acceptodds
Under review as a conference paper at ICLR 2027

EditR: Robust Image Watermarking by Learning How Generative Editing Transforms the Watermark

Abstract

Instruction-guided generative editing replaces objects or repaints scenes, often without visible traces—challenging invisible watermarking, whose value depends on recovering provenance and ownership after modification. Robust watermarking trains an embed–distort–extract pipeline against simulated attacks in a differentiable distortion layer. A generative editor, however, cannot take that place: its sampling pipeline resists stable gradient-based training. A natural remedy is to simulate the editor, yet what matters for recovery is not the edited image’s appearance but how editing transforms the watermark perturbation. We therefore identify the learning target as the editor’s response to the embedded perturbation and introduce DERS, a Differentiable Editing-Response Simulator. Trained on paired clean and watermarked edits, this differentiable surrogate predicts only this response, adds it to the clean edit—content generation inherited, not simulated—to take the editor’s place in the distortion layer. The watermark, however, is embedded at 256×256, while the editor works at native resolution on its own working grid, where naive resizing breaks the watermark’s message-bearing structure; we therefore introduce Content- and Geometry-aware Residual Transfer (CART) to adapt it to image content and the editor’s preprocessing geometry. We present EditR, an end-to-end framework integrating the two in two stages: the watermark is first trained against common distortions, and DERS is then pretrained on the paired edits and frozen. The embedder, extractor, and CART are trained against this fixed simulator—faithful to the editor’s observed response rather than drifting with the watermark it judges. On five Qwen-Image-Edit operations, EditR reaches 90.76% mean bit accuracy, an absolute gain of 12.00 percentage points over the strongest evaluated baseline, while maintaining 47.33 dB PSNR. It remains stable across the tested native resolutions and transfers without retraining, reaching 87.76% and 84.40% mean bit accuracy on JoyAI-Image-Edit and Doubao-Seedream-4.5.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Robust Image Watermarking, Instruction-Guided Generative Editing, Editing-Response Simulation

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 4 reviews, above 68% of the others.
Reviewer_ZxMzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EditR², which trains a differentiable simulator to predict how generative editing transforms watermark perturbations. The simulator provides feedback for training the watermarking network, while a residual-transfer module adapts the watermark to native-resolution images. Experiments show promising bit accuracy after instruction-guided editing.

Critical strengths
  1. Predicting the watermark-induced editing response provides a well-motivated learning target for the simulator.
  2. The paper includes component ablations and evaluations on real editors, unseen editing categories, and different image resolutions.
Critical weaknesses
  1. The decision to freeze DERS is insufficiently justified. The embedder and CART continue to change during training, potentially shifting the residual distribution away from DERS’s training data. This shift is not directly evaluated, and alternatives such as alternating updates of the simulator and watermarking network are insufficiently discussed.

  2. Complete-message recovery rates are missing. The evaluation primarily reports average bit accuracy, leaving unclear how often an entire message survives editing. This limits the assessment of practical provenance-tracing capability.

  3. Evidence of generalization across editing models is limited. Although results are reported on two additional editors, comparable baseline results are missing. The relative advantage of the method across editors therefore remains unclear.

Questions for authors
  1. How accurately does frozen DERS predict real editing responses for residuals produced during later training stages? Would periodically updating DERS improve its fidelity or final watermark recovery?
  2. What are the complete-message recovery rates after editing?
  3. How does EditR² compare with the strongest baselines on the two additional editors under the same evaluation settings?
5 comments
AuthorsOpenReview ↗

Response to Reviewer ZxMz

We sincerely thank the reviewer for the careful review of our paper and the valuable feedback.

Q1: How accurately does frozen DERS predict real editing responses for residuals produced during later training stages?

A1:

We evaluate DERS using the response-matching error (RESP) defined in Eq. (5), which measures the discrepancy between predicted and real editing responses (lower is better).

DERS is pretrained on paired edits generated by Qwen-Image-Edit, with 400 pairs per seen editing category. To evaluate its prediction fidelity after Stage-II optimization, we construct new paired editing samples using residuals generated by the final Stage-II embedder and CART, and test on 300 unseen images for each of Qwen-Image-Edit, JoyAI-Image-Edit, and Doubao-Seedream-4.5.

As shown in Table 1, the test RESP values are comparable to the training value. These results demonstrate that frozen DERS maintains good response-prediction fidelity after Stage-II optimization, including on unseen editors.

Table 1: Response-matching error (RESP ↓) of frozen DERS on training and held-out images.

Data SplitEditorRESP ↓
TrainingQwen-Image-Edit0.2333
TestQwen-Image-Edit0.2185
TestJoyAI-Image-Edit0.2166
TestDoubao-Seedream-4.50.2036

Q2: Would periodically updating DERS improve its fidelity or final watermark recovery?

A2:

In the end-to-end embed–distort–extract training framework, the distortion layer should remain independent of the watermarking task to faithfully simulate external distortions in real-world scenarios. Generative image editing, as a form of distortion, follows the same principle. In practice, generative image editing occurs independently of watermark embedding and extraction, and real editors do not adapt themselves to improve watermark recovery. Therefore, maintaining the independence of the distortion layer better reflects realistic image editing scenarios.

If DERS is jointly optimized with the embedder, extractor, and CART under the same watermark recovery objective, the distortion layer may gradually adapt to facilitate watermark extraction rather than faithfully simulate real editing distortions. In extreme cases, it might even learn to apply no effective distortion. Even with alternating updates, this risk cannot be eliminated as long as DERS is optimized under the same watermark recovery objective. Therefore, we pretrain DERS on paired real-editor responses and freeze its parameters during Stage II, while allowing gradients to propagate through DERS to optimize the embedder and CART, with the extractor jointly trained.

Q3: What are the complete-message recovery rates after editing?

A3:

The exact-message recovery rate (EMR) is less commonly used as the primary metric for multi-bit watermarking, as even a single bit error causes the entire message to be considered incorrect. In contrast, bit accuracy (ACC) provides a more informative measure of watermark robustness under distortions.

Nevertheless, we additionally evaluate EMR after generative editing, following the same evaluation protocol as in the main paper. As shown in Table 2, EditR² achieves the highest EMR, demonstrating its relative advantage in complete-message recovery.

Table 2: Exact-message recovery rate (EMR, %) after generative editing.

MetricEditR²InvisMarkPixelSealChunkySealTrustMark-QVideoSealVINE-RRobust-WideMBRSDWT-DCT
EMR (%) ↑7.000.600.000.000.600.402.601.200.000.00

Q4: How does EditR² compare with the strongest baselines on the two additional editors under the same evaluation settings?

A4:

To ensure a fair comparison, we evaluate EditR² and the baselines on JoyAI-Image-Edit and Doubao-Seedream-4.5, following the same evaluation protocol and settings as in the main paper.

As shown in Table 3, EditR² achieves the highest mean bit accuracy on both editors, outperforming the strongest baseline, Robust-Wide, by 4.86 and 5.55 percentage points, respectively. These results further demonstrate the effectiveness of EditR² and its cross-editor generalization without retraining.

Table 3: Cross-editor comparison under identical evaluation settings. Mean bit accuracy (ACC, %) ↑.

MethodJoyAI-Image-EditDoubao-Seedream-4.5
EditR²87.7684.40
InvisMark67.7361.33
VINE56.8655.23
TrustMark-Q73.7771.90
VideoSeal80.0075.26
Robust-Wide82.90‾\underline{82.90}78.85‾\underline{78.85}
MBRS59.2556.32
DWT-DCT50.4350.33

Thank you for the detailed response. However, regarding Q2, we are discussing an alternating-update scheme with two stages:

  • DERS update: The watermark embedder is frozen, and DERS is trained to fit the changes caused by watermark embedding. Watermark extraction accuracy is not used as its loss.
  • Watermark training: DERS is frozen, while the watermarking network is updated.

This procedure is similar to alternating optimization in GANs. I do not see a risk of information leakage in this scheme.

Thank you for the clarification. Our previous response conflated your proposal with a different update rule in which DERS itself is optimized using the watermark-recovery objective. Our concern in that different setting was that DERS could adapt to facilitate extraction rather than faithfully model real editing responses; we incorrectly carried this concern over to your scheme. In your scheme, the watermark embedder is frozen while DERS is updated using real paired-edit responses, without using watermark extraction accuracy as its training objective. DERS is then frozen while the watermarking network is updated. We therefore agree that the response-supervised alternating scheme you describe does not inherently involve information leakage.

Our original motivation for freezing DERS was to maintain a fixed differentiable surrogate throughout watermark training, avoiding repeated calls to the black-box editor to generate paired responses for recalibration. This was a design choice to simplify the optimization process, rather than an assertion that response-supervised alternating updates are problematic.

To directly examine this alternative, we followed the two-stage alternating-update procedure you described for three rounds. We evaluated all four settings on the same held-out test set using the same evaluation protocol. As shown in the table below, periodic recalibration reduced the response-matching error (RESP) from 0.2185 for the frozen-DERS baseline to 0.2055–0.2061, corresponding to reductions of approximately 5.7%–5.9%. Meanwhile, Bit ACC remained close to the baseline value of 90.76%, ranging from 90.73% to 90.94% across the three rounds.

Evaluation of periodic response-supervised DERS recalibration.

SettingBit ACC (%) ↑RESP ↓
Frozen-DERS baseline90.760.2185
DERS-190.730.2055
DERS-290.940.2061
DERS-390.870.2059

DERS-1, DERS-2, and DERS-3 denote the results after the first, second, and third rounds of the complete alternating-update procedure, respectively.

These results indicate that periodic recalibration improves DERS’s response-prediction fidelity on unseen samples, but does not necessarily lead to improved final watermark recovery: the lower RESP did not yield a consistent gain in Bit ACC in our experiments.

Given this trade-off and the additional black-box editor calls required to generate paired responses, we retain frozen DERS as a simple default in our current framework, while recognizing periodic recalibration as a reasonable alternative when further improving response-prediction fidelity justifies the additional editor calls.

We will revise the manuscript to clarify the distinction between the two update strategies, explain the motivation for freezing DERS in our current framework, and report the results of this alternating-update evaluation.

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

This paper proposes EditR2, a framework for robust invisible image watermarking against instruction-guided generative editing. It introduces DERS, a differentiable simulator that learns how editing transforms watermark perturbations from paired clean and watermarked edits, and CART, which adapts watermark residuals to native image resolutions and the editor's preprocessing geometry. The method is trained in two stages, using a frozen DERS to optimize watermark embedding without repeatedly invoking the actual editor. Experiments on five editing operations show that EditR2 achieves 90.76% mean bit accuracy, outperforming the strongest evaluated baseline by 12 percentage points while maintaining high visual fidelity.

Critical strengths

1. The design of the proposed method is novel and well-motivated. The paper introduces DERS to model how generative editing transforms watermark perturbations rather than predicting the entire edited image, enabling differentiable optimization without repeatedly invoking the actual editor. Combined with CART for native-resolution adaptation, this provides a coherent solution to two important challenges in editing-robust watermarking. 2. Strong empirical performance. EditR2 achieves 90.76% mean bit accuracy, outperforming the strongest evaluated baseline by 12 percentage points. The results across multiple editing operations, image resolutions, and additional editing models provide encouraging evidence of the method's effectiveness and generalization potential.

Critical weaknesses
  1. Potential surrogate mismatch during joint optimization. DERS is pretrained using residuals generated by the Stage-I embedder and then frozen while the embedder and CART are jointly optimized in Stage II. This creates a potential distribution shift between the residuals used to train DERS and those produced by the optimized model. The embedder may exploit inaccuracies in the frozen simulator, yielding improvements under simulated editing that do not fully transfer to the actual editor. Although the reported real-editor results are encouraging, the paper does not sufficiently validate the reliability of DERS on residuals produced by the final model, leaving a key assumption of the proposed training strategy insufficiently supported.
  2. Limited isolation of the contribution of the proposed design. EditR2 uses paired edits from the target editor to train DERS and additional data to optimize the watermark embedder and CART, whereas the baselines are evaluated using their officially released pretrained models. Although this comparison demonstrates strong practical performance, it does not clearly isolate the benefits of the proposed design from those of additional target-editor supervision and model adaptation. Consequently, the extent to which the reported gains arise from the proposed components rather than the additional supervision remains unclear.
Questions for authors
  1. Can the authors evaluate DERS on held-out paired edits using residuals generated by the final Stage-II embedder and CART, and quantify the discrepancy between its predicted responses and those of the actual editor? This would help establish whether the frozen simulator remains reliable after joint optimization.
  2. Were the baselines evaluated under comparable target-editor supervision and training-data budgets? Could the authors provide controlled comparisons or ablations to distinguish the gains from the proposed components from those attributable to additional target-editor supervision?
1 comment
AuthorsOpenReview ↗

Response to Reviewer YShS

We sincerely thank the reviewer for the careful review of our paper and the valuable feedback.

Q1: Can the authors evaluate DERS on held-out paired edits using residuals generated by the final Stage-II embedder and CART, and quantify the discrepancy between its predicted responses and those of the actual editor? This would help establish whether the frozen simulator remains reliable after joint optimization.

A1:

To evaluate the reliability of frozen DERS after Stage-II optimization, we use the response-matching error (RESP) defined in Eq. (5), which measures the discrepancy between predicted and real editing responses (lower is better).

DERS is pretrained on paired edits generated by Qwen-Image-Edit, with 400 pairs per seen editing category. To further evaluate its prediction fidelity, we construct new paired editing samples using watermark residuals generated by the final Stage-II embedder and CART, and evaluate DERS on 300 unseen images for each of Qwen-Image-Edit, JoyAI-Image-Edit, and Doubao-Seedream-4.5.

As shown in Table 1, the training RESP is 0.2333, while the test RESP values are 0.2185, 0.2166, and 0.2036, respectively. These results indicate that frozen DERS maintains good response-prediction fidelity for residuals generated by the final model. Furthermore, EditR² achieves mean bit accuracies of 90.76%, 87.76%, and 84.40% under the three real editors, respectively, demonstrating that the robustness improvements also hold under actual generative editing.

Table 1: Response-matching error (RESP ↓) of frozen DERS on training and held-out images.

Data SplitEditorRESP ↓
TrainingQwen-Image-Edit0.2333
TestQwen-Image-Edit0.2185
TestJoyAI-Image-Edit0.2166
TestDoubao-Seedream-4.50.2036

Q2: Were the baselines evaluated under comparable target-editor supervision and training-data budgets? Could the authors provide controlled comparisons or ablations to distinguish the gains from the proposed components from those attributable to additional target-editor supervision?

A2:

In the main experiments, we evaluate the baselines using their officially released pretrained models, following their respective training protocols. Therefore, the baselines do not necessarily have identical target-editor supervision or training-data budgets.

To provide a more controlled comparison, we follow the official MBRS implementation and training protocol. Specifically, we construct paired editing data using MBRS-watermarked images, following the same data budget as EditR² (400 pairs per seen editing category). We independently pretrain an MBRS-specific DERS and incorporate it into the MBRS distortion layer. Under comparable image quality and identical evaluation settings, we evaluate robustness against Qwen-Image-Edit.

As shown in Table 2, incorporating DERS improves MBRS's mean bit accuracy from 58.11% to 72.13%, demonstrating the effectiveness of introducing DERS as a differentiable distortion layer for watermark training. EditR² further achieves 90.76%. Together with the component ablations in Table 5 of the main paper, these results indicate that the improvements arise not only from additional target-editor supervision but also from the proposed framework.

Table 2: Controlled comparison under Qwen-Image-Edit.

MethodDERSMean ACC (%) ↑
MBRS✗58.11
MBRS + DERS✓72.13
EditR²✓90.76
Reviewer_tyhaOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces EditR2, a watermark recovery method designed for generative image editing. It learns how editors transform watermark signals and restores them without disrupting the requested edits. Experiments show strong recovery accuracy, high visual quality, and generalization across editing operations, resolutions, and different editors.

Critical strengths
  • This paper is targeting a timely and important problem of recovering the watermarking even through the series of instruction-based editing.
  • It demonstrates a end-to-end system framework with two major techniques that can handle the recovering.
Critical weaknesses
  • The paper writing is poor and very hard to follow.
  • It is unclear about the new contribution of this paper compared to those existing solutions.
  • The evaluation is limited to one type of model.
Questions for authors

Thanks for your submission to ICLR'2027, below are detailed comments.

  • The evaluation is limited to one type of editing model Qwen-Image-Edit, will this proposed framework be generalized towards other similar models, such as Step1X-Edit (https://github.com/stepfun-ai/Step1X-Edit) and OmniGen2 (https://github.com/VectorSpaceLab/OmniGen2)

  • For extra large image resolution like 4K or 8K, will increase the number of GPUs (even though currently using one RTX Pro 6000) will help with the speed of such processing? i.e., does the current framework support multi-GPU/server-nodes scaling?

  • The paper is very hard to read and follow, as there are too many jargons jumping out at the same time in both abstract and introduction without giving sufficient context or background. This makes it hard to identify the key problem of this paper are targeting at. While at the technical sections, the paper is too soon to dive into the details with too many equation and symbols without a high-level design walkthrough.

  • What is the major reasons for relatively inferior performance of the proposed solutions for common distortions other the first three columns in Table 4?

  • I notice that the demonstrated examples of Figures are relatively simple in terms of its foreground and background elements, I am wondering when such assumption change with more diverse types of elements/objects in its foreground and background, will the currently proposed solution would still remain effective?

4 comments
AuthorsOpenReview ↗

Response to Reviewer tyha

We sincerely thank the reviewer for the careful review of our paper and the valuable feedback.

Q1: Can EditR² generalize to other generative editing models, such as Step1X-Edit and OmniGen2?

A1:

The main paper already evaluates Qwen-Image-Edit, JoyAI-Image-Edit, and Doubao-Seedream-4.5. We further test EditR² on Step1X-Edit v1p2 and OmniGen2 without retraining, using the same 300 images and prompts. Table 1 shows high accuracy across five edit types, further supporting cross-editor generalization.

Table 1: Cross-editor bit accuracy (ACC, %) without retraining.

Editing ModelObject AdditionObject ReplacementObject RemovalBackground ReplacementAppearance RecoloringMean
Step1X-Edit v1p297.8396.0292.3387.8299.7394.75
OmniGen282.9384.7286.0386.4279.3383.89
Qwen-Image-Edit-251194.2490.1790.1385.3893.8790.76
JoyAI-Image-Edit95.5588.9086.1778.7089.4787.76
Doubao-Seedream-4.591.5683.8379.1175.7291.7884.40

Q2: Can EditR² efficiently process 4K/8K images and scale to multiple GPUs or server nodes?

A2:

We measure 4K/8K efficiency using one or two NVIDIA RTX PRO 6000 GPUs, processing images in parallel. Table 2 shows that two GPUs nearly double throughput at both resolutions.

Table 2: Processing efficiency at 4K and 8K.

ResolutionGPUsEmbedding (ms)Extraction (ms)Editing (s)Throughput (img/s) ↑
4K (3840×2160)1153.1010.1139.020.0256
4K (3840×2160)2114.836.0737.930.0527
8K (7680×4320)1440.196.9937.910.0264
8K (7680×4320)2330.334.2036.850.0543

Q3: Could the authors clarify the core problem, main contributions, and overall framework design before introducing the technical details?

A3:

The key challenge is integrating large-scale generative editing into end-to-end watermark training: the distortion layer must be differentiable, while large editors are costly and offer limited gradient access.

DERS learns the editor's response to watermark perturbations from paired real edits. Its Clean-Edit bypass reuses the precomputed clean edit, predicting only the watermark-induced response rather than reconstructing the full edit. CART preserves message-bearing residuals during native-resolution transfer and editor resampling by adapting them to image content and preprocessing geometry.

Together, they form the two-stage EditR² framework: Stage I trains under common distortions; Stage II optimizes through frozen DERS for editing robustness while preserving image fidelity. We will clarify this design before the technical details.

Q4: Could the authors further analyze EditR²'s robustness across different common distortions and provide a more comprehensive comparison with the baselines?

A4:

Table 4 of the main paper reports over 98% ACC for EditR² under every common distortion, though some baselines lead on individual cases. Stage I trains the embedder and extractor under common distortions; Stage II uses pretrained DERS as a differentiable distortion layer for generative editing.

We add five distortions (erasing, noise, blur, compression, and color). Table 3 shows the highest mean ACC across 15 conditions. Together with 90.76% under Qwen-Image-Edit, this supports robustness to both distortion types.

Table 3: Common-distortion ACC (%); mean over 15 conditions (10 original and 5 additional).

MethodSun FlareSalt-and-Pepper NoiseDefocus BlurWebP CompressionRandom ErasingMean (15) ↑
DWT-DCT73.8278.9758.9352.5684.9072.83
MBRS96.8283.1363.4980.1098.9887.73
VideoSeal94.5168.6297.9493.8199.5091.81
Robust-Wide98.4399.25‾\underline{99.25}99.30‾\underline{99.30}99.04‾\underline{99.04}99.86‾\underline{99.86}92.65
InvisMark97.5698.9387.4567.1399.3096.17‾\underline{96.17}
VINE99.68‾\underline{99.68}97.1596.0499.0199.4592.73
TrustMark-Q77.2446.3697.7488.2398.0488.16
PixelSeal97.4463.0495.5289.0499.5891.45
EditR²100.0099.9899.9999.8299.9899.81

Q5: Does EditR² remain effective on images with more complex foregrounds and backgrounds, such as scenes containing multiple objects or people?

A5:

We group complex scenes (crowded shops, multiple people, and cluttered dining tables) by original image content. Table 4 shows high ACC in both groups, supporting robustness to generative editing in visually complex scenes.

Table 4: ACC (%) by scene complexity.

Original Image ContentMean ACC (%) ↑
Multi-object / Multi-person Scenes89.93
Cluttered Scenes87.22

Thanks for more detailed evidence, I have some follow-up questions.

  • Does the reported two-GPU result only increase throughput by processing separate images in parallel, or can multiple GPUs (e.g., more than 4 GPUs) accelerate a single 4K/8K image?

  • Does the framework support multi-machine (where each machine has multiple GPUs) execution?

AuthorsOpenReview ↗

Response to Reviewer tyha (Part 1/2): Multi-GPU and Multi-Machine Execution

Thank you for the follow-up questions. We appreciate the opportunity to clarify the scope of our previous multi-GPU results and provide additional measurements. We also take this opportunity to further clarify the technical contributions of EditR².

1. Multi-GPU parallelism and single-image acceleration

Our previous two-GPU results used image-level data parallelism, with different images assigned to different GPUs. Thus, the reported improvement reflected increased aggregate throughput across multiple images, rather than acceleration of a single image.

To clarify this distinction, we conducted additional measurements with four GPUs at 4K and 8K resolutions. The results are summarized below. Here, E, Editing, and X denote watermark embedding, external generative editing, and watermark extraction, respectively. The external editor is Qwen-Image-Edit.

ResolutionProcessing modeGPUsEmbedding (ms)Extraction (ms)External editing (s)E+Editing+X throughput (img/s)E+X throughput (img/s)
4KSingle image1112.246.6426.690.03718.41
4KSingle image4114.077.1828.710.03458.25
4KMultiple images in parallel4115.656.8226.780.147032.22
8KSingle image1433.566.4127.140.03552.27
8KSingle image4436.887.0729.140.03312.25
8KMultiple images in parallel4436.277.3727.310.13998.98

For single-image execution, four GPUs did not reduce EditR²’s E+X latency. At 4K, the latency was 118.88 ms on one GPU and 121.25 ms on four GPUs; at 8K, it was 439.97 ms and 443.95 ms, respectively. The single-image E+Editing+X throughput likewise did not increase. Therefore, under our tested configuration, the measured benefit of using multiple GPUs for EditR² is increased aggregate throughput through concurrent processing of independent images, rather than reduced single-image E+X latency. We evaluated configurations with up to four GPUs and do not claim measured scaling beyond this setting.

When multiple images were processed in parallel, four GPUs increased the aggregate E+X throughput from 8.41 to 32.22 img/s at 4K (3.83×) and from 2.27 to 8.98 img/s at 8K (3.96×).

When the external editing stage was also included, the measured E+Editing+X system throughput increased from 0.0371 to 0.1470 img/s at 4K (3.96×) and from 0.0355 to 0.1399 img/s at 8K (3.94×). These full-pipeline results include Qwen-Image-Edit and should be understood as system-level throughput, rather than acceleration contributed by EditR² itself.

For single-image execution, throughput is reported separately for the E+X and E+Editing+X measurement scopes and is based on the latency of the corresponding scope. For multi-image execution, throughput denotes the measured aggregate rate across all four GPUs.

We also note that the measurements in our previous response were obtained while other jobs were running on the machine, whereas the new measurements were conducted on an otherwise idle machine. We therefore base the comparisons in this response on the internally consistent measurements reported above, rather than directly comparing absolute latencies across the two rounds.

2. Multi-machine execution

We have not implemented or experimentally evaluated multi-machine execution. Therefore, we do not claim existing multi-machine support or measured multi-machine scaling. Conceptually, the image-level parallelism described above could be extended by distributing independent images across multiple machines. However, the resulting efficiency would additionally depend on inter-machine scheduling and data-transfer overheads, which remain outside the scope of our current implementation and evaluation.

AuthorsOpenReview ↗

Response to Reviewer tyha (Part 2/2): Technical Contributions and Manuscript Revisions

3. Clarification of the technical contributions

We would also like to clarify the motivations and respective roles of DERS and CART, as we recognize that the original manuscript may not have made the underlying ideas sufficiently clear.

DERS: Learning the watermark-induced editing response. Conventional distortion-based training alone cannot adequately capture the content-dependent changes introduced by generative editing. Moreover, directly predicting the full watermarked edited image requires modeling both the edited image content and the changes to the watermark.

DERS instead learns the watermark-induced editing response from paired clean and watermarked outputs generated by real editors using the same source image, editing instruction, and random seed. Since the precomputed clean edit yy already provides the edited image content, DERS predicts only the additional watermark-induced response r^\hat{r} and forms the predicted watermarked edit through the clean-edit bypass:

y^w=y+r^.\hat{y}_w=y+\hat{r}.

This decomposition avoids requiring DERS to reconstruct image content already provided by the clean edit. It therefore provides a more targeted learning objective and differentiable feedback for optimizing the embedder.

CART: Transferring watermark residuals to native resolution. CART addresses a complementary challenge. Simple resizing alone does not account for the effects of editor preprocessing on the watermark information carried by the residual. In particular, image content and the editor’s preprocessing geometry affect residual transfer, while the watermark-induced latent differences should remain consistent across canonical and native views.

CART therefore combines three components: content-aware refinement, geometry alignment, and latent-difference consistency. Together, they improve the transfer of watermark residuals from the canonical representation to native-resolution inputs.

Our ablation results support the contribution of each component. Removing content-aware refinement, geometry alignment, or latent-difference consistency reduces bit accuracy from 90.76% to 80.72%, 85.56%, or 80.24%, respectively. These results show that each component contributes to the final native-resolution performance.

Together, DERS and CART address two distinct but related challenges: learning watermark robustness to generative editing and transferring the learned watermark to native-resolution images. Under otherwise matched training settings and at comparable source-image PSNR, response prediction with CART achieves higher real-editor bit accuracy than full-edit prediction with CART in every evaluated edit category. This controlled comparison, together with the DERS–CART ablations, provides further empirical support for our design choices.

4. Planned manuscript revisions

To address the concerns about presentation and the clarity of our contributions, we will revise the abstract and introduction to state the research problem, the motivations for response learning and native-resolution residual transfer, and the respective roles of DERS and CART more explicitly. We will also add a high-level walkthrough of the framework before introducing the technical details, so that readers can understand the overall pipeline without reconstructing it from separate sections.

In addition, we will incorporate the broader evaluation results into the revised manuscript, including experiments on five generative editors—Qwen-Image-Edit, JoyAI, Doubao, Step1X-Edit v1p2, and OmniGen2—15 common-distortion conditions, and complex scenes involving multiple objects, multiple people, and cluttered backgrounds. These revisions will make both the technical contributions and the empirical basis for our claims more explicit.

Thank you again for the constructive feedback.

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

The paper builds an image watermarking method designed to resist different image manipulations. Aside from the common image transformations, a focus of the work is on instruction-based generative image editing which is particularly challenging to defend against because of the sampling process in the image editor.

Critical strengths
  • The paper has technical novelty. One of their technical contributions is to train a model to simulate how the watermark propagates through editing which helps address the challenge of optimizing over the image editor due to both stochasticity and inefficiency.

  • The method achieves strong results relative to the baselines they consider especially under editing based manipulations. The gap in Table 2 is quite significant.

Critical weaknesses
  • One of the motivations for DERS that the authors mention is computational efficiency. However, the authors didn't include any experiments regarding that.

  • There is some prior work [1] on robust watermarking that the authors didn't evaluate against which is more recent than the other baselines.

  • The authors did not evaluate on other datasets that were included in prior work such as SA-1b which is for example used in PixelSeal and [1].

References

[1] https://arxiv.org/pdf/2605.16720

Questions for authors

Did the authors consider multiple edits to the image similar to the composition of transformations from [1]?

1 comment
AuthorsOpenReview ↗

Response to Reviewer rVjU

We sincerely thank the reviewer for the careful review of our paper and the valuable feedback.

Q1: One of the motivations for DERS that the authors mention is computational efficiency. However, the authors didn't include any experiments regarding that.

A1:

The primary motivation of DERS is not computational efficiency, but to incorporate large-scale generative editing attacks into end-to-end watermark training. Direct integration is impractical because closed-source editors are black-box models that do not provide gradients, while large open-source editors incur prohibitive computational costs, making end-to-end training nearly infeasible. DERS addresses this challenge by learning the editor's response to watermark perturbations, providing a differentiable surrogate for watermark training.

Computational efficiency is an additional advantage. Existing work, such as Robust-Wide, incorporates InstructPix2Pix (IP2P) into training. IP2P is an earlier-generation image editing model with relatively limited editing quality and instruction-following capabilities compared to modern large-scale generative editors. Table 1 shows that DERS has 288 times fewer parameters than IP2P, while EditR² improves Qwen-Image-Edit mean bit accuracy over Robust-Wide by 16.38 percentage points.

Table 1: Editing component parameters and robustness under Qwen-Image-Edit.

MethodEditing ComponentParameters (M) ↓Qwen-Image-Edit Mean ACC (%) ↑
Robust-WideIP2P1,066.2574.38
EditR²DERS3.7090.76

Q2: Could the evaluation include a comparison with the recent PixelSeal+CAT [1] method and additional experiments on SA-1B, a dataset used in prior work?

A2:

We additionally evaluate EditR² against PixelSeal+CAT [1] on SA-1B. Considering the computational cost of large-scale generative editing (approximately one minute per image), and given that [1] also uses 1,000 SA-1B images for evaluation, we randomly sample 1,000 images from 100 SA-1B shards (10 images per shard). Following the same evaluation protocol as in the main paper (including evaluation metrics and attack settings), we compare EditR² with PixelSeal+CAT [1] in terms of image quality, common distortions, and Qwen-Image-Edit attacks.

Tables 2–3 show that EditR² outperforms PixelSeal+CAT [1] in PSNR and robustness to common distortions and Qwen-Image-Edit.

Table 2: Image quality (PSNR) and bit accuracy (%) under common distortions.

MethodPSNR ↑│NoneFlipCropGBJPEGGNB/CResizePosterizeRotation
PixelSeal+CAT [1]43.2002│99.0275.7171.6199.0497.3698.9397.5698.5697.0161.10
EditR²46.7990│99.8299.2999.0099.4698.0799.3999.2899.4399.4199.54

Table 3: Bit accuracy (%) under Qwen-Image-Edit.

MethodAdditionReplacementRemovalBackground ReplacementRecoloringMean
PixelSeal+CAT [1]52.0351.2651.3351.1651.8651.53
EditR²93.6789.7189.6284.7993.4390.24

Q3: Did the authors consider multiple edits to the image similar to the composition of transformations from [1]?

A3:

Our work primarily focuses on incorporating large-scale generative editing into end-to-end watermark training, whereas [1] investigates compositions of common distortions. To provide a more comprehensive evaluation, we conduct additional experiments on both settings. Both experiments follow the same evaluation scale (300 images) and attack settings as in the main paper.

For multiple generative edits, we sequentially apply Qwen-Image-Edit to the same image. Table 4 shows that EditR² maintains reasonable robustness after multiple editing rounds.

For composed common distortions, we evaluate sequential combinations of horizontal flipping with Gaussian blur, JPEG compression, and resizing. As shown in Table 5, EditR² consistently outperforms PixelSeal+CAT [1], demonstrating strong robustness to composed common distortions.

Table 4: Robustness under multiple rounds of generative editing.

Method1 Round2 Rounds3 Rounds
EditR²90.7680.6672.96

Table 5: Robustness under composed common distortions.

MethodHFlip + Gaussian Blur (k=3)HFlip + JPEG (q=60)HFlip + Resize (scale=0.7)
PixelSeal+CAT [1]99.6498.8599.64
EditR²99.8799.9399.87

References

[1] Anirudh Satheesh, Michael-Andrei Panaitescu-Liess, Andrew Xu, Georgios Milis, Heng Huang, Zikui Cai, and Furong Huang. Compositional adversarial training for robust visual watermarking. arXiv preprint arXiv:2605.16720, 2026. https://arxiv.org/abs/2605.16720.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.