acceptodds
Under review as a conference paper at ICLR 2027

Reference-State Decoupling Regeneration for Restoring Image Editability

Abstract

Protective perturbations can severely disrupt downstream image editing while remaining visually similar to clean images. Existing diffusion-based purification methods typically initialize generation from a noised protected image, causing image content and protective perturbations to enter the initial generative state together. Stronger noising may suppress the perturbations but also discard the content required for faithful editing. To better understand this trade-off, we distinguish two roles of the protected image: state initialization, which directly introduces protected pixels into the generative state, and reference conditioning, which provides image information to guide generation. We investigate how these distinct input roles affect editability restoration under protective perturbations. A diagnostic experiment shows that progressively reducing the protected image's contribution to the initial state improves editability restoration while maintaining comparable reconstruction fidelity, with independent initialization achieving the highest Recovery. Motivated by this finding, we propose Reference-State Decoupled Regeneration (RSDR), which retains the protected image as a fixed reference while initializing the evolving pixel state from independent Gaussian noise. Time-dependent reference features guide the state trajectory through a lightweight interface, enabling content-preserving regeneration without directly injecting protected pixels into the initial state. Experiments across four protection methods demonstrate consistent improvements in editability restoration over existing baselines. These results demonstrate the effectiveness of decoupling reference conditioning from state initialization for restoring image editability under protective perturbations. Code will be released upon acceptance.

Primary area:
generative models
Keywords:
protective perturbations, image purification, image editability, diffusion-based image editing, generative purification, reference-state decoupling

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Z3ZqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper contributes in exploring image protection from generative model based editing by proposing a method to by-pass (in a sense) the perturbation-based safeguards for image editing. In perturbation based safeguards, a virtually imperceptible perturbation is added to images that disrupt the downstream image editing performance with generative models. The common paradigm to attack such a safeguard is to denoise the image, and/or use basic image transforms before performing editing. This works reasonably well but causes quality degradation. The key idea is to separate denoising process (for perturbation cleansing) from the pixel values through a separate stream that only provides time modulated latents to the denoising process as guidance. This is taken to the extreme where even the initial state of denoising is sampled from a Gaussian (independent of the target image). A method is developed that enables such denoising by training an interface model between the latent-providing stream and the denoising stream. The training of the interface model is kept pragmatic, using synthetic distortions. Before presenting the method, the paper empirically establishes that keeping the initial (and early) state(s) of the denoising process independent of the protected image helps the cause of editability recovery. Experiments establish the main contribution of the paper in a convincing manner.

Critical strengths

The paper is well-written and easy to follow. It introduces an interesting idea of by-passing perturbation-based safe guards against generative editing by decoupling the initial state of denoising from the protected image. The methods seems practical. The paper provides a clear motivation and intuition behind its key idea with an empirical evaluation before detailing the method. Overall performance is good.

Critical weaknesses

As such the reviewer did not find any 'critical' weakness that should cause a rejection. However, the paper could benefit from discussing the limitations and failure cases, a discussion currently missing.

Questions for authors
  • On line 185, it is mentioned hat a separate model is trained for each η\eta value. Could the authors clarify why was that necessary.
  • Did the authors observed any challenging cases where the approach struggled?
  • The CLIP-Dir metric seems important. In that metric PDM-Pure is often comparable (or better) than the proposed method. What advantage does RSDR provide over PDM-Pure beyond the metrics in Table1?
3 comments
AuthorsOpenReview ↗

We welcome any further discussion or feedback from the reviewer Z3Zq

Dear Reviewer Z3Zq,

We sincerely thank you for your time and effort in reviewing our paper and providing valuable feedback.

As the rebuttal and discussion period is approaching its end, we would be grateful if you could kindly take a moment to review our responses. We have carefully addressed your concerns and hope that our clarifications have helped resolve the issues raised in your review.

If any concerns remain, we would be more than happy to provide further clarification or additional evidence. We would greatly appreciate your updated assessment in light of our responses.

Thank you again for your valuable time and consideration.

Best regards,

The Authors

Thank you for your responses. My concerns are addressed.

AuthorsOpenReview ↗

Thank You for Your Positive Assessment

Reviewing many assigned papers is a time-consuming task and so we would like to sincerely thank you again for your time and effort to review the paper.

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

The paper studies restoring the editability of images protected by adversarial perturbations. It anaylzes state initialization and reference conditioning and shows through a diagnostic experiment that reducing the protected image’s contribution to the initial state improves editability restoration. Based on this, the authors propose RSDR, which initializes with the Gaussian noise and use the protected image only as a reference. RSDR uses a lightweight AdaLN interface on a pretrained pixel-space flow model with reconstruction losses. Experiments cover four protection methods with inpainting-based editing.

Critical strengths
  1. The idea is simple and clear. Separating whether the protected image enters the initial state or is used only as a reference is an easy way to look at generative purification.

  2. The improvements are consistent. RSDR outperforms the baselines across four protection methods, and the gains remain on held-out editing masks.

  3. The training is lightweight, since only a small interface is trained with synthetic degradations, without protection-specific data.

Critical weaknesses
  1. The motivation for decoupling the two roles seems insufficient. Reference conditioning has already been used in conditional restoration, so the diagnostic experiment mainly re-confirms that it works better than state initialization in this task, and such a reference-conditioned restoration method is not included as a baseline. In addition, the experiment shows that the initial state is not needed to preserve content when the reference is available, but this does not explain why the two roles have to be decoupled. It only shows that the protected image can be removed from the initial state without losing content, not that the perturbation entering through the initial state is what limits editability restoration.

  2. Only one way of reference conditioning (AdaLN modulation) is examined. The reference could also be given by concatenation with the input state, cross-attention, or ControlNet, so it is hard to conclude from this setting alone that the initial state is not useful in general.

  3. The interface is trained with synthetic degradations instead of protected images, but the paper does not compare it with training on protected images. In addition, the training data contains only FFHQ faces, while the evaluation includes humans, animals, and objects.

  4. In Figure 7(b), using the clean reference instead of the protected one gives almost the same Recovery. This suggests that the gain may come from the reference encoder being insensitive to the perturbation rather than from the decoupling itself.

  5. (Minor) The paper only varies the initial state while the reference is always the protected image. It would be interesting to see whether the initial state becomes important again when the reference is degraded, for example by adding noise.

Questions for authors

Refer to the weaknesses.

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

The paper addresses the problem of restoring the editability of images containing protective perturbations. The authors propose RSDR, which regenerates an image starting from independent Gaussian noise while using the protected image as a reference. The goal is to reconstruct the original content without preserving the protective perturbations, allowing the image to be edited again.

Critical strengths

The problem is interesting given the increasing use of protective perturbations against unauthorized image editing.

The distinction between state initialization and reference conditioning is useful. The authors provide an empirical analysis of this design choice and evaluate the method against four protection settings.

Critical weaknesses

I am not familiar with the field nor aware of the previous works. Yet, here are my comments:

  • I will be surprise if the idea of generating an image from independent noise while conditioning on a reference image is new. Reference-conditioned generation is already established in image restoration, such as DiffBIR, which the authors cite.

Therefore, the author seems to apply this existing principle to protective perturbation removal. It is unclear what the fundamental methodological contribution is beyond this application.

  • I am also not familiar with the details of the experiments and baseline. It is therefore difficult to determine whether the improvements come from the proposed reference-state decoupling or from the generative backbone used. A comparison using the same backbone as previous work would help isolate the contribution.

  • The trade-off between removing perturbations and preserving image content is intuitive. However, the quantitative results do not fully establish whether the proposed method offers a better trade-off compared to previous work. For example, Table 3 shows that RSDR achieves substantially better editability recovery but lower image fidelity, particularly in SSIM and LPIPS. It is unclear whether the improvement comes from effectively removing protective perturbations or simply regenerating a semantically similar but structurally different image. Maybe previous work could have the same performance...A comparison at matched reconstruction quality would be more informative.

  • The qualitative results are very limited. I was hoping to see more examples in the appendix. I could not find them. I hope i didn t miss the appendix. More examples involving human faces would be useful to assess whether identity and fine details are preserved. This is important because a regenerated image may remain semantically similar while changing important visual attributes.

  • The method requires an additional generative process before editing, using 25 sampling steps. However, the paper does not provide a quantitative comparison of inference time or computational cost against the baselines. This makes it difficult to assess the practical benefits of the approach.

Questions for authors

To sum up all the points above:

  1. Novelty: is there methodological contribution? Or just suggesting to an existing reference-conditioned image regeneration method as a new baseline to remove protective perturbation.
  2. Same-backbone comparison: Could the authors compare RSDR with a standard reference-conditioned regeneration method using the same pretrained backbone? This would help isolate the contribution of the proposed design.
  3. Fidelity vs. editability: Table 3 shows a substantial improvement in Recovery but a decrease in SSIM and LPIPS. Could the authors compare the methods at matched reconstruction fidelity to better evaluate this trade-off?
  4. Content preservation: Could the authors provide more examples involving human faces and evaluate identity preservation?
  5. Efficiency: What is the computational overhead of RSDR? How does its inference time compare with the other methods?
1 comment
AuthorsOpenReview ↗

We welcome any further discussion or feedback from the reviewer DTWR

Dear Reviewer DTWR,

Reviewing many assigned papers is a time-consuming task and so we would like to sincerely thank you again for your time and effort to review the paper.

As the rebuttal and discussion period is approaching its end, we would be grateful if you could kindly take a moment to review our responses and additional experimental results. We have carefully addressed your concerns and hope that our clarifications have helped resolve the issues raised in your review.

If any concerns remain, we would be more than happy to provide further clarification or additional evidence. We would greatly appreciate your updated assessment in light of our responses.

Thank you again for your valuable time and consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.