experiments should be interpreted as restricted-access attack simulations in which the attack procedure is not given the key arrays, not as proof of secrecy when the key-generation procedure and artifacts are public.

Changes in the manuscript

  • Section 5.2, “Keyed terminal embedding” (p. 6): defines the experimental pilot/dither seeds and the effective QIM phase.
  • Section 3.4, “Threat model” (p. 3), and Section 7 (p. 10): state the protected-key assumption and clarify that the experiments do not provide cryptographic key secrecy.
  • Supplementary Material, Appendix K.5 (p. 40): details the information and access available to each attacker.

Question b: Generalization beyond SD3.5 Medium

C-TQIM operates on the terminal latent and the corresponding VAE round-trip channel. Its carrier calibration is therefore model/VAE-specific. When the generator or VAE changes, the latent shape and coordinate statistics, candidate directions and validity mask, source scales σjsrc\sigma_j^{\mathrm{src}}, RGB distortion costs qjq_j, affine reader parameters (αj,βj)(\alpha_j,\beta_j), fixed-reader errors νj2\nu_j^2, ordered carrier set, and operating threshold must be recomputed for the new pipeline. The QIM embedding and analytic detection equations remain unchanged once these quantities are calibrated.

We provide three forms of evidence. First, SD1.5 is evaluated with its own VAE-specific calibration and obtains a macro TPR of 90.20% and maximum observed FPR of 0.875% on 800 test hosts. Second, an SD3.5 abstract-prompt domain is evaluated with the original frozen configuration and with separate domain recalibration; the frozen configuration has a maximum observed FPR of 3.75%, while the recalibrated setting has 1.50% on a separate 200-host test cohort. Third, an SD3.5 native-512 configuration is calibrated for its latent geometry and evaluated on 200 hosts, obtaining a macro TPR of 92.96% and maximum observed FPR of 1.50%.

We also test negative-only images from SD1.5 and Z-Image-Turbo, in addition to real photographs and Oxford-IIIT Pet images. These results characterize false-positive behavior on external sources; they are not positive embedding evaluations on those generators. We have not yet evaluated a separate rectified-flow generator. We therefore claim applicability after model/VAE-specific calibration only for the evaluated SD1.5 and SD3.5 settings, not universal transfer to arbitrary latent-diffusion or rectified-flow models.

Changes in the manuscript

  • Section 6.6 (p. 10): reports SD1.5 with model/VAE-specific calibration, SD3.5 domain transfer and recalibration, native-512 evaluation, and negative-only external sources.
  • Supplementary Material, Appendix C (pp. 18–19) and Appendix K.3 (pp. 37–38): documents the baseline/model-specific adaptation protocols and the separate cross-model/domain cohorts.
Reviewer_tLXHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes C-TQIM, a watermarking method that embeds a fixed keyed pilot into the terminal latent of a diffusion model after sampling and before VAE decoding. The generator and VAE remain frozen, and detection requires one VAE encoding followed by analytic computation. The method selects DCT carriers using a criterion that combines estimated RGB distortion cost with readout error under a fixed, calibrated detector. On SD3.5 Medium, it achieves 95.40% macro TPR across 13 conditions, with a maximum observed FPR of 0.775% and paired PSNR of 33.36 dB. The main contribution is the combination of terminal quantization and calibration of the frozen VAE channel to balance image preservation and watermark readability.

Critical strengths
  1. The method is simple and technically coherent.
  2. The experimental results shows the performance is competitive.
Critical weaknesses

The closest experimental comparison is missing. GROW, discussed in the related-work section, also uses frozen model weights and inversion-free detection through a single VAE encoding. Without a direct comparison, it is difficult to determine the advantage of terminal QIM over an existing method satisfying most of the same deployment constraints. The main comparisons against adapted noise-domain methods do not resolve this question.

The proposed carrier-selection criterion remains a heuristic, and the single-factor controls could be stronger. Quantization steps vary with each carrier’s source standard deviation, but the cost-only and MSE-only controls rank carriers using unscaled distortion coefficients and unnormalized errors. Comparisons against selection rules that account for the actual quantization steps would better establish the benefit of the proposed joint criterion.

The incremental contribution of some emphasized components is limited. In Table 2 and Table 29, Full improves over MSE-only selection by 0.78 percentage points, while its 0.27-point improvement over centered-error selection has a confidence interval containing zero. These results support the overall method but provide weaker evidence for cost weighting and bias retention as substantial independent contributions. The matched comparison also covers only one distortion budget.

The evaluation distribution is narrow. Experiments use one generator, one resolution, and one caption domain, while negative examples come from the same generator. False-positive behavior on real photographs and outputs from other generators remains unknown. The transformation suite also excludes cropping, translation, rotation, regeneration, and composed attacks, limiting the evidence for broader robustness.

The stability and reproducibility of carrier calibration require further clarification. Appendix A.2 estimates RGB distortion costs using only four calibration images, but does not establish whether the selected carriers remain stable across different probe samples. Exact probe displacements, numerical validity rules, and RGB conversion conventions should also be specified clearly.

Questions for authors

N/A

4 comments
AuthorsOpenReview ↗

Direct comparison with GROW

Comment 1: Missing comparison with GROW

We agree that GROW is a particularly relevant comparator because it also uses frozen model components and a single VAE encoding at detection. We have therefore expanded the main-text comparison in two complementary parts: a consolidated comparison with eight baselines, followed by a development-PSNR-matched direct comparison with GROW.

The new main-text Table 3 includes Gaussian Shading, HSQR, HSTR, PRC, RingID, Tree-Ring, WIND, and Stable Signature. Seven methods are evaluated on the original shared 4,000-host test cohort; Stable Signature is shown in a separately qualified later split because 600 source IDs were exposed during development. Tree-Ring is explicitly identified as a modified local-template variant. The seven shared-cohort methods use a common 10-step, empty-prompt reverse FlowMatch/Euler-style detection adaptation without CFG; we describe them as documented SD3.5 adaptations rather than claiming that each is an unchanged, individually optimized native implementation. On the shared cohort, Gaussian Shading and WIND attain higher macro and worst-condition TPR than C-TQIM, while PRC is close in macro TPR but has a higher maximum FPR. C-TQIM has the highest paired PSNR among the seven adapted noise-domain baselines. We do not treat the Stable Signature row as a fully host-matched ranking.

The GROW comparison is then conducted on 800 shared test hosts, with development PSNR matched within 0.2 dB. GROW achieves macro AUC/TPR and worst-condition TPR of 0.999924/99.9231%/99.3750%, compared with 0.992794/93.5481%/53.1250% for C-TQIM-B1. C-TQIM has a lower observed maximum FPR (0.625% vs. 0.750%) and lower LPIPS (0.01575 vs. 0.01867), while its test PSNR is also lower (34.916 vs. 35.623 dB). The two methods also have different payload objectives: GROW recovers a known 32-bit message, whereas C-TQIM detects a fixed 144-symbol pilot.

The measured detection latencies are similar (58.725 and 59.425 ms/image under the recorded timing boundary), so we do not claim a meaningful speed advantage. B2 and same-200 follow-up comparisons are included in the Supplementary Material.

Changes in the manuscript

  • Section 6.3, “Comparison with Existing Watermarking Methods” (pp. 7–8; Table 3, p. 8): adds the eight-baseline comparison. The seven SD3.5 adaptations share the original 4,000-host test cohort and use the documented common 10-step empty-prompt reverse FlowMatch/Euler-style detection protocol; Stable Signature is placed in a separately qualified block because its later split includes 600 development-exposed source IDs. Appendix C (pp. 18–19) details the method-specific adaptations.
  • Section 6.4, “Direct Comparison with GROW” (p. 9): presents the 800-host shared-test comparison, development-PSNR matching, detection metrics, image-quality trade-off, and task-semantics distinction.
  • Supplementary Material, Appendix C (pp. 18–19) and Appendix K.1 (p. 35): documents baseline adaptations and the full GROW B1/B2/same-200 results.
AuthorsOpenReview ↗

Carrier-selection criterion and component contributions

Comment 2: Carrier-selection criterion and step-aware controls

We agree that the original raw cost-only and raw MSE-only controls did not account explicitly for carrier-dependent quantization steps. We have added step-aware controls that use the actual step Δj=τσjsrc\Delta_j=\tau\sigma_j^{\mathrm{src}}:

C_j^{\mathrm{StepMSE}}=\frac{\nu_j^2}{\Delta_j^2}.$$ These are compared with the proposed joint criterion $C_j^{\mathrm{Full}}=q_j\nu_j^2$ and a centered-error variant. The joint score can be written as the product of the step-scaled distortion term and the step-normalized reader-error term. We now present it as a practical ranking criterion rather than a claim of global optimality. To address the single-budget limitation with a new study, we recomputed B1 and B2 on the same 200 test hosts used by B3. All 13 conditions are retained within each host, and paired bootstrap resampling is performed at the host level (10,000 replicates). Each ranking rule uses its own calibrated frozen threshold, so the comparison is between operating points rather than at an exactly matched test FPR. The results show that Full is consistently stronger than StepMSE: the macro-TPR gains are 6.65, 10.08, and 18.38 percentage points at B1, B2, and B3, respectively, with 95% confidence intervals excluding zero. However, Full does not consistently outperform StepCost. At B1, Full is lower by 0.65 points (95% CI: [-1.23, -0.08]); at B2 and B3, the corresponding intervals include zero. The centered-error comparison also does not show a consistent positive gain. We have revised the contribution statement accordingly. The evidence supports combining distortion cost with fixed-reader reliability over the step-normalized MSE-only control in the tested budgets, but does not establish consistent superiority over StepCost. **Changes in the manuscript** - **Section 5.1, “Offline calibration and carrier ranking” (p. 5):** defines the RGB distortion and fixed-reader error terms and explains the joint score as a practical ranking criterion. - **Section 6.5, “Step-aware carrier-selection ablations” (p. 9; Table 4):** adds StepCost, StepMSE, and Centered controls at three budgets, all evaluated on the common 200-host subset. - **Supplementary Material, Appendix K.2 (pp. 36–37):** provides the common-cohort results and host-paired bootstrap intervals. --- ### Comment 3: Incremental contribution of cost weighting and bias retention The historical submitted matched-control analysis reports Full--MSE-only = 0.7788 percentage points (95% CI [0.3077, 1.2500]) and Full--Centered-error = 0.2692 points (95% CI [-0.2019, 0.7599]), consistent with the reviewer's summary. These are retained in Appendix J as historical aggregate results; the original per-sample export is unavailable, so they cannot be independently reaggregated in this revision. The new November step-aware study is a separate cohort and should be interpreted as an additional, not replacement, analysis. Across the common 200-host cohort, Full improves over StepMSE by 6.65, 10.08, and 18.38 percentage points at B1–B3, respectively. In contrast, its difference from StepCost is budget-dependent: Full is lower at B1, while the B2 and B3 confidence intervals include zero. Full also does not show a consistent advantage over Centered. These results lead us to narrow the interpretation of the components. We retain the joint criterion because it combines an image-distortion proxy with the deployed reader's error, and it consistently improves over StepMSE in the tested budgets. We no longer present bias retention as a separately established source of substantial performance gain. Rather, retaining the mean error is consistent with evaluating the actual fixed reader, whose systematic offset is part of its deployed behavior. **Changes in the manuscript** - **Section 6.5 (p. 9):** reports the three-budget contrasts and narrows the interpretation: Full consistently improves over StepMSE, but its advantage over StepCost depends on the budget; Centered does not show a consistent gain. - **Introduction (pp. 1–2) and Section 7, “Limitations and Conclusion” (p. 10):** align the contribution and conclusion with these results. - **Supplementary Material, Appendix J (pp. 32–35) and Appendix K.2 (pp. 36–37):** distinguishes the historical submitted matched-control aggregates from the newly executed step-aware cohort.
AuthorsOpenReview ↗

Evaluation coverage, generalization, and attack robustness

Comment 4: Evaluation distribution and attack coverage

We agree that the original evaluation did not sufficiently characterize model, domain, and transformation shifts. We have added several complementary evaluations and now distinguish cross-model positive tests from negative-only external-source tests.

First, we evaluate SD1.5 with its own VAE-specific calibration. On 800 test hosts, it obtains a macro AUC of 0.9806, a macro TPR of 90.20%, and a maximum observed FPR of 0.875%. This is evidence of applicability after model-specific recalibration, rather than zero-shot transfer of SD3.5 calibration parameters.

Second, we evaluate an SD3.5 abstract-prompt domain under both the original frozen configuration and a separately recalibrated configuration. With the original configuration, the maximum observed FPR is 3.75%; after domain recalibration, it is 1.50% on a separate 200-host test cohort. A native-512 configuration is also evaluated on 200 hosts, with a macro TPR of 92.96% and a maximum observed FPR of 1.50%.

Third, we add external negative-source tests using a frozen threshold: COCO real photographs (7/800 false accepts), unmarked SD1.5 generations (7/800), Oxford-IIIT Pet images (1/200), and Z-Image-Turbo images (3/200). These are negative-only tests and therefore characterize FPR rather than TPR or AUC.

Finally, we add geometric, regeneration, and composed transformations. Under the frozen D threshold on 800 positive and 800 negative images, resized-crop area 0.95 yields 97.75% TPR, resized-crop area 0.90 yields 87.50%, and rotation by +3 degrees yields 88.38%. Additional custom conditions include translation, reverse rotation, VAE reconstruction, diffusion regeneration, gain/contrast changes, and composed JPEG-plus-blur transformations. A separate T09-E evaluation covers 33 conditions with 200 positive and 200 negative hosts per condition; its macro TPR is 79.9394%, and no false positives are observed in the 200 negatives per condition. The recovered protocol manifest identifies 21 new operator/parameter pairs and 12 reused condition IDs. Per-condition TP counts are now included in Appendix K.4. They range from 12/200 (translate_right005) and 27/200 (rotate_m5) to 200/200 for several VAE, gain, and regeneration conditions. All 33 conditions have 0/200 observed false positives. The macro TPR remains 79.9394%. We report T09-E as a separate evaluation with its own threshold, rather than merging it with the original 13-condition benchmark or the WAVES operator subset.

These additions broaden the evidence while also showing that performance depends on the model, domain, and transformation. We have revised the robustness discussion to report these conditions directly rather than make a blanket robustness claim.

Changes in the manuscript

  • Section 6.6, “Cross-model, domain, resolution, and negative-domain tests” (p. 10): reports SD1.5, abstract-domain, native-512, and external-negative results with their separate calibration regimes.
  • Section 6.7, “Geometric, regeneration, and external-operator evaluation” (p. 10): reports the three official WAVES operators and the separate T09-E evaluation.
  • Section 6.8, “Adaptive attacks and cross-key behavior” (p. 10): reports the four attacker-access settings and fixed-pair cross-key confirmation.
  • Supplementary Material, Appendix K.3–K.5 (pp. 37–40): provides the corresponding cross-domain, condition-level WAVES/T09-E, and adaptive-attack details. Appendix K.4, Table 40 (p. 39) gives the 33 TP/FP counts.
AuthorsOpenReview ↗

Probe calibration, numerical validity, and reproducibility

Comment 5: Calibration stability and reproducibility

We agree that the original description did not provide enough detail about the probe protocol or the stability of the resulting carrier ranking. We have clarified the implementation and added a probe-size sensitivity study.

The symmetric probe displacement is ηj=0.25σjsrc\eta_j=0.25\sigma_j^{\mathrm{src}}, where the source standard deviation is computed using the sample convention (ddof=1)(ddof=1). RGB distortion is measured on the continuous clipped RGB output before uint8 conversion. The recovered implementation specifies the fixed-Identity-reader C-TQIM-D mask over the 256 candidate IDs: finite and positive qjq_j; finite αj\alpha_j with ∣αj∣>10−8|\alpha_j|>10^{-8}; σjsrc>10−8\sigma_j^{\mathrm{src}}>10^{-8}; and finite qjνj2q_j\nu_j^2. The selected 144 candidates are obtained by a stable ascending sort of the valid IDs, with candidate-ID order retained for ties. A zero ranking cost is not itself an exclusion criterion. The archived C-TQIM-D run has 256 valid candidates, and the resulting ordered 144 IDs match the saved deployment ranking.

We also distinguish this rule from the legacy mask used by the original 4,000-host main experiment. That historical rule requires strictly positive ranking cost and uses the per-condition source-noise construction; it must not be conflated with the fixed-Identity-reader C-TQIM-D rule. The probe protocol is specified as ηj=0.25σjsrc\eta_j=0.25\sigma_j^{\mathrm{src}}, sample standard deviation with ddof=1ddof=1, and continuous clipped RGB measured before uint8 conversion.

The later stability study compares probe pools of 4, 8, 16, and 32 images. The Spearman rank correlations are 0.9971, 0.9986, 1.0000, and 1.0000, respectively; the corresponding top-set Jaccard similarities are 0.9700, 0.9835, 1.0000, and 1.0000. We clarify that the 16- and 32-probe comparisons include self-comparison with the reference pool, so they should not be interpreted as independent replications. The results support ranking stability under the tested resampling protocol.

The original per-image, per-direction probe export was not retained. Accordingly, the later stability experiment is reported as a separate study, not as a reconstruction of the historical four-image probe records.

Changes in the manuscript

  • Section 5.1, “Offline calibration and carrier ranking” (p. 5): specifies the symmetric probe displacement ηj=0.25σjsrc\eta_j=0.25\sigma_j^{\mathrm{src}} and continuous clipped-RGB measurement path.
  • Supplementary Material, Appendix A.3 (p. 14): gives the exact fixed-Identity-reader C-TQIM-D validity mask, stable sorting/tie-breaking rule, 256-valid/144-selected check, and the distinct legacy mask for the original main experiment.
  • Supplementary Material, Appendix A.2 (p. 13) and Appendix K.6 (pp. 40–41): specify the four-host probe, ηj=0.25σjsrc\eta_j=0.25\sigma_j^{\mathrm{src}}, ddof=1ddof=1, and continuous clipped-RGB measurement. The historical per-image/per-direction probe export remains unavailable; the later probe-size study is kept separate.

We thank the reviewer again for identifying these issues. The revision adds the requested direct comparison and stronger controls, broadens the evaluation, and clarifies the calibration protocol. It also narrows the claims where the new evidence does not support a general advantage.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.