- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Key-Recoverable Conditional Flow Hijacking, which enables authorized users to perform reference-consistent editing on a protected image while preventing unauthorized users from obtaining reference-consistent edits. The key idea is to steer the editing trajectory so that the edited result of the perturbed image falls into the region that satisfies prompt consistency but not reference consistency. To do this, the authors optimize the perturbation by minimizing the hijacking loss between the context steering of the protected image and that of a hijack anchor, along with a repulsion loss from the clean context steering and an attention preservation loss. They further optimize a key embedding appended to the prompt embedding so that the edited results with the correct key fall into the region satisfying both the reference condition and the text condition. The perturbation and the key are optimized alternately while the pretrained editor is kept frozen. To demonstrate the effectiveness of the method, they conduct experiments on FLUX.1-Kontext with CelebA-HQ and VGGFace2, and further evaluate key selectivity, cross-model transferability, and robustness to image transformations and purification.
Critical strengths
- By introducing key parameters directly into the existing text path of the Diffusion Transformer (DiT), this method eliminates the need for additional modules or structural model modifications while generating distinct key values for each image.
- The underlying objectives are well designed through the steering decomposition in Equation 4, where the specific role of each loss component is thoroughly validated across transplantation, steering replacement, and ablation experiments.
- The approach demonstrates exceptional cross-model generalization where keys trained on FLUX.1-Kontext operate seamlessly across four other un-retrained editors, significantly boosting ICEdit's performance from 0.49 to 0.77.
Critical weaknesses
- Although the loss formulation seems appropriate, the core novelty of the loss formulation appears limited. The attention loss merely reverses the sign of the metric used in DeContext. Furthermore, and share the exact two-loss structure of Semantic Injection (Lee et al., 2026) which pulls toward the target semantics while suppressing the source semantics.
- The evaluation lacks robustness against adaptive attacks. Because the key operates as a continuous embedding space, an attacker could directly optimize a key to maximize identity similarity with the protected image. Merely testing against Random Key and Swap Key baseline scenarios fails to establish security against such tailored, white-box adaptive threats.
- The experimental scope and overall performance present notable weaknesses. The evaluation relies on a restricted dataset of only 50 face images per set, and selectivity is validated using just two prompts on CelebA-HQ. Even under Correct Key conditions, the protected outputs exhibit noticeable quality degradation compared to clean images, dropping in both ISM (from 0.78 to 0.68) and CLIP-IQA (from 0.81 to 0.70).
- Cross-model transfer experiments were mostly conducted on FLUX-family models, making it difficult to consider them true black-box transferability evaluations. Additionally, there is no explanation of how the key was applied to Step1X-Edit, which uses a different text encoder, nor is there a comparison with the transfer performance of baseline methods.
Questions for authors
Does protection and key restoration hold on Stable Diffusion based editing models like InstructPix2Pix or models with different architectures and text encoders like Qwen-Image-Edit? In particular, because UNet-based models utilize a cross-attention structure rather than joint attention, the proposed steering decomposition does not directly apply. Can you demonstrate whether the same image–key pairs remain functional in this setting, or if model-specific re-optimization is required?
5 comments
2、Response to Weakness 3、Weakness 4 and Question [1/2]
Response to Weakness 3: Experimental Scope and Performance
Reviewer’s Concern: Limited evaluation scale and degraded authorized editing quality.
Response:
We thank the reviewer for raising concerns about evaluation scale, key selectivity, and authorized editing quality.
Regarding evaluation scale, we acknowledge that 50 reference images per dataset is relatively limited. Nevertheless, this is comparable to DeContext and NullEdit [1], which each evaluated 50 images per dataset, and CaD [2] and IDGuardian [3], which used 50 test images. Additionally, we evaluate 60 diverse prompts unseen during optimization, yielding 3,000 image-prompt combinations per dataset. Although broader identity coverage is desirable, evaluating 60 unseen prompts per reference provides substantial diversity in editing scenarios. We are currently expanding the datasets and sample sizes to further assess generalization.
Regarding Key Selectivity, we extend the original two-prompt evaluation to 60 diverse editing prompts, comparing No Key, Random Key, Swap Key, and Correct Key on CelebA-HQ.
| Setting | FDFR | ISM ↑ | CLIP-IQA ↑ | CLIP-Score ↑ |
|---|---|---|---|---|
| Clean | 0.00 | 0.78 | 0.81 | 26.56 |
| No Key | 0.09 | 0.26 | 0.76 | 27.67 |
| Random Key | 0.09 | 0.26 | 0.74 | 27.41 |
| Swap Key | 0.00 | 0.28 | 0.72 | 27.22 |
| Correct Key | 0.00 | 0.68 | 0.70 | 26.46 |
Across 60 diverse unseen prompts, Correct Key achieves an ISM of 0.68, substantially outperforming No Key (0.26), Random Key (0.26), and Swap Key (0.28). This confirms the image-specific selectivity of the learned keys.
Regarding authorized editing quality, we acknowledge the remaining gap from Clean (ISM: 0.78 vs. 0.68; CLIP-IQA: 0.81 vs. 0.70). Our key restores reference-consistent editing rather than removing the perturbation. Since authorized editing still uses the perturbed reference, residual identity deviations may contribute to the ISM gap, while perturbation artifacts may affect CLIP-IQA.
Appendix A.4.1 further demonstrates the protection–quality trade-off. Increasing the perturbation budget ε from 0.02 to 0.10 reduces No-Key ISM from 0.71 to 0.14, while Correct-Key ISM decreases from 0.97 to 0.77 and CLIP-IQA from 0.75 to 0.70. A smaller budget can improve authorized editing quality but may weaken protection.
We are investigating ways to improve recovery and reduce visual artifacts while maintaining protection.
Response to Weakness 4 and Question: Cross-Model Transferability and Key Adaptation [1/2]
Reviewer’s Concern: The cross-model transfer evaluation lacks comparisons with existing protection baselines.
1. Comparison with Baseline Protection Transfer
We additionally evaluate AdvDM and DeContext on Step1X-Edit and ICEdit. Since these baselines provide protection without key-based recovery, we compare their No-Key transfer performance.
| Target Editor | Method | FDFR | ISM ↓ | CLIP-IQA ↑ | CLIP-Score ↑ |
|---|---|---|---|---|---|
| Step1X-Edit | Clean | 0.00 | 0.83 | 0.75 | 24.36 |
| AdvDM | 0.00 | 0.50 | 0.60 | 24.06 | |
| DeContext | 0.00 | 0.60 | 0.59 | 24.34 | |
| Ours (No Key) | 0.00 | 0.42 | 0.54 | 23.73 | |
| ICEdit | Clean | 0.00 | 0.91 | 0.78 | 23.46 |
| AdvDM | 0.00 | 0.82 | 0.57 | 24.10 | |
| DeContext | 0.00 | 0.69 | 0.64 | 23.70 | |
| Ours (No Key) | 0.00 | 0.49 | 0.49 | 23.57 |
Our method achieves stronger identity disruption on both editors. On Step1X-Edit, ISM is 0.42, compared with 0.50 for AdvDM and 0.60 for DeContext. On ICEdit, it is 0.49, compared with 0.82 and 0.69, respectively. We compare all methods on the same editors using the same identity metric. Since AdvDM and DeContext lack key-based recovery, this comparison concerns unauthorized protection only.
Unlike these protection-only baselines, our method also supports recovery: Correct Key increases ISM from 0.42 to 0.63 on Step1X-Edit and from 0.49 to 0.77 on ICEdit, without target-specific re-optimization.
Reviewer’s Concern: Explanation of how the key was applied to Step1X-Edit, which uses a different text encoder.
2. Key Injection Across Different Text Encoders and Architectures
For Step1X-Edit, we truncate the FLUX.1-Kontext key from 16 × 4096 to 16 × 3584, then append it as additional tokens to the encoded text-conditioning sequence.
We use the same strategy for Qwen-Image-Edit and InstructPix2Pix: the original FLUX key is truncated to match the target text-conditioning dimension and appended to its encoded text embeddings. This reuses the original key without additional trainable adapters or target-editor modifications. The same dimension-matching truncation is applied to the other editors.
[1] Huang et al. NullEdit. arXiv:2608.10870, 2026.
[2] Abdel Magid et al. CaD. arXiv:2610.35566, 2026.
[3] Xiong et al. IDGuardian. CVPR, 2026.
3、Response to Weakness 4 and Question [2/2]
Response to Weakness 4 and Question: Cross-Model Transferability and Key Adaptation [2/2]
Reviewer’s Question: Does protection and key recovery transfer across different architectures, such as Qwen-Image-Edit and InstructPix2Pix, or is model-specific re-optimization required?
Reponses3-4:
3. Protection and Authorized Recovery Across Different Architectures
We further evaluate transferability on Qwen-Image-Edit and the U-Net-based InstructPix2Pix. All protected images and keys are optimized on FLUX.1-Kontext and reused directly, without target-specific re-optimization.
| Target Editor | Setting | FDFR ↓ | ISM ↑ | CLIP-IQA ↑ | CLIP-Score ↑ |
|---|---|---|---|---|---|
| Qwen-Image-Edit | Clean | 0.00 | 0.67 | 0.73 | 23.90 |
| No Key | 0.00 | 0.31 | 0.71 | 24.66 | |
| Correct Key | 0.45 | 0.50 | 0.65 | 22.25 | |
| InstructPix2Pix | Clean | 0.02 | 0.97 | 0.78 | 23.53 |
| No Key | 0.16 | 0.71 | 0.46 | 23.27 | |
| Correct Key | 0.18 | 0.84 | 0.67 | 23.43 |
These results show partial protection and recovery transfer, but also limitations across heterogeneous architectures. On Qwen-Image-Edit, No-Key ISM decreases from 0.67 to 0.31, with CLIP-IQA remaining close to Clean. The Correct Key raises ISM to 0.50 among face-detected outputs; however, FDFR reaches 0.45, meaning 45% of outputs contain no detectable face and recovery remains unstable.
On InstructPix2Pix, No-Key ISM decreases from 0.97 to 0.71, and the Correct Key raises it to 0.84. However, No-Key CLIP-IQA drops to 0.46; the Correct Key improves it to 0.67, but it remains below Clean.
We acknowledge that our Context Steering decomposition, formulated for DiT joint attention, does not directly apply to the cross-attention structure of U-Net models. Differences in visual conditioning, text representations, and denoising dynamics also make cross-architecture transfer challenging. Thus, while protective perturbations show partial transfer, directly reusing the FLUX key does not guarantee stable, high-quality recovery.
4. Discussion on Cross-Model Authorization and Potential Key Adaptation
We regret that our method still has limitations in cross-model generalization, particularly across editing architectures and text-conditioning interfaces that differ substantially from those evaluated in our experiments.
We further clarify the role of cross-model transfer in the server-side authorization scenario illustrated in Fig. 1. In our framework, image owners publicly release protected images, while authorized users hold access credentials. The Protection Service Provider verifies user authorization and manages the keys associated with each protected image. From this perspective, cross-model transferability is a desirable capability rather than a strict requirement of our authorization mechanism. The essential requirement is to verify whether a user is authorized to edit a protected image, rather than to ensure that a single fixed key embedding is universally compatible with every editor. If different editors require different key embeddings, the Protection Service Provider can associate model-specific keys with the same user credential and provide the appropriate key for the selected editor. In this way, authorized users can access different editing models using their corresponding keys, without republishing the protected image or changing the overall authorization workflow.
Motivated by this consideration, we have further thought about a potential direction for reducing the cost of cross-model adaptation. One possible approach is Warm-Started Cross-Model Key Adaptation, which initializes the target-model key from the existing FLUX key and learns only a model-specific residual:
Here, adjusts the original key to the target conditioning dimension, while is a residual embedding optimized for the target editor. The protected image and pretrained target editor remain frozen, and only the residual key is optimized. Specifically, the target editor's predictions conditioned on the clean reference image can serve as supervision to align the editing dynamics of the protected image with those of clean editing. In this way, the existing FLUX key provides a warm start, potentially reducing the adaptation effort compared with optimizing a new key from random initialization.
We emphasize that this is a preliminary idea rather than an established solution, and its effectiveness across different architectures and text encoders remains to be investigated. We will add the results once they become available.
We hope that the above clarifications and additional experiments address the reviewer's concerns. We will incorporate the corresponding explanations, experimental results, and discussions into the revised manuscript.
1、Response to Weakness 1 and Weakness 2
Response to Weakness 1: Novelty of the Loss Formulation
Reviewer’s Concern: Clarify the novelty of the proposed loss formulation and its technical differences from DeContext and Semantic Injection.
Response:
We thank the reviewer for the thoughtful analysis. We would like to clarify that our contribution is not a simple combination or sign reversal of existing losses, but a recoverable image editing protection mechanism based on Conditional Flow Hijacking, enabling unauthorized protection and key-based recovery without modifying the pretrained DiT or introducing additional auxiliary models.
Our protection mechanism fundamentally differs from DeContext. DeContext suppresses Target-to-Context Attention to weaken reference information propagation. As acknowledged in its Appendix D, this strategy becomes less effective in complex scene edits where generation is dominated by textual conditions and relies less on the visual reference. In contrast, our method preserves context interaction while actively redirecting reference-induced steering through the hijacking loss (Lhij) and the representation loss (Lrep). Thus, the attention loss (Lattn) serves to maintain context engagement rather than directly enforce protection.
Our method also differs substantially from Semantic Injection (Lee et al., 2026), despite the similar attraction–repulsion formulation. Semantic Injection optimizes U-Net cross-attention outputs to inject predefined target semantics, providing one-way image protection. In contrast, our method redirects reference-induced Context Steering in DiT-based in-context editing to disrupt unauthorized identity preservation. More importantly, we jointly optimize an image-specific key embedding through the native text-conditioning pathway to restore reference-consistent editing. This asymmetric, key-recoverable control is not addressed by Semantic Injection.
The novelty of our method lies in integrating Context Steering Redirection, Context Engagement Preservation, and Key-conditioned Recovery to achieve image-level authorization without modifying the pretrained editor, rather than in the mathematical form of individual loss terms. We will clarify these distinctions in the revised manuscript.
Response to Weakness 2: Robustness Against Adaptive Attacks
Reviewer’s Concern: Random Key and Swap Key tests are insufficient to establish robustness against white-box adaptive attacks that optimize alternative key embeddings.
Response:
We thank the reviewer for this valuable suggestion. We agree that Random Key and Swap Key primarily evaluate key selectivity and do not fully assess robustness against adaptive attacks on continuous key embeddings. We therefore conducted an additional white-box adaptive key optimization experiment.
Specifically, we designed an attack in which the attacker has access to the publicly available protected image and the editing model, but not the clean image or the correct key. The attacker optimizes an alternative key embedding by maximizing ArcFace facial feature similarity between the edited output and the protected image, while keeping the editor frozen.
We evaluate the optimized key on CelebA-HQ using the fixed prompt "a photo of this person", with the following results:
The optimized alternative key increases ISM only slightly from 0.14 to 0.21, remaining well below the 0.77 achieved with the correct key. This indicates that the attacker fails to recover effective identity-preserving editing under the tested white-box attack setting.
We appreciate the reviewer for suggesting this meaningful attack scenario, which also opens up interesting directions for future research on adversarial attacks against multimodal models. Our results demonstrate robustness against the tested adaptive attack, and we will include this experiment and discussion in the revised manuscript.