acceptodds
Under review as a conference paper at ICLR 2027

GuardFlow: Image Immunization via Trajectory Disruption against Flow-Based Image Editing

Abstract

Text-guided image editing models enable powerful manipulation of visual content but also introduce serious security risks by allowing unauthorized or malicious edits. Image immunization aims to protect images by adding imperceptible perturbations that induce edit failure. However, existing methods are primarily designed for GAN or diffusion models and do not generalize to flow-based generative systems. In this work, we propose GuardFlow, the first immunization framework that guards images from flow-based editing models. Our key insight is to exploit the trajectory-driven nature of flow models, where small perturbations accumulate and alter the generation path. We further derive a theoretical perturbation bound that ensures effective disruption while preserving perceptual quality. Experiments on major flow-based editing frameworks show that GuardFlow effectively impedes text-guided editing while maintaining high visual fidelity, outperforming existing baselines in both effectiveness and robustness. The code will be released upon publication.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Immunization, Flow Matching, Text-guided Image Editing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_94irOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GuardFlow generates small image perturbations to inhibit flow-based editing through velocity-field disruption, terminal regression, and an image-specific perturbation lower bound. The experiments demonstrate competitive protection against the targeted models, while its practical value across unseen models and methodological novelty require further validation.

Critical strengths
  1. Competitive protection against targeted flow editors. GuardFlow consistently outperforms the evaluated baselines on SD3 and FLUX while maintaining comparable input visual quality. For example, on SD3 inpainting, it reduces CLIP-T from the strongest baseline’s 25.46 to 22.89, providing meaningful empirical evidence for addressing an important image-protection problem.

  2. Efficient perturbation generation. Once trained, the immunizer generates protection in a single forward pass, with a reported runtime of approximately 0.01 seconds per image. This substantially improves scalability over per-image iterative optimization and supports potential use in batch image protection.

Critical weaknesses
  1. Practical protection against unknown editors remains insufficiently established. The main results evaluate protection against the backbone used to train the immunizer, whereas deployment requires protection against editors unknown at publication time. Although Table 8 demonstrates some transferability, performance weakens across different latent representations, and transfer to SD3.5 uses a shared VAE; consequently, the evidence does not yet establish broadly applicable protection.

  2. The methodological contribution beyond adapting existing techniques is unclear. The velocity-residual objective, SDS/RFDS-inspired update, terminal reconstruction constraint, and global immunizer closely relate to established approaches. Existing comparisons do not sufficiently isolate the proposed flow-specific design from architecture and training differences, leaving uncertainty about whether the gains constitute a substantive methodological advance.

Questions for authors
  1. Clarification and targeted validation of transferability: How does GuardFlow compare with the strongest applicable baseline under the same SD3→FLUX transfer setting, with both methods generating perturbations using only SD3? A limited comparison on the existing validation data would materially affect my assessment of whether GuardFlow offers a distinctive benefit against unknown editors.

  2. Targeted validation of methodological novelty: Under the same immunizer architecture, training data, and optimization budget, how does GuardFlow compare with a directly adapted SDS/RFDS objective? Please also clarify whether the image-specific bound provides a meaningful additional benefit over removing LBL_B or using a tuned fixed bound, since these comparisons would determine which proposed components substantiate the claimed advance.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 94ir (1/3)

We appreciate the reviewer’s thoughtful assessment and constructive comments, which have helped us clarify the contribution of each objective and strengthen our evaluation of transferability to unseen editors. We hope these responses address the reviewer's concerns. We welcome any further questions or discussion.


W1, Q1) Regarding SD3.0 → FLUX Cross Transferability

To directly assess transferability from SD3.0 to FLUX, we compare GuardFlow with all evaluated baselines, including SDS and DiffVax, the strongest competitors in our main experiments (Tab. D). All perturbations are generated using only SD3.0, without access to or adaptation on FLUX. (All methods follow the same perturbation budget and editing settings as in Tab. 1.)

Table D. Cross-model transferability from SD3.0 to FLUX (inpainting). CLIP-T reduction is measured relative to clean editing (higher reduction indicates stronger editing disruption).

Source → TargetMethodCLIP-T ↓CLIP-T Reduction ↑
SD3.0 → FLUXClean25.64—
Random25.050.59
EditShield25.020.62
DiffusionGuard25.190.45
SDS25.090.55
PhotoGuard25.120.52
DiffVax24.830.81
Ours22.633.01

The results in Tab. D show that GuardFlow achieves the largest CLIP-T reduction among all evaluated methods when transferred to FLUX. The observed transferability may be attributed to the disruption of intermediate states in the editing process, rather than the final output of a specific editing model. Additional cross-transfer results on Qwen-Image and diffusion-based editors are provided in Tab. B in our response to Reviewer B9Gc, further supporting GuardFlow's effectiveness across different model architectures and generative paradigms.

We acknowledge that achieving model-agnostic immunization remains challenging as generative models continue to evolve. While GuardFlow demonstrates effective transfer across the evaluated editors, its effectiveness cannot be guaranteed for arbitrary unseen or future models. Developing global immunizers with stronger cross-model generalization remains an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer 94ir (2/3)

W2, Q2) Regarding GuardFlow Objective Contribution and Design

While our approach builds upon established optimization techniques, its objectives are specifically formulated to address the trajectory dynamics of flow-based editing. Together, the three objectives address trajectory disruption, output-level guidance, and image-specific perturbation bounds. To isolate their contributions from architectural and training differences, we conducted additional controlled ablation studies using the same UNet++ immunizer, training data, optimization settings, and perturbation budget, varying only the training objectives (Tab. E). We elaborate on these design choices and their empirical contributions below.

Table E: Controlled comparison of objectives and lower-bound designs (SD3.0, CCP inpainting). All immunizers share the same UNet++ architecture, training data, number of epochs, learning rate, and ℓ∞ budget β = 16/255; only the training objective differs.

Training objectiveCLIP-T ↓SSIM(Noise) ↑∥δ∥2/d\Vert\delta\Vert_2/\sqrt{d}
No immunization28.46––
LRFDS+LT+LB\mathcal{L}_{\mathrm{RFDS}}+\mathcal{L}_T+\mathcal{L}_B25.400.9094.36/255
Ltraj+LT\mathcal{L}_{\mathrm{traj}}+\mathcal{L}_T (w/o LB\mathcal{L}_B)24.240.9283.53/255
Ltraj+LT+LBfixed\mathcal{L}_{\mathrm{traj}}+\mathcal{L}_T+\mathcal{L}_B^{\mathrm{fixed}} (ϵ=2/255\epsilon=2/255)24.160.9054.15/255
Ltraj+LT+LBfixed\mathcal{L}_{\mathrm{traj}}+\mathcal{L}_T+\mathcal{L}_B^{\mathrm{fixed}} (ϵ=4/255\epsilon=4/255)23.840.8765.55/255
Ltraj+LT+LBfixed\mathcal{L}_{\mathrm{traj}}+\mathcal{L}_T+\mathcal{L}_B^{\mathrm{fixed}} (ϵ=6/255\epsilon=6/255)24.390.8286.08/255
Ltraj+LT+LBfixed\mathcal{L}_{\mathrm{traj}}+\mathcal{L}_T+\mathcal{L}_B^{\mathrm{fixed}} (ϵ=8/255\epsilon=8/255)24.230.7338.61/255
LT+LB (w/o Ltraj)\mathcal{L}_T+\mathcal{L}_B\ (\text{w/o}\ \mathcal{L}_{\text{traj}})26.080.9094.38/255
Ltraj+∥δ∥22+LB\mathcal{L}_{\mathrm{traj}}+\Vert\delta\Vert_2^2+\mathcal{L}_B24.410.8894.82/255
Ltraj+LT+LB\mathcal{L}_{\mathrm{traj}}+\mathcal{L}_T+\mathcal{L}_B (Ours)22.890.9004.58/255

(1) Trajectory disruption objective (Ltraj\mathcal{L}_{\mathrm{traj}}).

  • Unlike RFDS, which measures velocity residuals against the perturbed image's own trajectory, Ltraj\mathcal{L}_{\mathrm{traj}} uses the fixed clean-source trajectory as its reference. This removes the direct dependence of the reference trajectory on the perturbation. Under a first-order expansion of vθv_\theta around the clean trajectory (Appendix A), the perturbation changes the Ltraj\mathcal{L}_{\mathrm{traj}} residual by tJvθδtJ_{v_\theta}\delta and the RFDS residual by (tJvθ−I)δ(tJ_{v_\theta} - I)\delta. Hence, when Jvθ≈IJ_{v_\theta} \approx I, the shift of the reference largely cancels the change in the predicted velocity, reducing the RFDS term to −(1−t)δ-(1-t)\delta, whereas Ltraj\mathcal{L}_{\mathrm{traj}} is free of this cancellation.

  • Accordingly, replacing Ltraj\mathcal{L}_{\mathrm{traj}} with LRFDS\mathcal{L}_{\mathrm{RFDS}} increases CLIP-T from 22.89 to 25.40 under the same training conditions and a similar perturbation magnitude (Tab. E).

  • In addition, while the Jacobian-free update follows SDS/RFDS, their original formulations do not characterize whether discarding the Jacobian preserves the optimization direction. Appendix B derives this condition for rectified flow (Eqs. (39) and (54)) and verifies it on the pre-trained SD3 vector field, where the exact and Jacobian-free directions are positively aligned for 99.2% of samples (Table 5).

(2) Terminal objective (LT\mathcal{L}_T).

  • Unlike conventional output-level objectives such as PhotoGuard-D, which push the edited output toward a fixed target by backpropagating through the full sampling trajectory, LT\mathcal{L}_T suppresses the intended editing changes by guiding the terminal estimate toward the source image. The one-step terminal estimate in Eq. (12) provides this output-level signal from a single timestep, which substantially reduces peak GPU memory (Tab. 1).

  • LT\mathcal{L}_T also differs from a conventional magnitude penalty, as the regularization it induces on the perturbation is direction-dependent. As shown in Appendix C, expanding LT\mathcal{L}_T in δ\delta yields a linear term that pulls the predicted edit toward the source and an anisotropic quadratic penalty δ⊤Aδ\delta^\top A\delta, where A=Et[Mt⊤Mt]A = \mathbb{E}_t[M_t^\top M_t] and Mt=t (I+(1−t)Jvθcfg)M_t = t\,(I + (1-t)J_{v^{\mathrm{cfg}}_\theta}) weights each perturbation direction by how strongly it changes the predicted edit.

  • Replacing LT\mathcal{L}_T with a simple magnitude penalty ∥∥δ∥∥22\|\|\delta\|\|_2^2 weakens editing disruption (CLIP-T 24.41 vs. 22.89) despite a larger perturbation, and reduces source preservation within the regenerated region (Tab. E; Tab. C in our response to Reviewer B9Gc). This supports the contribution of source-directed terminal guidance beyond magnitude regularization.

AuthorsOpenReview ↗

Response to Reviewer 94ir (3/3)

(3) Image-specific lower bound (LB\mathcal{L}_B).

  • The perturbation magnitude required for effective disruption varies substantially across images. On PIE-Bench++, our estimate ϵ∗\epsilon^* ranges from 0.77/255 to 13.44/255, with correlations to image structure such as Canny edge density (Fig. 5; Appendix F.1). This is consistent with the image-dependent nature of effective perturbation magnitudes suggested by Yuan et al. (TPAMI 2024) [1]. Accordingly, LB\mathcal{L}_B incorporates a theoretically derived necessary lower bound on perturbation magnitude based on image-specific vector-field sensitivity.

  • The ablation results show that removing LB\mathcal{L}_B weakens editing disruption, while increasing the fixed bound (ϵ∈{2,4,6,8}/255\epsilon \in \{2, 4, 6, 8\}/255) leads to larger perturbations without monotonic improvements in editing disruption. Note that ϵ=4/255\epsilon = 4/255 is the fixed bound selected by hyperparameter search in the manuscript (Tab. 3). These results highlight the importance of image-specific perturbation bounds, as our bound achieves stronger editing disruption and higher SSIM(Noise) than the best fixed-bound LB\mathcal{L}_B, despite a smaller perturbation magnitude (Tab. E).

Altogether, these controlled comparisons demonstrate the complementary contributions of our trajectory-oriented objectives and support their effectiveness beyond architectural or training differences.


[1] Yuan et al. Adaptive perturbation for adversarial attack. IEEE Transactions on Pattern Analysis and Machine Intelligence. (2024)

Reviewer_B9GcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GuardFlow, an image protection framework designed to hinder unauthorized editing by flow-based generative models. GuardFlow trains an image-conditioned immunizer to generate small perturbations using a joint objective that disrupts the predicted velocity field while encouraging an estimate of the final edited output to remain close to the original image. The method incorporates an image-specific perturbation lower bound into its regularization and applies a Jacobian-free approximation to reduce the training cost of the trajectory-disruption term. After training, the immunizer protects unseen images in a single forward pass. Experiments demonstrate stronger edit-disruption performance than the evaluated baselines and robustness against various image transformations and purification methods.

Critical strengths
  1. The paper connects mathematical analysis of flow dynamics with a concrete protection objective. It derives an image-specific perturbation lower bound based on local vector-field sensitivity under a first-order approximation and incorporates this estimate into the training objective. The ablation studies support the benefits of image-adaptive lower-bound regularization and the complementary roles of trajectory disruption and terminal regression.

  2. Unlike methods requiring iterative optimization for each image, the trained immunizer generates image-dependent perturbations in a single forward pass. Although training incurs an initial offline computational cost, this feed-forward design supports efficient protection of large collections of images.

  3. The evaluation covers both inpainting and instruction-based editing. The authors test robustness against JPEG compression, rotation, cropping, and purification-based perturbation removal. Additional cross-dataset qualitative results and cross-model transfer experiments provide useful evidence about generalization across the evaluated conditions.

Critical weaknesses
  1. Since the owner of an image cannot control which editing model an adversary employs, cross-model transferability is important for assessing the practical scope of protection. Although the authors evaluate transfer from SD3.0 to SD3.5-Large and FLUX, the evaluation remains limited to a small set of flow-based backbones. SD3.5 shares its VAE with SD3.0, while transfer to FLUX is less consistent, particularly for inpainting. Additional transfer experiments involving distinct architectures and VAEs, as well as diffusion-based editors, would help establish the protection available when the target editor is unknown. These experiments should include baseline methods under matched perturbation budgets and editing settings, without target-specific retraining or perturbation re-optimization.

  2. The primary quantitative evaluation measures differences between clean-image and protected-image editing outputs, together with reductions in CLIP-T scores. These metrics support editing disruption but do not directly establish whether the final edited outputs preserve the original content, as encouraged by the terminal-regression objective. Visual degradation can also increase differences from clean edits and reduce semantic alignment with the prompt. Comparing the final edited outputs with the original source images and supplementing the evaluation with human-assessed instruction-following success rates would clarify whether the method preserves source content while preventing the intended edits.

Questions for authors
  1. Could the authors extend cross-model transfer evaluation to additional backbones with distinct architectures and VAEs, including diffusion-based editors, alongside baselines under matched perturbation budgets and editing settings without target-specific adaptation?

  2. Could the authors directly evaluate source-content preservation in the final edited outputs and report human-assessed edit success rates to determine whether GuardFlow preserves the original content while preventing the requested edits?

4 comments
AuthorsOpenReview ↗

Response to Reviewer B9Gc (1/2)

We sincerely thank the reviewer for their thoughtful and constructive feedback, which has helped us strengthen our evaluation of cross-model transferability and source preservation.


W1, Q1) Regarding Cross-Model Transferability to Unseen Model Architectures and VAEs

To further evaluate cross-model transferability, we extend our existing evaluation to Qwen-Image [1], which employs a distinct architecture and VAE, as well as diffusion-based editors SD1.5 [2] and SDXL [3] (Tab. B). All methods are evaluated under matched perturbation budgets and editing settings, without target-specific retraining or perturbation re-optimization.

Table B: Cross-transfer evaluation across unseen editing models. CLIP-T scores (↓) are reported for inpainting and instruction-based editing across flow-based and diffusion-based editors. Values in parentheses indicate the reduction in CLIP-T relative to clean editing (↑). Lower CLIP-T indicates stronger editing disruption. Bold indicates the best performance in each row, italics indicate the strongest baseline, and — denotes inapplicable settings.

SettingGroupTargetCleanRandomEditShieldDiffusionGuardSDSPhotoGuardDiffVaxOurs
InpaintingWhite-boxSD3.028.4628.02 (0.44)26.39 (2.07)26.15 (2.31)25.49 (2.97)26.38 (2.08)25.46 (3.00)22.89 (5.57)
InpaintingFlowFLUX25.6425.05 (0.59)25.02 (0.62)25.19 (0.45)25.09 (0.55)25.12 (0.52)24.83 (0.81)22.63 (3.01)
InpaintingFlowQwen-Image30.3530.18 (0.17)30.13 (0.22)30.19 (0.16)30.23 (0.12)30.13 (0.22)27.98 (2.37)26.29 (4.06)
InpaintingDiffusionSD1.527.1226.93 (0.19)26.84 (0.28)26.68 (0.44)26.73 (0.39)27.02 (0.10)26.06 (1.06)24.24 (2.88)
InpaintingDiffusionSDXL27.9927.65 (0.34)27.64 (0.35)27.68 (0.31)27.78 (0.21)27.58 (0.41)27.05 (0.94)25.02 (2.97)
InstructionWhite-boxSD3.036.2635.64 (0.62)34.90 (1.36)—35.40 (0.86)34.54 (1.72)—30.50 (5.76)
InstructionFlowFLUX33.5533.47 (0.08)33.46 (0.09)—33.24 (0.31)33.17 (0.38)—31.39 (2.16)
InstructionFlowQwen-Image36.5036.13 (0.37)36.13 (0.37)—36.07 (0.43)36.03 (0.47)—33.44 (3.06)
InstructionDiffusionSD1.532.1231.84 (0.28)31.88 (0.24)—31.80 (0.32)31.90 (0.22)—29.14 (2.98)
InstructionDiffusionSDXL31.5231.16 (0.36)31.32 (0.20)—31.19 (0.33)31.18 (0.34)—28.31 (3.21)

The results demonstrate GuardFlow's cross-model transferability across distinct architectures and generative paradigms. Specifically, GuardFlow achieves the largest CLIP-T reduction on Qwen-Image in both editing settings, despite its distinct architecture and VAE. Furthermore, although our objectives are formulated for flow-based editing, GuardFlow consistently outperforms the evaluated baselines across all diffusion-based settings. We have incorporated these additional cross-transfer results into the revised manuscript (Appendix E.4).

AuthorsOpenReview ↗

Response to Reviewer B9Gc (2/2)

W2, Q2) Regarding LT\mathcal{L}_T's Source Preservation Effect

LT\mathcal{L}_T is designed to discourage intended editing changes by guiding the terminal estimate toward the source image. Beyond editing disruption, this source-directed guidance can also encourage greater source-content retention in the edited outputs. To examine this effect, we provide an additional source-preservation evaluation comparing edited images with their original source images (Tab. C).

Table C: Source preservation of edited outputs. All metrics compare the final edited output with the original source image xsrcx^{\mathrm{src}}. Clean edit denotes editing the unprotected image and serves as a reference. The second row replaces LT\mathcal{L}_T with a magnitude penalty ∥δ∥22\|\delta\|_2^2, as in Tab. 3. Human evaluation of source preservation against the baselines is reported in Tab. A in our response to Reviewer RSxL.

Method∥δ∥2/d\|\delta\|_2/\sqrt{d}LPIPS ↓SSIM ↑PSNR ↑DINO ↑
Clean edit (no immunization)–0.55130.27759.350.4943
Ltraj+∥δ∥22+LB\mathcal{L}_{\mathrm{traj}}+\|\delta\|_2^2+\mathcal{L}_B4.82/2550.52450.322511.760.5443
Ltraj+LT+LB\mathcal{L}_{\mathrm{traj}}+\mathcal{L}_T+\mathcal{L}_B (Ours)4.58/2550.43070.475415.260.6556

We evaluate images edited with SD3.0 by measuring LPIPS, SSIM, PSNR, and DINO similarity against the original source image xsrcx^{\mathrm{src}} within the regenerated regions. We assess source preservation across three settings: clean-image editing, GuardFlow-immunized editing without source-directed guidance, and GuardFlow-immunized editing. The results in Tab. C show that GuardFlow achieves higher source similarity across all four metrics, consistent with the source-directed guidance provided by LT\mathcal{L}_T.

Tab. A in our response to Reviewer RSxL further reports human-assessed editing success and source-preservation vote shares, directly evaluating whether the intended edits are applied and how well the source content is retained.


[1] Wu et al. "Qwen-image technical report." arXiv preprint arXiv:2508.02324 (2025).

[2] Rombach et al. High-Resolution Image Synthesis with Latent Diffusion Models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022).

[3] Podell et al. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. International Conference on Learning Representations (ICLR 2024).

AuthorsOpenReview ↗

Follow-up on our Response

Dear Reviewer B9Gc,

Thank you again for your thoughtful feedback. In response to your comments, we (1) extended the cross-model transfer evaluation to Qwen-Image, SD1.5, and SDXL, with all baselines evaluated under matched perturbation budgets and without target-specific adaptation (Table B), and (2) directly evaluated source preservation of the final edited outputs using LPIPS, SSIM, PSNR, and DINO, together with human-assessed edit success (Tables A and C).

As the discussion period ends on November 18 (AoE), we would greatly appreciate it if you could take a look at our responses and let us know whether they address your concerns. We would be happy to provide any further clarification.

Best regards,

Authors of Submission 6063

AuthorsOpenReview ↗

Update: Full Human Evaluation Results

Dear Reviewer B9Gc,

We would like to briefly update you on the human evaluation you requested in Q2. We have now included the full results from 40 participants in the revised manuscript (App. E.7), updating the preliminary results (Table A) referenced in our earlier response. GuardFlow achieves the lowest human-assessed edit success rate (41.7% vs. 64.2–71.7% for the baselines) and the highest source-preservation vote share (58.6%), complementing the quantitative source-preservation results in Table C.

As the discussion period ends in less than 10 hours, we would be grateful if you could take a moment to review these results and let us know whether they address your concerns. If you have any further questions, we would be happy to clarify.

Best regards,

Authors of Submission 6063

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

This paper proposes GuardFlow, an image immunization framework designed to protect images against unauthorized editing by flow-based generative models. The proposed method combines a trajectory disruption objective with a terminal regression objective that encourages edited outputs to resemble the original images. Experiments on SD3 and FLUX demonstrate that GuardFlow outperforms existing immunization methods in terms of editing disruption, imperceptibility, and efficiency.

Critical strengths
  • The paper addresses a timely and practically meaningful problem of protecting images against unauthorized editing by flow-based generative models.

  • The experimental evaluation is comprehensive, covering multiple editing settings, model backbones, and robustness experiments.

  • The proposed method consistently achieves strong performance compared with the reported baselines.

Critical weaknesses
  • The paper lacks a human evaluation of immunization effectiveness and perceptual quality. While automatic metrics demonstrate strong performance, a human preference study would provide more convincing evidence that GuardFlow prevents intended edits while preserving image quality.
Questions for authors
  • Could the authors conduct a human evaluation to assess both editing failure and the perceptual quality of immunized images?
4 comments
AuthorsOpenReview ↗

Response to Reviewer RSxL

We sincerely thank the reviewer for their valuable feedback and careful evaluation of our work. We appreciate the reviewer’s insightful comments, which have helped us strengthen the empirical support for our method, particularly through the additional human evaluation.

Regarding human evaluation.

We conducted a blinded user study with 27 participants on 67 randomly sampled image-prompt pairs. For each baseline, participants rated the edited results in terms of prompt alignment on a 1–5 scale and selected the result that best preserves the source image among the four immunization methods. To assess imperceptibility, participants were further asked to identify the perturbed image from a pair of clean and immunized images.

Each metric is defined as follows: (1) Edit Success, the proportion of edited results rated 3 or higher on the 1–5 prompt-alignment scale; (2) Src. Preservation, the vote share of each method when participants select the one edited result that best preserves the source image among the four immunization methods; and (3) Perceptibility, the accuracy of identifying the immunized image from a pair of clean and immunized images. Lower is better for Edit Success and higher is better for Src. Preservation, whereas Perceptibility closer to 50% (chance level) indicates a less perceptible perturbation.

Table A. Human evaluation of editing disruption, source preservation, and imperceptibility.

MethodEdit Success (%) ↓Src. Preservation (%) ↑Perceptibility (%) →50
Clean76.9——
DiffVax68.517.060.0
SDS71.312.757.8
PhotoGuard75.010.861.5
Ours45.459.654.1

As shown in Tab. A, GuardFlow achieves the lowest edit success rate (45.4%) and the highest source-preservation vote share (59.6%), substantially outperforming the evaluated baselines. Meanwhile, its perceptibility score (54.1%) remains close to chance level, indicating that the perturbations are difficult for participants to distinguish from clean images. These findings complement our quantitative results with human evaluations of editing disruption and visual imperceptibility.

We have included the detailed evaluation protocol, survey layout, and questionnaire in the revised manuscript (Appendix E.7). As data collection is ongoing, we will report the results with the full set of participants, together with confidence intervals and inter-rater agreement, in the final version.

AuthorsOpenReview ↗

Follow-up on our Response

Dear Reviewer RSxL,

Thank you again for your constructive feedback. Following your suggestion, we conducted a blinded human evaluation of editing success, source preservation, and perceptibility (Table A in our response). GuardFlow achieves the lowest edit success rate (45.4% vs. 68.5–75.0% for the baselines) and the highest source-preservation vote share (59.6%), while its perceptibility remains close to chance level (54.1%).

If our response has addressed your concern, we would be grateful if you could consider reflecting this in your final assessment. We would be happy to provide any further clarification before the discussion period ends on November 18 (AoE).

Best regards,

Authors of Submission 6063

I thank the authors for their response. The human evaluation results address my concerns, showing the lowest edit success rate and perceptibility choice among the baselines. I will raise my score from 3 to 4, with the expectation that the authors will include the human evaluation results in the revised manuscript (which currently only includes the user study interface and the questions used), along with the performance scores for the other human eval. questions described in the Appendix.

Thank you for your thoughtful follow-up and for raising your score. As requested, we have updated the revised manuscript to include the full human evaluation results in Appendix E.7. Table 12 now reports the results for all evaluated questions, including Visual Quality, together with 95% confidence intervals, and Fig. 9 reports the pairwise comparison results.

These results are based on the full set of 40 participants, updated from the 27 participants available at the time of our initial response. The trends remain consistent: GuardFlow achieves the lowest edit success rate (41.7% vs. 64.2–71.7% for the baselines) and the highest source-preservation vote share (58.6%), while its perceptibility remains close to chance (55.3%).

We are grateful for your careful review and for taking the time to revisit our responses.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank all reviewers for their careful evaluation and constructive feedback. We appreciate the positive assessment of GuardFlow's practical significance, methodological contributions, and empirical effectiveness. In particular, the reviewers highlighted the following strengths:


Practical Relevance and Methodology:

  • “a timely and practically meaningful problem” (Reviewer RSxL)
  • “connects mathematical analysis of flow dynamics with a concrete protection objective” (Reviewer B9Gc)

Empirical Evaluation:

  • “The experimental evaluation is comprehensive” (Reviewer RSxL)
  • “consistently achieves strong performance compared with the reported baselines” (Reviewer RSxL)
  • “Competitive protection against targeted flow editors” (Reviewer 94ir)
  • “provide useful evidence about generalization across the evaluated conditions” (Reviewer B9Gc)

Efficiency and Scalability:

  • “supports efficient protection of large collections of images” (Reviewer B9Gc)
  • “substantially improves scalability over per-image iterative optimization” (Reviewer 94ir)

We also carefully considered the reviewers' concerns regarding human evaluation, cross-model transferability, source preservation, and methodological novelty. To address these concerns, we conducted additional experiments and analyses, summarized below.

ConcernResponseEvidence
Human evaluation (RSxL, B9Gc)Conducted a blinded human evaluation of editing success, source preservation, and immunization perceptibility.Table A
Cross-model transferability (B9Gc, 94ir)Extended evaluation to Qwen-Image, SD1.5, and SDXL, and compared all baselines under SD3.0→FLUX transfer without target-specific adaptation. GuardFlow achieves the largest CLIP-T reduction across all evaluated settings.Tables B, D
Source preservation (B9Gc)Evaluated source similarity using LPIPS, SSIM, PSNR, and DINO. GuardFlow achieves higher source similarity than clean editing across all four metrics, and LT\mathcal{L}_T yields consistent gains over a simple magnitude penalty ∥δ∥22\Vert\delta\Vert_2^2.Table C
Methodological novelty (94ir)Conducted controlled comparisons with LRFDS\mathcal{L}_{\mathrm{RFDS}} and fixed lower bounds LBfixed\mathcal{L}_B^{\mathrm{fixed}} under matched architecture and training settings. GuardFlow achieves stronger editing disruption than all other objective configurations, supporting the contributions of trajectory disruption, terminal guidance, and image-specific lower-bound regularization.Table E

All corresponding revisions and additional experimental results have been incorporated into the updated manuscript, with the changes highlighted in blue\color{blue}{\text{blue}} for clarity. We hope these additional evaluations address the reviewers' concerns and clarify the scope and contributions of GuardFlow. We would be happy to address any further questions or comments.

AuthorsOpenReview ↗

Final Summary of the Author-Reviewer Discussion

Dear Area Chair and Reviewers,

As the discussion period draws to a close, we provide a concise summary mapping each concern raised in the original reviews to the corresponding evidence. Reviewer RSxL confirmed that our response addressed their concern, and no other reviewers have raised any new concerns regarding our responses.

Reviewer RSxL

  • Human evaluation: In a blind study with 40 participants (App. E.7), GuardFlow yields 41.7% edit success versus 64.2–71.7% for the baselines, the highest source-preservation vote share (58.6%), and near-chance perceptibility (55.3%). The reviewer confirmed that these results address their concern.

Reviewer 94ir

  • W1/Q1 (protection against unknown editors): As the reviewer noted that a comparison under the same SD3→FLUX setting would materially affect their assessment, we have provided additional cross-transfer results. With perturbations generated using only SD3.0, GuardFlow reduces FLUX CLIP-T by 3.01, compared with 0.81 for the strongest baseline (Table D). GuardFlow also achieves the largest reduction in all eight transfer settings, including Qwen-Image (distinct architecture and VAE), SD1.5, and SDXL (Table B).
  • W2/Q2 (contribution beyond adapting existing techniques): Under an identical UNet++ immunizer, training data, and budget, replacing L_traj with an RFDS objective raises CLIP-T from 22.89 to 25.40, and replacing L_T with a magnitude penalty raises it to 24.41. The image-specific bound outperforms both removing L_B (24.24) and the best tuned fixed bound (23.84) (Table E).

Reviewer B9Gc

  • W1/Q1 (transfer to distinct architectures, VAEs, and diffusion editors under matched budgets): Table B, as above.
  • W2/Q2 (source preservation and human-assessed edit success): GuardFlow improves source similarity on all four metrics (Table C) and achieves the lowest human-assessed edit success rate (App. E.7).

All results have been incorporated into the revised manuscript (App. E.4, E.7). We have also clarified the scope of our transferability claim, acknowledging that protection against arbitrary future editors cannot be guaranteed.

We sincerely thank the Area Chair and all reviewers for their time and consideration.

Best regards,

Authors of Submission 6063

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.