acceptodds
Under review as a conference paper at ICLR 2027

FROST:Improving Image Fidelity and Watermark Robustness without Updating the Base Model

Abstract

Invisible image watermarking aims to preserve the visual quality of cover images while enabling reliable recovery of embedded messages after image transformations. In autoencoder-based systems such as RoSteALS, output distortion can arise from both message embedding and image reconstruction. We introduce FROST (Frozen-Reader Output-Space Tuning), a framework for improving image fidelity and message recovery robustness without updating the pretrained watermarking model. Its trainable component, the FROST Module, is attached after the image decoder and predicts an additive residual conditioned on the watermarked image and the embedded message. All existing image and message encoders and decoders remain frozen. The complete FROST Adapter combines this module with reconstruction-error correction and high-frequency residual suppression at inference time. Component-wise comparisons show that the module improves bit accuracy under image corruptions, whereas the subsequent residual processing improves fidelity. With 100-bit messages, the complete framework increases PSNR on CLIC from 29.10 to 33.66 dB. Across CLIC and MetFaces, the reported mean bit accuracy over 15 ImageNet-C corruptions increases from 94.18% to 95.74% at severity 3 and from 89.17% to 91.67% at severity 5. Experiments with a KLf8 backbone also improve PSNR, SSIM, and average bit accuracy under corruption. However, LPIPS improvements are not consistent across datasets and backbones. These results demonstrate that learned output adjustment and inferencetime residual processing can improve evaluated watermarking systems while preserving their pretrained components.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Watermarking, Invisible Watermarking

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_efzoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

FROST improves watermark recovery without modifying the pretrained encoder or decoder, requiring only a lightweight additional module.

The complete framework improves CLIC PSNR from 29.10 to 33.66 dB and achieves modest gains in average bit accuracy under the evaluated corruptions.

Critical strengths

FROST improves watermark recovery without modifying the pretrained encoder or decoder, requiring only a lightweight additional module.

Critical weaknesses
  1. Experiments mainly compare against RoSteALS (Workshop paper), without evaluating stronger recent watermarking methods such as TrustMark, VINE, MBRS (MM21) and FIN (AAAI23). The results therefore do not establish competitiveness with the current state of the art.

  2. The FROST Module alone decreases PSNR, while most fidelity improvements arise after reconstruction-error correction and residual filtering. A crucial baseline applying these two operations directly to RoSteALS, without the learned module, is missing. Additionally, the ablations use different effective batch sizes, and all test images share a single fixed 100-bit message. These limitations weaken the conclusions about individual components and message generalization.

Questions for authors

Can the authors provide comparisons with stronger recent methods under matched image resolution, payload, image quality, and JPEG compression settings (e.g., Q50 and Q70)?

What performance can be achieved by applying reconstruction-error correction and B0 removal directly to RoSteALS, without training the FROST Module?

2 comments
AuthorsOpenReview ↗

Response to Reviewer efzo

We appreciate the reviewer's valuable suggestions regarding modern baselines and evaluation fairness. We added further comparisons and controlled experiments to clarify the strengths and limitations of FROST.

1. Recent methods, matched conditions, and JPEG Q50/Q70

We re-evaluated seven systems on one server with a shared evaluation codebase. Selected CLIC results are shown below. PSNR is in dB; accuracy is in %. IC-all averages 15 ImageNet-C types at severities 3 and 5; explicit JPEG Q50/Q70 are separate tests.

SystemPSNRIC-allQ50 / Q70
RoSteALS28.142791.48197.341 / 98.514
FROST VQ (A_orig)33.660692.99398.203 / 98.767
TrustMark-Q41.621171.79197.544 / 98.836
WAM37.017186.01996.055 / 99.424
VINE-R35.174986.42399.723 / 99.930
PixelSeal44.174366.42184.523 / 93.876

These are measured operating points, not fully matched payload/quality comparisons. The common ImageNet-C branch retains the predefined 256→224→corruption→256 operator used in FROST training; explicit JPEG uses a separate branch. Removing resizing changes that attack. Each pretrained model retains its original training augmentations.

Quality-matched evidence. A separate CLIC-530 study uses A_ref, distinct from A_orig above. At PSNR 35 dB, FROST−VINE-R is +4.8 pp [4.5, 5.1] in IC-all but −2.1 pp [−2.4, −1.9] in mean JPEG Q50/Q70. Both share the 256-pixel, 100-bit setting. Against TrustMark-Q at 35 dB, Q50 is −0.78 pp [−1.11, −0.45] and Q70 is −0.64 pp. Brackets are 95% paired image-bootstrap CIs.

Auxiliary IC-all differences against MBRS-256, FIN-HEAVY, and FIN-JPEG are +14.9/+8.0/+9.5 pp at 35 dB. MBRS uses 256 bits (100 input bits plus deterministic extra bits); FIN uses 128-pixel images and 64 bits, so these are not fully payload/resolution-matched comparisons. PSNR matching also does not match LPIPS. These results identify protocol-specific strengths and limitations rather than universal superiority.

2. REC+B0 without the learned module

We directly compared module-free REC+B0 (R2) with Full FROST (R5), sharing the same reconstruction correction and filter. This separate VQ-GAN control uses A_ref, paired images/messages/attacks, 256-pixel images, and image-specific random 100-bit messages.

On CLIC-530, R2 gives PSNR 34.95 dB, LPIPS 0.0120, and IC-all 90.22%; R5 gives 34.29 dB, 0.0186, and 92.27%. R5−R2 is +2.05 pp [1.97, 2.13]. On MetFaces-1,336, the gain is +2.32 pp [2.27, 2.37], again at worse PSNR/LPIPS. These CIs refer to this paired control, not the latest benchmark.

Thus fixed processing provides most fidelity gains, while the learned module improves recovery at these operating points. This control alone does not establish an equal-distortion advantage.

3. Random messages

Original training already sampled a random 100-bit message per sample; the fixed seed-42 message was an evaluation restriction. We now assign a different random message to each image and repeat assignments five times on CLIC-530 and four on MetFaces-1,336. Methods share image/message/attack pairs; attacks on one watermarked image recover the same embedded payload.

A_ref's mean IC-all gains over R2 are +2.00/+2.32 pp on CLIC/MetFaces and are positive in every assignment. This tests message generalization with fixed checkpoints; it does not measure training-seed uncertainty.

4. Matched-budget ablations

We retrained VQ-GAN variants using global batch 4, 20,000 updates/80K image presentations, the same MIR list, AdamW, effective learning rate 4e-4, and frozen RoSteALS weights. These controlled checkpoints differ from both A_orig and A_ref.

Across three default-model seeds, IC-all ranges are 0.58/0.57 pp on CLIC/MetFaces; most component differences are smaller. Removing external LPIPS raises recovery accuracy at a substantial perceptual cost, with only one seed for that variant. Matching the budgets removes the original optimization confound, but the results do not establish that every internal component is necessary or every variant has converged.

Clarification on VQ-GAN Adapter Checkpoints (A_orig and A_ref)

We would like to clarify the training configurations of the two VQ-GAN adapter checkpoints referenced in our responses.

Both checkpoints were trained for 20,000 optimization steps using AdamW (weight decay 1e-4, betas 0.9/0.999). However, A_orig used a global batch size of 8 and learning rate of 8e-4 (160K image presentations), whereas A_ref used a global batch size of 4 and learning rate of 4e-4 (80K image presentations).

A_orig was used for the latest cross-system benchmark, while A_ref was used for the separate VQ-GAN module-free controls, random-message evaluations, and quality-matched analyses.

These checkpoints are not training-seed replicates or equal-budget runs. All R0/R2/R5 comparisons within the A_ref control study share the same frozen RoSteALS backbone and evaluation conditions. We do not combine results across these checkpoints to estimate the contribution of the learned module.

This clarification does not change any reported experimental results or conclusions.

Reviewer_LveyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces FROST, a post-hoc framework designed to simultaneously improve image fidelity and message recovery robustness for pretrained autoencoder-based image watermarking systems (specifically RoSteALS) without modifying any pretrained model weights.

Critical strengths

(+) Practical and Non-Invasive Adaptation Formulation: Freezing the entire watermarking pipeline (image encoder/decoder and message encoder/decoder) ensures backward compatibility with already-deployed message extraction endpoints while avoiding catastrophic forgetting or costly full-pipeline retraining.

(+) Highly Parameter- and Compute-Efficient Tuning: The FROST Module requires only ~316K trainable parameters and approximately 7.2 hours of training on 2× RTX 3090 GPUs (160K image presentations), which is orders of magnitude lower than training deep watermarking models from scratch (e.g., Pixel Seal or Video Seal).

Critical weaknesses

(-) In Table 1, the FROST Module alone (yMy_M) actually degrades image fidelity relative to the RoSteALS baseline (CLIC PSNR drops from 29.10 dB to 28.67 dB; LPIPS worsens from 0.0406 to 0.0570). All PSNR and SSIM gains are exclusively introduced by the non-learned, inference-time heuristics (REC and B0B_0 removal). However, the paper never evaluates applying REC and B0B_0 removal directly to the vanilla RoSteALS output without the FROST module (i.e., xF=Π[x+BσB(y−x0)]x_F = \Pi[x + \mathcal{B}_{\sigma_B}(y - x_0)]). Without this baseline, it is impossible to determine how much of the final fidelity gain is simply an artifact of subtracting the autoencoder's reconstruction error versus an effect of the learned adapter.

(-) While pixel-level metrics (PSNR, SSIM) improve, perceptual metrics show severe degradation on continuous-latent models. In Table 4 (KL-f8), the FROST Adapter degrades LPIPS from 0.0387 to 0.0487 on CLIC and causes an 85% relative increase in perceptual distortion on MetFaces (0.0410 to 0.0762). The authors acknowledge "purple-green color artifacts in low-texture regions." For invisible watermarking, human perceptual distortion and chromatic artifacts are critical failure modes that undermine the validity of PSNR gains.

(-) In Section 4.1.2 and Table 3, the default configuration (EXP-A) was trained with a global batch size of 8 (160K cumulative image presentations across 2 GPUs), whereas the ablated variants (EXP-B, EXP-D, EXP-E) were trained with a global batch size of 4 (80K cumulative image presentations on 1 GPU). Because the optimization budget, learning dynamics, and gradient batch noise were halved for the ablation variants, the performance variations in Table 3 cannot be rigorously attributed to the individual architectural components (FiLM, smoothing, or internal LPIPS).

(-) To perform Reconstruction-Error Correction (REC), the embedding pipeline requires computing x0=G(z)=G(E(x))x_0 = G(z) = G(E(x)). This introduces an additional full forward pass through both the heavy image encoder EE and image decoder GG purely to obtain the unwatermarked reconstruction. This effectively doubles the autoencoder compute cost during watermark embedding.

Questions for authors

Please see above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Lvey

We sincerely thank the reviewer for highlighting important concerns regarding module effectiveness, perceptual quality, and computational cost. We conducted additional experiments to provide a more comprehensive evaluation.

1. REC+B0 without the learned module

We added the requested module-free control. In the paired VQ-GAN study, CLIC-530 images at 256×256 receive independent random 100-bit messages; configurations share the image/message/attack realizations. Full uses A_ref, which differs from the A_orig performance checkpoint and the newly trained ablation checkpoints. IC-avg averages 15 corruptions at severities 3 and 5; table accuracy is rounded to one decimal place.

ConfigurationPSNR (dB)LPIPSIC-avg (%)
RoSteALS R029.170.040391.3
REC+B0, no module R234.950.012090.2
Full FROST R534.290.018692.3

Full−R2 improves IC-avg by 2.05 pp, with a 95% paired image-bootstrap CI of [1.97, 2.13], while Full−R0 is +0.98 pp [0.90, 1.06]. These intervals concern image sampling at fixed checkpoints, not training-seed variation. The control confirms that most fidelity gains come from REC and filtering; the learned module improves recovery at these operating points with worse fidelity than R2.

Quality-strength sweeps clarify this trade-off. For CLIC, Full−R2 is +1.14 pp at matched PSNR 34 dB but −0.38 pp at matched LPIPS 0.02, using interpolation over overlapping ranges. There is no consistent LPIPS-matched advantage. In a separate KL-f8 control, Full gains 5.38 pp over default R2 at severity 5, but stronger module-free controls narrow or reverse the gain; exact test-quality matching was not achieved.

2. LPIPS and color artifacts

We agree that higher PSNR alone does not establish better perceptual quality. The separate KL-f8 control uses CLIC-102/MetFaces-1,336, five random messages per image, and equal weighting of dataset means. LPIPS worsens from 0.040 for R0 to 0.058 for Full. The observed purple-green artifacts in low-texture regions remain a substantive limitation.

Color measurements support this concern: in the separate VQ-GAN A_ref study, low-frequency color error worsens relative to R0 for 85.7% of MetFaces images, and Full has larger mean low-frequency color error than R2 on both datasets. We therefore limit the claim to measured recovery–fidelity trade-offs and do not claim uniform perceptual improvement or artifact-free outputs.

These artifacts limit perceptual fidelity despite improvements in PSNR, and addressing them requires additional constraints or architectural changes beyond the present study.

3. Ablations with matched optimization budgets

We repeated the VQ-GAN ablations with global batch 4, 20,000 updates (80K image presentations), AdamW, effective learning rate 4e-4, the same MIR training list, and frozen RoSteALS weights. Evaluation uses CLIC-530/MetFaces-1,336 and two random messages per image. Controlled-A and the ablated variants are independently trained checkpoints, distinct from A_orig and A_ref.

Across 12 runs, default A has three training seeds; the internal-LPIPS and FiLM variants and both smoothing variants have two each, and the external-LPIPS variant has one. Default A's IC-avg seed ranges are 0.58/0.57 pp on CLIC/MetFaces. The mean differences for the internal-component ablations are smaller than these ranges. Removing external LPIPS increases recovery at a clear perceptual cost. Thus the new comparison removes the original budget confound, but does not establish that every component is necessary, isolate the effect of doubling the original budget, or establish convergence of every variant.

4. REC computation and measured overhead

REC reuses the already computed latent z=E(x) to obtain the unwatermarked reconstruction G(z), including the same latent realization for a stochastic encoder. It requires one extra image-decoder pass, without rerunning the encoder.

On an RTX 3090 (FP32, batch 1, 256×256), end-to-end mean latency rises from 96.9±3.4 ms for RoSteALS to 163.0±5.5 ms for Full, approximately 1.68×. Peak allocated memory, including resident models, changes from 1038.5 to 1039.2 MiB. Module-only latency is 99.9±3.4 ms; REC accounts for most of the added cost. The adapter's 315,652 trainable parameters therefore do not imply negligible full-pipeline inference overhead.

Clarification on VQ-GAN Adapter Checkpoints (A_orig and A_ref)

We would like to clarify the training configurations of the two VQ-GAN adapter checkpoints referenced in our responses.

Both checkpoints were trained for 20,000 optimization steps using AdamW (weight decay 1e-4, betas 0.9/0.999). However, A_orig used a global batch size of 8 and learning rate of 8e-4 (160K image presentations), whereas A_ref used a global batch size of 4 and learning rate of 4e-4 (80K image presentations).

A_orig was used for the latest cross-system benchmark, while A_ref was used for the separate VQ-GAN module-free controls, random-message evaluations, and quality-matched analyses.

These checkpoints are not training-seed replicates or equal-budget runs. All R0/R2/R5 comparisons within the A_ref control study share the same frozen RoSteALS backbone and evaluation conditions. We do not combine results across these checkpoints to estimate the contribution of the learned module.

This clarification does not change any reported experimental results or conclusions.

Reviewer_h6g2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FROST, a post-decoder adaptation method for invisible image watermarking. A trainable 316K-parameter module predicts a message-conditioned residual after a pretrained RoSteALS decoder, while the original image and message encoders/decoders remain frozen. The complete FROST Adapter additionally applies reconstruction-error correction and high-frequency residual suppression at inference time. On VQ-GAN-based RoSteALS, the adapter improves PSNR and average bit accuracy on CLIC and MetFaces under selected ImageNet-C-style corruptions. A second KL-f8 configuration shows similar PSNR, SSIM, and average robustness gains, although LPIPS deteriorates. The paper also reports ablations for the perceptual term, residual smoothing, and FiLM conditioning.

Critical strengths
  1. The method is simple and practically motivated: it improves a deployed watermarking system without updating the pretrained encoder and decoder. The separation between the learned module and inference-time residual processing is clearly described.
  2. The paper reports component-wise results and evaluates a second KL-f8 backbone. It also acknowledges important failure cases, including LPIPS degradation, color artifacts, and poor performance under saturation and elastic transformations.
Critical weaknesses
  1. The evaluation uses one fixed 100-bit message, generated with seed 42, for all evaluation images. Since FROST is explicitly message-conditioned, this leaves open the possibility that the reported gains depend on the particular message and does not establish robustness across payloads.
  2. The ablation evidence is not fully controlled. EXP-A uses a global batch size of 8 and 160K image presentations, whereas EXP-B, EXP-D, and EXP-E use a batch size of 4 and 80K image presentations. Consequently, the reported component differences confound the removed component with optimization budget. The study also compares only against RoSteALS variants and does not establish competitiveness with recent watermarking systems.
Questions for authors
  1. Can you report results over multiple independently sampled messages, including clean-image bit accuracy and confidence intervals?
  2. Can the component ablations be repeated with matched image presentations, batch size, and optimization budget, or can you quantify how much of the difference is attributable to training duration?
  3. How does FROST compare with recent image-watermarking methods such as TrustMark, WAM, or other robust watermarking baselines under the same datasets, payload length, and corruption protocol?
  4. What are the inference-time latency and memory costs of the FROST Module, reconstruction-error correction, and residual filtering? Does the method transfer to a new pretrained watermarking backbone without retraining the module?
2 comments
AuthorsOpenReview ↗

Response to Reviewer h6g2

We appreciate the reviewer's constructive comments on experimental reliability and fair comparisons. We expanded our evaluations and ablations to address these concerns.

1. Random messages and uncertainty

Training already sampled a random 100-bit message per sample; the fixed seed-42 message was an evaluation restriction. We now repeat independent per-image message assignments, pairing images, messages and attacks across methods. Let R0 be RoSteALS, R2 REC+filtering without the learned module, and Full the complete adapter.

With the VQ-GAN reference checkpoint A_ref, five CLIC-530 assignments and four MetFaces-1,336 assignments give mean IC-avg gains over R2 of +2.00/+2.32 pp, positive in every assignment. Clean accuracy decreases by 0.34/0.28 pp relative to R0. In the separate paired-control evaluation, Full−R2 is +2.05 [1.97, 2.13] and +2.32 [2.27, 2.37] pp. These 95% image-bootstrap CIs measure image uncertainty, not training-seed variability. Evaluation-message SDs for A_ref are 0.041/0.017 pp. A separate five-message KL-f8 control gives clean accuracy 98.85% for R0 and 97.91% for Full (−0.94 pp, CI [−1.04, −0.85]): improved corrupted-image recovery does not imply improved clean recovery.

2. Matched-budget ablations

We retrained all variants with global batch 4, 20,000 updates (80K image presentations), AdamW, effective learning rate 4e-4, the same MIR images and frozen VQ-GAN base. Evaluation uses two random messages per image on CLIC-530/MetFaces-1,336. These Controlled-A/ablation checkpoints are distinct from A_ref and the main A_orig checkpoint.

The 12 runs comprise A (default, three seeds); B (no internal LPIPS), Da (no smoothing), Db (both removed), and E (no FiLM), two seeds each; and F (no external LPIPS), one seed. Default IC-avg is 92.22/92.27%; removing FiLM gives 92.28/92.24%. The observed differences for individual internal components are smaller than the default model's across-seed ranges (0.58/0.57 pp), providing insufficient evidence to attribute the gains reliably to any single component. Removing external LPIPS gives 93.35/93.55% but worsens LPIPS to 0.057/0.087, versus 0.020/0.029 for A. Thus the new runs control the budget confound, but do not establish every component's necessity, isolate the effect of doubling the original budget, or prove convergence.

3. Recent systems and matched conditions

We re-evaluated seven systems, including TrustMark-Q, WAM, VINE-R and PixelSeal, on one server with shared evaluation code. IC-avg covers 15 corruption families at severities 3 and 5. The common ImageNet-C branch preserves the predefined 256→224→corruption→256 operator used in FROST training; removing resizing defines a different attack. Explicit JPEG and geometric tests use separate branches. Common test code does not equalize original training augmentations, payloads or output quality.

Latest CLIC PSNR (dB)/IC-avg (%) pairs are 33.6606/92.993 (FROST A_orig), 41.6211/71.791 (TrustMark-Q), 37.0171/86.019 (WAM), 35.1749/86.423 (VINE-R), and 44.1743/66.421 (PixelSeal).

A separate CLIC-530 comparison uses A_ref and VINE-R at 256 pixels, 100 bits and PSNR 35 dB. Full−VINE-R is +4.8 pp [4.5, 5.1] for IC-avg, but −2.1 pp [−2.4, −1.9] for JPEG Q50/Q70. These 95% paired image-bootstrap CIs belong to that study. This supports a protocol-specific trade-off, not superiority at matched LPIPS or across attacks. The seven-system A_orig benchmark and the A_ref controls remain separate experiments.

4. Latency, memory and transfer

REC reuses z=E(x) and computes G(z), requiring one extra image-decoder pass. RTX 3090, FP32, 256×256, batch-1 end-to-end results are:

ConfigurationLatency, mean ± SD (ms)Peak memory (MiB)
R096.9 ± 3.41038.5
Module only99.9 ± 3.41038.6
R2162.7 ± 5.11039.3
Full163.0 ± 5.51039.2

Isolated calls take 3.83 ms for the module, 92.7 ms for extra G, and 0.20 ms for filtering; these do not add to end-to-end latency. Memory includes resident models, not isolated component costs. Full costs about 1.68× baseline time.

Reliable transfer without adapter retraining is not established. Applying VQ-GAN-trained A_ref to KL-f8 changes IC-s5 by +0.69 pp [0.65, 0.73] for the module alone and −0.93 pp for Full versus that run's baseline. A same-run R2 control is absent. Positive native KL-f8 results use a separately trained adapter with the pretrained base frozen.

Under matched training budgets, the current experiments do not establish a statistically reliable contribution from FiLM, smoothing, or internal LPIPS individually. We therefore treat these components as architectural choices rather than independently validated sources of improvement.

Clarification on VQ-GAN Adapter Checkpoints (A_orig and A_ref)

We would like to clarify the training configurations of the two VQ-GAN adapter checkpoints referenced in our responses.

Both checkpoints were trained for 20,000 optimization steps using AdamW (weight decay 1e-4, betas 0.9/0.999). However, A_orig used a global batch size of 8 and learning rate of 8e-4 (160K image presentations), whereas A_ref used a global batch size of 4 and learning rate of 4e-4 (80K image presentations).

A_orig was used for the latest cross-system benchmark, while A_ref was used for the separate VQ-GAN module-free controls, random-message evaluations, and quality-matched analyses.

These checkpoints are not training-seed replicates or equal-budget runs. All R0/R2/R5 comparisons within the A_ref control study share the same frozen RoSteALS backbone and evaluation conditions. We do not combine results across these checkpoints to estimate the contribution of the learned module.

This clarification does not change any reported experimental results or conclusions.

Reviewer_ZTyHOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper presents a framework (FROST) to improve existing image watermarking approaches without retraining their encoder and/or decoders. This is demonstrated for a particular watermarking scheme "RoSteALS". The solution adds a small trained network after the watermark is embedded followed by two fixed steps: 1) add only the watermark's change on top of the original image, 2) smooths that change. The paper shows that image quality rises from 29.1 to 33.7 dB PSNR, and the average bit accuracy increases 1-2% under 15 common corruptions. Perceptual quality metric (LPIPS) does not improve consistently and gets worse.

Critical strengths
  1. Well written and motivated. Upgrading a deployed watermark without retraining it or changing its decoder is useful in production environment
  2. Simple and efficient solution
  3. Extensive ablations and trade-off discussions (LPIPS vs bit accuracy etc)
Critical weaknesses
  1. Missing the key baseline. The trained module alone lowers PSNR; the gain comes from the fixed "add the change to the original image" step, which could be applied to the baseline directly.

  2. Robustness gains might simply be a quality trade, not an actual improvement. The module trains on the same attack family it's tested on. No comparison at the same image quality.

  3. Paper studies RoSteALS watermarking approach only, whose visual quality is particularly low ( around 29 dB); Other well-known watermarking schemes such as TrustMark, WAM, Pixel Seal and InvisMark were not compared. They have higher visual quality.

  4. One particular message is used for watermarking every image. Message-conditioned module could simply learn one pattern. There is a need for training and testing with random patterns.

  5. The attack suite is narrow. No cropping, resizing, rotation or AI regeneration. Only 256x256 images are used across the board, and this may pose a limitation to the learnings.

  6. Perceptual quality loss (LPIPS) is an important finding, which brings an important question mark about the usefulness of the approach.

Questions for authors

The paper has several short-coming in its current form and is not ready for acceptance as described in the weaknesses section. The improvements below could make the paper stronger.

  1. Add the missing baseline. Apply the "add change to original" and smoothing steps to the plain baseline, without the trained module. This proves the value of the trained module itself.
  2. Establish the visual quality vs robustness tradeoff clearly, ideally with a plot (PSNR vs bit accuracy, LPIPS vs bit accuracy).
  3. Use random messages. Evaluate with a different random message per image and attack. Use more images. Higher resolution images (not just 256p). Repeat experiments with confidence intervals reported.
  4. Compare with modern watermarks. Include TrustMark, WAM, Pixel Seal or Video Seal at matched payload and quality.
  5. Expand the attacks and metrics. Add crops, resizing, rotation and regeneration, and report detection at a fixed false-positive rate.
2 comments
AuthorsOpenReview ↗

Response to Reviewer ZTyH

We sincerely thank the reviewer for the detailed feedback on module effectiveness and robustness evaluation. We conducted additional controlled experiments to clarify our contributions and limitations.

1. REC and filtering without the module

We evaluated RoSteALS (R0), REC+filtering alone (R2), and Full FROST (R5) with the same frozen base and paired images/messages/attacks. This VQ-GAN study uses the reference checkpoint A_ref, separately from the A_orig checkpoint in the latest system comparison. CLIC results are below; table accuracy is rounded.

ConfigurationPSNR (dB)LPIPSIC-all (%)
R029.170.040391.3
R2: REC+filtering34.950.012090.2
R5: Full34.290.018692.3

IC-all averages 15 corruptions at severities 3 and 5. R5−R2 is +2.05 pp on CLIC-530 (95% paired image-bootstrap CI [1.97, 2.13]) and +2.32 pp on MetFaces-1,336 [2.27, 2.37]. The learned module improves recovery at these settings; REC/filtering supply most fidelity gains. R2 has better PSNR and LPIPS than R5, so this control alone does not establish an equal-quality advantage.

2. Quality–robustness trade-off and LPIPS

Residual-strength sweeps with fixed A_ref/R2 checkpoints give a CLIC IC-all gain of +1.14 pp at matched PSNR 34 dB, but −0.38 pp at matched LPIPS 0.02. These are interpolated comparisons; small differences require caution. A separately calibrated, strengthened KL-f8 R2 gives Full−R2 IC-s5 of −0.57/+0.68 pp on CLIC/MetFaces, with imperfect test-quality matching.

In the separate KL-f8 control (equally weighted dataset means), Full improves IC-s5 by 5.38 pp over default R2 but increases LPIPS from 0.040 to 0.058 relative to R0. Color artifacts remain a limitation. We therefore claim selected recovery/fidelity trade-offs, not a uniform perceptual-quality improvement.

3. Random messages and repetitions

Original training already sampled a random 100-bit message per sample; the fixed seed-42 message was an evaluation restriction. We repeated independent per-image message assignments five times on CLIC-530 and four on MetFaces-1,336, pairing messages/attacks across methods. A_ref's IC-all gain over R2 is positive in every assignment (mean +2.00/+2.32 pp). Payloads stay fixed when attacking the same watermarked image. Message repetitions assess evaluation variability; the above bootstrap CIs assess image-sampling uncertainty, not training-seed variability.

4. Recent systems and common evaluation

We reran seven systems on one server with one evaluation codebase. Representative latest CLIC results are:

SystemPSNR (dB)IC-all (%)
FROST VQ A_orig33.660692.993
TrustMark-Q41.621171.791
WAM37.017186.019
VINE-R35.174986.423
PixelSeal44.174366.421

These operating points differ in quality and payload; they do not establish a matched-quality ranking. In a separate 256px/100-bit, PSNR-35-dB CLIC comparison, A_ref exceeds VINE-R in IC-all by 4.8 pp [4.5, 5.1], but performs worse in mean JPEG Q50/Q70 accuracy (FROST−VINE-R: −2.1 pp [−2.4, −1.9]).

The original evaluation stayed close to RoSteALS for controlled comparison. The common ImageNet-C branch retains our predefined training/evaluation operator: 256→224→corruption→256. Removing resizing changes the attack. Explicit JPEG/geometric tests are separate; pretrained systems retain their own training augmentations. Earlier controls and their CIs remain separate from the latest benchmark.

5. Additional attacks and resolution

We added crops, resizing, rotation, unseen corruptions, and regeneration. FROST remains near 50% under crop 0.75 and 15° rotation; WAM/PixelSeal are stronger in several geometric settings. On four unseen corruption families, A_ref Full−R0 averages only +0.20/+0.71 pp on CLIC/MetFaces. Under SD1.5 img2img at strength 0.2, it gains approximately +4.5/+6.5 pp over R0 at default embedding strength; VAE-only results are mixed and this comparison is not quality-matched.

For 512px (530 images) and 1024px (100 images), we evaluate the transfer of residuals generated at 256px, rather than native high-resolution watermark embedding. Full−R2 improves IC-all by +2.10/+2.12 pp, with PSNR reductions of 0.66/0.67 dB. Native high-resolution training and embedding remain untested.

6. Detection at a fixed false-positive rate

Our KL-f8 known-message verification study calibrates on 701 negatives and tests on 737 disjoint negatives. At nominal FPR 0.001, clean false positives are 3/737 and worst-attack FPR is 0.54%; attack-averaged TPR is approximately 83%. Thus the requested low empirical FPR is not validated. This is claimed-message verification, not blind detection or multiple-message search.

Clarification on VQ-GAN Adapter Checkpoints (A_orig and A_ref)

We would like to clarify the training configurations of the two VQ-GAN adapter checkpoints referenced in our responses.

Both checkpoints were trained for 20,000 optimization steps using AdamW (weight decay 1e-4, betas 0.9/0.999). However, A_orig used a global batch size of 8 and learning rate of 8e-4 (160K image presentations), whereas A_ref used a global batch size of 4 and learning rate of 4e-4 (80K image presentations).

A_orig was used for the latest cross-system benchmark, while A_ref was used for the separate VQ-GAN module-free controls, random-message evaluations, and quality-matched analyses.

These checkpoints are not training-seed replicates or equal-budget runs. All R0/R2/R5 comparisons within the A_ref control study share the same frozen RoSteALS backbone and evaluation conditions. We do not combine results across these checkpoints to estimate the contribution of the learned module.

This clarification does not change any reported experimental results or conclusions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.