acceptodds
Under review as a conference paper at ICLR 2027

Self-Editing for Safe Image Generation with Unified Multimodal Models

Abstract

Unified multimodal models (UMMs) combine visual understanding with image generation, but this integration does not necessarily translate into more rational generation behavior, e.g., in safety-sensitive scenarios. Specifically, UMMs often fail to prevent safety violations during generation, yet can recognize them in their own outputs. In this paper, we propose **Safety-Guided Self-Editing** (**SGSE**), a framework that connects safety understanding with safe image generation in UMMs. **SGSE** constructs and distills self-correction trajectories using a UMM's own safety assessments. For each image generated by the UMM, **SGSE** uses the same model to identify how safety violations manifest and formulate an editing instruction describing a benign target scene tailored to the image. It then edits toward this target and reassesses the result, repeating the correction if a violation remains. Accepted trajectories supervise both assessment and editing, allowing the model to learn from iterative correction while retaining the same corrective loop at inference time. This process requires neither external safety assessments nor externally provided reference images. Across diverse harmful concepts, **SGSE** reduces the average attack success rate (ASR) on I2P from 31.31% for the base model to 7.84% with at most one edit, achieving safety comparable to that of the untrained loop with up to three edits while requiring 46.0% fewer edits per image. Moreover, **SGSE** generalizes to a concept-erasure task unseen during safety training while largely preserving general image generation capabilities.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI safety, Safe generation, Text-to-image generation, Image editing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_S9nzOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Safety-Guided Self-Editing (SGSE), a framework for improving image generation safety in unified multimodal models. SGSE uses a model's own safety assessments to identify harmful content, generate image-grounded editing instructions, and iteratively correct unsafe outputs. It further employs self-distillation on successful correction trajectories to improve assessment and editing capabilities while reducing inference overhead. Experiments across different multimodal models and safety benchmarks demonstrate substantial improvements in generation safety while largely preserving general image generation capabilities. The authors also evaluate editing efficiency and generalization to unseen object-erasure tasks.

Critical strengths
  1. The paper addresses the practical challenge of improving image generation safety in unified multimodal models. SGSE offers an interesting approach by leveraging a model's own safety assessments and successful correction trajectories to improve safe image generation while reducing the need for repeated editing.

  2. The paper evaluates SGSE on different unified multimodal models across diverse safety benchmarks. The experiments also include comparisons with existing safety methods, ablation studies, and analyses of editing efficiency and general generation capabilities, providing useful evidence of the method's effectiveness and generalizability.

Critical weaknesses
  1. The proposed self-assessment and editing framework is closely related to prior work on iterative safety editing and feedback-driven image correction. SafeEditor [1] already investigates multi-round post-hoc safety editing, while UniEdit-I [2] explores iterative understanding, editing, and verification in unified models. Furthermore, Xue et al. [3] leverage a model's own safety judgments to improve image generation. Although SGSE introduces endpoint self-distillation, its distinct contribution is not fully isolated. The comparison with SafeEditor does not use matched backbones and training conditions, and the endpoint-trimming ablation does not completely separate the benefits of multi-step distillation from those of training-data selection. Moreover, most retained editing trajectories require only a single correction. More controlled comparisons would help establish whether the improvements specifically arise from the proposed distillation strategy rather than conventional editing fine-tuning.

  2. The paper demonstrates substantial reductions in unsafe image generation and provides additional independent safety evaluation. However, it does not sufficiently examine whether safety editing preserves the legitimate content and semantic intent of the original images. The reported COCO and GenEval results mainly reflect overall generation performance, while only a small proportion of images undergo editing, potentially masking significant changes in the edited subset. Furthermore, SGSE relies on predefined harmful categories and uses the model's own safety judgments to select successful correction trajectories, leaving the reliability of these decisions insufficiently validated. Although these assumptions are reasonable for targeted safety correction, further evaluation of over-editing, preservation of benign content, and independently verified safety decisions would strengthen the evidence supporting the framework's practical effectiveness.

References

[1] SafeEditor: Unified MLLM for Efficient Post-hoc T2I Safety Editing. arXiv:2510.24820, 2025.

[2] UniEdit-I: Training-free Image Editing for Unified VLM via Iterative Understanding, Editing and Verifying. CVPR 2026.

[3] Safe Autoregressive Image Generation with Iterative Self-Improving Codebooks. ICML 2026.

Questions for authors

Please see the Critical Weaknesses above for my main concerns and related questions.

  1. Could the authors further clarify the methodological differences between SGSE and the closely related works [1–3], particularly SafeEditor [1]? Given the differences in backbone models and training conditions, could the authors explain whether more controlled comparisons under matched settings are feasible?

  2. Although the endpoint-trimming ablation demonstrates improved performance, it remains unclear whether the gains arise from compressing multi-step corrections or simply excluding potentially unsafe intermediate training targets. Since most retained editing trajectories require only a single correction, could the authors provide further analysis to isolate the contribution of multi-step self-distillation?

  3. Could the authors further evaluate whether SGSE preserves legitimate visual content during safety editing? In particular, could they report image quality and semantic preservation specifically for edited samples, rather than relying mainly on aggregate COCO and GenEval results?

  4. The proposed endpoint self-distillation assumes that the final editing instruction, originally generated for an intermediate image, can effectively guide correction directly from the original image. However, the paper provides limited theoretical justification for this assumption. Could the authors clarify the conditions under which this strategy is expected to work and provide further theoretical analysis or supporting evidence to explain its effectiveness and limitations?

6 comments
AuthorsOpenReview ↗

Response to Reviewer S9nz (1/2)

We sincerely appreciate your time and effort for reviewing our paper. We respond to each of your questions one-by-one in what follows. Please let us know if you have additional questions or concerns.


(W1-1, Q1). Discussion of related work [1-3]; matched comparison with SafeEditor [1]?

While [1-3] are related to SGSE, we clarify that none of them amortizes the UMM’s own self-correction loop into the UMM by training on its accepted trajectories, which is a distinctive feature of SGSE.

  • SafeEditor [1] trains a separate editor with GPT-4o annotations and target images generated by SD-3.5. SGSE instead requires neither external safety assessments nor reference images.
  • UniEdit-I [2] focuses on making an image edit follow an instruction given by the user more faithfully, through a loop of understanding, editing, and verification. In SGSE, the UMM itself decides whether an image is unsafe and formulates the editing instruction. A direct comparison is therefore not straightforward, as our safety benchmarks provide no editing instructions.
  • Safe-CB [3] focuses on the discrete visual codebook of autoregressive image generation, from which it removes the features of specified harmful concepts using the UMM’s own safety judgments. SGSE instead uses the UMM’s judgments to assess and correct each generated image through editing, and further amortizes this self-correction loop into the UMM. A direct comparison is not feasible on our backbones, as BAGEL and SenseNova generate images without a discrete visual codebook.

Our comparison with SafeEditor in Table 1 follows its plug-and-play design, in which a single trained editor is applied to outputs from different text-to-image models. We apply its released checkpoint to the initial images from each backbone, with the same budget of up to three edits as SGSE. We agree that a comparison with matched backbones would be informative. However, we could not retrain SafeEditor on our backbones as, to our knowledge, its released dataset contains only text annotations without the corresponding images, and the code for constructing its training data has not been released.

We have incorporated these discussions into Section 2 of our revised draft.


(W1-2, Q2). Contribution of multi-step self-distillation

Following your suggestion, we conducted two additional controlled comparisons. For accepted trajectories with multiple edits, we train on only the first edit (x0→x1x_0 \to x_1) or only the final edit (xT−1→x⋆x_{T-1} \to x^\star), instead of the endpoint pair (x0→x⋆x_0 \to x^\star). These trajectories account for 1,467 of the 6,506 editing pairs used for training (22.5%). The three models start from the same checkpoint and are trained for the same number of updates, and they differ only in the editing pair used for these trajectories. As shown in the table below, SGSE achieves the lowest ASR on every dataset with at most one edit. Although these trajectories are a minority, replacing their endpoint pairs with either single-step pair increases ASR on every dataset. In particular, last-step supervision uses the same accepted image x⋆x^\star as its target, so it also excludes unsafe intermediate targets and differs from SGSE only in the input image. The lower ASR of SGSE therefore indicates that the gain is not explained by excluding unsafe intermediate targets alone. We have added these results to Appendix A.10 of our revised draft.

MethodI2P-Sexual ↓RAB ↓MMA ↓T2I-RP ↓
First-step supervision (x0→x1x_0 \to x_1)9.1321.058.6012.79
Last-step supervision (xT−1→x⋆x_{T-1} \to x^\star)10.2025.268.9011.19
SGSE (Ours) (x0→x⋆x_0 \to x^\star)7.6315.797.709.33
AuthorsOpenReview ↗

Response to Reviewer S9nz (2/2)

(W2, Q3). Effect on benign content in edited images

To evaluate whether SGSE preserves legitimate content during safety editing, we measured semantic preservation specifically on edited samples, using unsafe prompts where a safety edit is required. We use ViSU [4], which pairs each benign COCO caption with an unsafe prompt derived from it, so the paired caption specifies the legitimate content that an edit should preserve. We generate an image from each unsafe prompt with BAGEL and take 100 randomly sampled images classified as unsafe by NudeNet or Q16. To compare only the editing of each method, every method is applied to these same images with at most one edit. As shown in the table below, SGSE achieves the highest CLIPScore with the paired benign caption among the editing methods (28.37), while its ASR is the lowest (15.00%). Its CLIPScore also increases from 26.93 before editing, which indicates that the edit removes the unsafe content while keeping the image close to the benign caption. For reference, the images generated directly from the benign caption achieve 31.62, whereas the editing methods are not given the benign caption. These results indicate that SGSE removes more unsafe content than the other editing methods without a larger loss of legitimate content. We have added this evaluation to Section 4.6 of our revised draft.

ASR (%) ↓CLIPScore with benign caption ↑
Before editing100.0026.93
Benign caption-31.62
One-time Editing70.0027.44
SafeEditor37.0026.99
Untrained loop30.0027.86
SGSE15.0028.37

To further examine the preservation of benign content, we have also added qualitative examples of the edits that SGSE applies to images generated from benign prompts (illustrated in Figure 8 of Appendix A.8). For example, for the prompt “A group of surfers change their clothes on a beach”, the generated image contains unexpected nudity, and SGSE adds swim trunks while keeping the surfers, the surfboards, and the beach.


(Q4). Justification for endpoint self-distillation

The final editing instruction is designed to describe the desired benign scene in a self-contained manner, rather than specify a change that depends on the intermediate image. As described in Section 3.2, this design allows the instruction to serve as a correction goal for the original image as well. The strategy is therefore expected to work when the instruction specifies the intended benign outcome without relying on changes already made in intermediate images.

To support this, we additionally evaluated the final instructions on the 1,467 accepted BAGEL training trajectories requiring multiple edits. Given an image, its assessment rationale, and the final instruction (but not the edited output), an independent GPT-5.6 Luna judge rated the instruction as appropriate for the original image in 87.93% of cases, compared with 87.05% for the image immediately preceding the final edit. These results support the applicability of the final instruction to the original image in most evaluated trajectories.

We have incorporated this evaluation and its prompt into Appendix A.3 of our revised draft.

SourceAppropriate (%) ↑
Original image (x0x_0)87.93
Final-step input image (xT−1x_{T-1})87.05

[1] Zhang et al., SafeEditor: Unified MLLM for Efficient Post-hoc T2I Safety Editing, arXiv 2025.

[2] Bai et al., UniEdit-I: Training-free Image Editing for Unified VLM via Iterative Understanding, Editing and Verifying, CVPR 2026.

[3] Xue et al., Safe Autoregressive Image Generation with Iterative Self-Improving Codebooks, ICML 2026.

[4] Poppi et al., Safe-CLIP: Removing NSFW Concepts from Vision-and-Language Models, ECCV 2024.

Reviewer_S9nzOpenReview ↗

Reply to Authors' Rebuttal

Thank you for your detailed responses and additional experiments. I appreciate the authors' efforts and have three additional suggestions:

  1. Inference Efficiency. Section 4.4 shows that SGSE (K=1) achieves comparable safety to the untrained loop (K=3), while reducing the average number of edits by 46.0%. Could the authors additionally report the end-to-end inference throughput of both methods, measured in samples per minute on the same GPU?

  2. Theoretical Analysis. As mentioned in my previous comment, including some theoretical analysis would further strengthen the paper.

  3. Minor Revisions. The table formatting is inconsistent across the manuscript, and some appendix pages contain excessive blank space (e.g., page 24). Please improve the formatting and overall layout.

Reviewer_Wn1fOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Self-Guided Self-Editing (SGSE), a framework that enables unified multimodal models to improve the safety of generated images through self-assessment and iterative editing. Instead of relying on external safety models or human annotations to supervise the editing process, SGSE uses the model’s own safety assessments to construct training data and distills multi-step correction trajectories into a more efficient editing procedure. Experiments on multiple image-generation safety benchmarks show that SGSE substantially reduces the attack success rate (ASR) while reducing the number of editing steps required at inference time. The paper also evaluates the method under an alternative safety evaluation protocol and investigates its generalization to object removal. While the empirical results are promising, the reliability of self-generated supervision, the alignment between the learned safety criterion and external evaluation standards, and the specific contribution of endpoint trimming remain insufficiently established.

Critical strengths
  1. Addresses an important practical problem. The paper studies how to improve the safety of unified multimodal models during image generation, with a focus on correcting unsafe outputs rather than relying exclusively on input filtering or external safety mechanisms.
  2. Presents a potentially useful self-supervised framework. SGSE combines model self-assessment, iterative image editing, and trajectory distillation to reduce dependence on external supervision. The idea of distilling multi-step correction into a more efficient editing process is practically appealing.
  3. Demonstrates promising empirical improvements. The experiments report substantial reductions in ASR across multiple unsafe-content categories. For example, the reported average ASR for SGSE is 6.28%, compared with 24.94% for SafeEditor under the main evaluation protocol. The alternative evaluation in Appendix A.6 also favors SGSE, suggesting that the observed improvement is not limited to a single evaluation setup.
  4. Considers inference efficiency. The method aims to reduce the number of editing steps needed to achieve safe outputs. This is relevant to practical deployment, where repeated image editing can introduce additional latency and computational cost.
  5. Includes useful ablations and supplementary evaluations. The paper examines endpoint trimming and evaluates generalization to object removal. Although these experiments do not fully establish the causal contribution or generality of the method, they provide a starting point for understanding its behavior beyond the primary results.
Critical weaknesses
  1. The reliability of self-generated supervision is insufficiently validated. SGSE constructs its training data using the model’s own safety assessments, but the paper does not independently establish the reliability of these assessments. In particular, the initial images classified as keep are excluded from the audit in Appendix A.3, leaving potential false negatives unexamined. If unsafe images are incorrectly accepted as safe, the resulting supervision may reinforce the model’s existing assessment errors. The paper does not quantify the reliability of these initial keep decisions or analyze how assessment errors contribute to the residual ASR. Consequently, the experiments demonstrate that self-generated supervision can improve benchmark performance, but do not establish that the underlying self-assessment process is consistently reliable.
  2. The relationship between the learned safety criterion and external evaluation standards remains unclear. SGSE constructs its supervision using the model’s own safety definitions, while the main experiments evaluate outputs using NudeNet and Q16. Appendix A.6 additionally reports results using Qwen2.5-VL-72B-Instruct under the UniSAFE rubric. Although SGSE performs well under both evaluation protocols, the substantial differences in absolute ASR—for example, SafeEditor’s reported ASR changes from 24.94% under the main protocol to 5.14% under UniSAFE—indicate that measured performance is sensitive to the evaluation setup. The paper does not provide sufficient analysis of evaluator agreement, category-level discrepancies, or human-verified errors. It therefore remains unclear to what extent SGSE learns a broadly applicable safety criterion rather than one that aligns particularly well with its own assessment procedure or the selected evaluation metrics.
  3. The contribution of endpoint trimming is not sufficiently isolated. The paper uses endpoint trimming to select completed correction trajectories for training, but the ablation does not cleanly isolate its contribution. Removing endpoint trimming changes not only the trajectory-selection strategy but also the completeness of the correction target, the relationship between images and editing instructions, and the distribution of training examples across editing stages. Moreover, the reported editing-appropriateness rate for intermediate editing pairs is substantially lower than that for endpoint pairs. This difference does not necessarily imply that intermediate edits are incorrect, since they may represent valid partial progress; however, the current analysis does not establish which properties of endpoint selection account for the observed improvements. More controlled comparisons are needed to demonstrate the specific benefit of endpoint trimming.
Questions for authors

See the weaknesses

2 comments
AuthorsOpenReview ↗

Response to Reviewer Wn1f

We sincerely appreciate your time and effort for reviewing our paper. We respond to each of your questions one-by-one in what follows. Please let us know if you have additional questions or concerns.


(W1). Reliability evaluation of initial KEEP supervision trajectories

Following your suggestion, we additionally audited the initial KEEP decisions used for training. We used the same GPT-5.6 Luna judge and decoding settings as in Appendix A.3. Each decision is one image-category assessment, and all initial KEEP decisions used for training are audited. As shown in the table below, 98.87% of the initial KEEP decisions from BAGEL and 98.78% of those from SenseNova are judged appropriate. These results support that unsafe images are rarely accepted as safe in the training data. We also note that the assessments become more reliable after training, where the rate of appropriate assessments on all I2P prompts increases from 78.70% for the untrained loop to 84.94% for SGSE (described in Table 8 in Appendix A.4). We have added these results and the evaluation prompt to our revised draft.

Appropriate (%) ↑BAGELSenseNova
Initial KEEP decisions98.8798.78

(W2). ASR discrepancies across evaluation protocols

We clarify that the two evaluation protocols, NudeNet/Q16 and UniSAFE, apply different safety criteria, so differences in their absolute ASR should be interpreted in light of these criteria. NudeNet and Q16, which are established evaluators in text-to-image safety research [1, 2, 3], detect selected exposed body parts and assess general inappropriateness learned from human socio-moral judgments, respectively. UniSAFE instead specifies prohibited content together with contextual exceptions. For example, it allows non-erotic nudity in medical or educational contexts. Following your suggestion, we further analyzed the agreement between the two protocols on the BAGEL outputs of Base Prompting, which averages 72.73% across the seven I2P categories, as shown in the table below.

HateHarassmentViolenceSelf-harmSexualDisturbingIllegal activityAverage
Agreement (%)74.8975.9772.8870.1679.7066.5968.9172.73

Under both protocols, SGSE achieves the lowest ASR in every evaluation column of Table 1 and Table 10 on BAGEL, and neither protocol is used to construct its training data. This also holds under a combined criterion, where an image is considered unsafe if either protocol classifies it as unsafe. As shown in the table below, SGSE achieves an average ASR of 7.17% over the ten evaluation columns of Table 1 under this criterion, compared with 26.98% for SafeEditor. We have added these analyses to our revised draft.

MethodNudeNet/Q16UniSAFECombined
Base Prompting39.8225.2845.76
Safety Guidance38.5529.1249.29
Negative Prompting30.4820.0536.77
One-time Editing43.7828.5648.68
SafeEditor24.945.1426.98
SGSE (Ours)6.281.197.17

(W3). Further ablation study on endpoint trimming

Following your suggestion, we conducted two additional controlled comparisons. For accepted trajectories with multiple edits, we train the base model on only the first edit (x0→x1x_0 \to x_1) or only the final edit (xT−1→x⋆x_{T-1} \to x^\star), instead of the endpoint pair (x0→x⋆x_0 \to x^\star). The three models start from the same checkpoint and are trained for the same number of updates, and they differ only in the editing pair used for these trajectories. As shown in the table below, SGSE achieves the lowest ASR on every dataset with at most one edit. Last-step supervision uses the same instruction and the same accepted image x⋆x^\star as SGSE, so the two differ only in the input image. The gap is largest on Ring-A-Bell (RAB), which has the highest ASR before editing (83.16% in Table 1). For such difficult cases, a single edit of the untrained UMM is often insufficient, as illustrated in Figure 3. Our endpoint trimming lets the UMM learn to perform the whole correction in a single edit, whereas neither variant provides such supervision. We have added these results to our revised draft.

MethodI2P-Sexual ↓RAB ↓MMA ↓T2I-RP ↓
First-step supervision (x0→x1x_0 \to x_1)9.1321.058.6012.79
Last-step supervision (xT−1→x⋆x_{T-1} \to x^\star)10.2025.268.9011.19
SGSE (Ours) (x0→x⋆x_0 \to x^\star)7.6315.797.709.33

[1] Schramowski et al., Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models, CVPR 2023.

[2] Tsai et al., Ring-A-Bell! How Reliable are Concept Removal Methods for Diffusion Models?, ICLR 2024.

[3] Yoon et al., SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation, ICLR 2025.

Dear Reviewer Wn1f,

Thank you again for your time and effort in reviewing our manuscript.

As the discussion period is coming to an end, we would like to kindly ask if there are any remaining questions or concerns that we could help clarify.

We appreciate your thoughtful feedback and hope that our responses have adequately addressed your earlier comments.

Sincerely,
Authors

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

This paper proposes a safety framework for UMMs, especially for the image generation part. The model assesses its generated image, constructs a benign target scene, and iteratively edits the image. Successful trajectories then supervise assessment and editing, using their final instructions and images to teach direct correction from the original image. The experiments show significant safety improvement.

Critical strengths
  1. The proposed method makes effective use of the UMMs' architecture. Using the same model to identify harmful visual content, formulate a correction, and perform editing is a coherent design. Constructing training targets from its own successful corrections avoids requiring an external safety teacher or externally supplied reference images.

  2. The safety improvements are significant and broad. On BAGEL, average I2P ASR decreases from 31.31% to 6.18% with at most three edits, demonstrating the effectiveness.

  3. The benefit of distilling iterative correction is demonstrated clearly. With one permitted edit, trained SGSE achieves similar I2P ASR to the untrained three-edit loop (7.84% v.s. 7.93%), while reducing average edits from 0.938 to 0.506.

Critical weaknesses
  1. The safety-utility evaluation does not cover all cases. Table 13 uses broad category definitions, including crying under violence and courtroom proceedings under illegal activity. These rules could alter legitimate content. However, benign evaluation assigns only one random category to each image.

  2. SafeEditor [1] already combines safety reasoning and iterative image editing in a unified model, while Self-Refine [2] establishes the broader pattern of iterative self-feedback. These methods should be compared.

[1] https://arxiv.org/abs/2510.24820

[2] https://arxiv.org/abs/2303.17651

Questions for authors

Please refer to the weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Cewi

We sincerely appreciate your time and effort for reviewing our paper. We respond to each of your questions one-by-one in what follows. Please let us know if you have additional questions or concerns.


(W1). Benign evaluation across all harmful categories

Following your suggestion, we have expanded the benign evaluation on COCO (reported in Table 2) by assessing each image against all seven I2P categories, rather than a single randomly assigned category. As shown in the table below, with at most one edit (i.e., K = 1), the CLIPScore remains between 31.75 and 32.08 across the categories, comparable to 31.99 for the pretrained BAGEL. SGSE edits fewer than 5% of the images in every category except Disturbing, where 9.4% of the images are edited, and the CLIPScore in this category is still 31.75. In this category, generated images with unnatural depictions also receive an edit (illustrated as the third example of Figure 8), which can lead to a relatively higher edit rate. We have added these results, a detailed analysis of the edited cases, and qualitative examples to Appendix A.8 of our revised draft.

COCOBAGELHateHarassmentViolenceSelf-harmSexualDisturbingIllegal activityAverage
CLIPScore31.9931.9831.9032.0632.0832.0231.7531.9731.97
Edited images (%)-2.74.40.80.32.79.43.03.3

(W2). Comparison with SafeEditor [1] and Self-Refine [2]?

We first note that our original evaluation already compares with SafeEditor [1] in Table 1. While we initially considered our untrained loop baseline to capture a similar feedback-and-refinement process to Self-Refine [2], we appreciate your suggestion and have now included an explicit comparison.

Since Self-Refine was originally proposed for text generation with LLMs, we adapted it to image generation with BAGEL by adopting its key mechanism, where the model refers to the results of its previous edits when formulating the next editing instruction (denoted as Self-Refine$^\star$). As shown in the table below, with up to three edits on BAGEL, SGSE achieves an average I2P ASR of 6.18%, compared with 26.73% for SafeEditor and 13.67% for Self-Refine$^\star$. Our untrained loop, which iteratively refines outputs using the model’s own feedback in a similar manner to Self-Refine, achieves 7.93%. SGSE further reduces ASR, as it amortizes the loop into the UMM by training on its own accepted self-correction trajectories. We have added the comparison with Self-Refine and a detailed discussion of both methods to Appendix A.9 of our revised draft.

MethodI2P ASR (%) ↓
SafeEditor26.73
Self-Refine$^\star$13.67
Untrained loop7.93
SGSE6.18

[1] Zhang et al., SafeEditor: Unified MLLM for Efficient Post-hoc T2I Safety Editing, arXiv 2025.

[2] Madaan et al., Self-Refine: Iterative Refinement with Self-Feedback, NeurIPS 2023.

Dear Reviewer Cewi,

Thank you again for your time and effort in reviewing our manuscript.

As the discussion period is coming to an end, we would like to kindly ask if there are any remaining questions or concerns that we could help clarify.

We appreciate your thoughtful feedback and hope that our responses have adequately addressed your earlier comments.

Sincerely,
Authors

Thank the authors for the rebuttal. I've read the newly added experiments, and they addresses my concerns. I will keep the positive score and increase my confidence.

Thank you for your follow-up and continued support! We are glad that our responses have helped address your concerns. Your thoughtful feedback has greatly helped us improve our manuscript.

Sincerely,
Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revision

Dear Reviewers,

We sincerely appreciate your time and effort in reviewing our paper. Your constructive comments have been invaluable in improving our manuscript.

We have addressed your comments in the revised manuscript, with all additions and revisions highlighted in blue. Specifically, we have made the following changes:

  • An extended benign evaluation on COCO with each of the seven categories assigned to every image (Cewi W1).
  • A comparison with Self-Refine and an additional discussion of SafeEditor, Self-Refine, UniEdit-I, and Safe-CB (Cewi W2; S9nz W1-1, Q1).
  • An additional audit of the initial KEEP decisions used for training (Wn1f W1).
  • An analysis of the agreement between NudeNet/Q16 and UniSAFE, with ASR under a combined criterion (Wn1f W2).
  • Two controlled comparisons on endpoint trimming, with first-step and last-step supervision (Wn1f W3; S9nz W1-2, Q2).
  • An evaluation of content preservation during safety editing in Section 4.6, with qualitative examples of the edits on benign prompts in Appendix A.8 (S9nz W2, Q3).
  • An additional analysis of how appropriate the final instruction is for the original image (S9nz Q4).

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.