- 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
-
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.
-
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
-
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.
-
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
-
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.
-
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 or using a tuned fixed bound, since these comparisons would determine which proposed components substantiate the claimed advance.
5 comments
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 objective | CLIP-T ↓ | SSIM(Noise) ↑ | |
|---|---|---|---|
| No immunization | 28.46 | – | – |
| 25.40 | 0.909 | 4.36/255 | |
| (w/o ) | 24.24 | 0.928 | 3.53/255 |
| () | 24.16 | 0.905 | 4.15/255 |
| () | 23.84 | 0.876 | 5.55/255 |
| () | 24.39 | 0.828 | 6.08/255 |
| () | 24.23 | 0.733 | 8.61/255 |
| 26.08 | 0.909 | 4.38/255 | |
| 24.41 | 0.889 | 4.82/255 | |
| (Ours) | 22.89 | 0.900 | 4.58/255 |
(1) Trajectory disruption objective ().
-
Unlike RFDS, which measures velocity residuals against the perturbed image's own trajectory, 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 around the clean trajectory (Appendix A), the perturbation changes the residual by and the RFDS residual by . Hence, when , the shift of the reference largely cancels the change in the predicted velocity, reducing the RFDS term to , whereas is free of this cancellation.
-
Accordingly, replacing with 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 ().
-
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, 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).
-
also differs from a conventional magnitude penalty, as the regularization it induces on the perturbation is direction-dependent. As shown in Appendix C, expanding in yields a linear term that pulls the predicted edit toward the source and an anisotropic quadratic penalty , where and weights each perturbation direction by how strongly it changes the predicted edit.
-
Replacing with a simple magnitude penalty 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.
Response to Reviewer 94ir (3/3)
(3) Image-specific lower bound ().
-
The perturbation magnitude required for effective disruption varies substantially across images. On PIE-Bench++, our estimate 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, incorporates a theoretically derived necessary lower bound on perturbation magnitude based on image-specific vector-field sensitivity.
-
The ablation results show that removing weakens editing disruption, while increasing the fixed bound () leads to larger perturbations without monotonic improvements in editing disruption. Note that 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 , 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)
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.
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).
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.