acceptodds
Under review as a conference paper at ICLR 2027

When Does an Edit Become Unsafe? Tracing Safety Emergence Along the Generative Trajectory of Image Editing Models

Abstract

Modern image editors do not produce an edit in one step: a source image and instruction are progressively transformed through a diffusion or flow trajectory. Yet existing safeguards mostly inspect the request before generation or the image after generation, leaving unanswered a mechanistic question: when does a policy- violating edit become internally represented and causally committed? We introduce TRACE-EDIT, a trajectory-level analysis that separates Safety Emergence (SEC), Causal Commitment (CCC), and Visual Evidence (VEC). SEC predicts the realized final unsafe-valid-edit outcome and is evaluated under held-out paraphrase families and same-instruction or different-seed outcomes. CCC is measured with both full- residual and low-rank projected donor interventions plus compatibility controls. Matched counterfactual edits derived from SafetyPairs allow paired trajectories to share a source, edit scope, and random seed. Their temporal ordering defines an Unsafe Commitment Window: a middle region where risk is causally actionable before it becomes visually obvious. We turn this observation into TRAJGUARD, a read-only three-checkpoint monitor with frozen checkpoint, block, threshold, and persistence rules. At the matched 5% benign false-block operating point, the common-protocol endpoint is 19.4% unsafe-valid-edit ASR versus 31.3% for Safety-Guided Flow, with a paired absolute gap of 11.8 pp [4.3,19.4]. A matched three-checkpoint visual monitor remains weaker (28.5% ASR), isolating the value of internal states from earlier temporal inspection alone. The study includes adaptive attackers, independent output/edit judges with human audit, category-held-out analysis, SafeEditBench transfer, and model-native replication on LongCat-Image-Edit and OmniGen2.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image Editing Models, Tracing, Generative Trajectory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_A1BoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TRACE-Edit, a framework for investigating when unsafe content emerges during the generation process of image editing models. The framework analyzes three complementary signals: Safety Emergence (SEC), Visual Evidence (VEC), and Causal Commitment (CCC), to identify an Unsafe Commitment Window (UCW). Based on these findings, the authors introduce TrajGuard, a three-checkpoint safety monitor that detects and blocks potentially unsafe edits before generation is completed. Experiments across multiple image editing models demonstrate improvements in safety detection and defense effectiveness.

Critical strengths
  • Important and practical research problem. The paper investigates safety risks during the generation process of image editing models, addressing the limitations of conventional input-side and output-side safety detection.

  • Interesting trajectory-level analysis. TRACE-Edit combines internal-state prediction, visual evidence, and causal interventions to investigate when unsafe editing behavior emerges. The experiments include multiple models, adaptive attacks, and causal ablations.

Critical weaknesses
  • Insufficient distinction from prior trajectory-level safety studies. Although the paper compares against Detect-and-Guide and Safety-Guided Flow, its methodological distinctions from prior studies remain insufficiently clarified, particularly UVR [1], which investigates unsafe information flows in both image generation and editing. The authors should better clarify the novel contributions of TRACE-Edit and provide a quantitative comparison with UVR under consistent evaluation settings.

  • Insufficient justification of the Unsafe Commitment Window design. The proposed UCW combines high SEC, high CCC, and low VEC to select monitoring checkpoints. However, the necessity of these three conditions remains unclear, particularly since TrajGuard primarily performs early abortion rather than intermediate-state correction. The paper lacks ablations examining whether SEC-only checkpoint selection or alternative combinations of these signals could achieve comparable performance.

  • Unclear method organization and naming. The organization of Section 4 is somewhat confusing. While the section is titled TRACE-Edit, Section 4.4 introduces TrajGuard under the ambiguous title "A Deployment Consequence", making the relationship between the two frameworks difficult to follow. The authors should more clearly distinguish the offline trajectory analysis performed by TRACE-Edit from the deployment-time monitoring performed by TrajGuard.

  • Missing conclusion section. The paper ends with a Discussion section but lacks a dedicated Conclusion section. A concise conclusion summarizing the main findings, contributions, and limitations would improve the completeness of the presentation.

[1] Yang, Xiang, et al. "Unified Safe In-context Image Generation in Multimodal Diffusion Transformers via Restricting Unsafe Information Flows." ICML, 2026.

Questions for authors
  • Relation to prior work. Could the authors clarify the methodological differences between TRACE-Edit and UVR [1], and provide a quantitative comparison under consistent evaluation settings?

  • UCW design. Why are high CCC and low VEC necessary for selecting monitoring checkpoints when TrajGuard ultimately relies on SEC-based detection and early abortion? Would SEC-only checkpoint selection achieve comparable performance?

[1] Yang, Xiang, et al. "Unified Safe In-context Image Generation in Multimodal Diffusion Transformers via Restricting Unsafe Information Flows." ICML, 2026.

Reviewer_Vj6WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies when unsafe image-editing outcomes become internally predictable, causally influenceable, and detectable in reconstructed images. TRACE-Edit measures SEC, donor-patching CCC, and VEC on matched trajectories, then defines an Unsafe Commitment Window used to select three checkpoints for TrajGuard. The primary intervention is early abort. At approximately 5% benign false-block rate, TrajGuard achieves 19.4% unsafe-valid-edit ASR versus 31.3% for Safety-Guided Flow.

Critical strengths

1.The study includes substantive controls against alternative explanations. Realized-outcome labels, held-out instruction families, identical-request/different-seed tests, projected interventions, and mismatched donors provide stronger evidence than probe accuracy alone. 2.The practical evaluation addresses temporal access and endpoint independence. A matched three-checkpoint visual monitor helps isolate representation type. Independent policy and edit-validity judges, human auditing, and paired uncertainty strengthen the comparison within the evaluated protocol.

Critical weaknesses

1.The safety specificity of the Unsafe Commitment Window remains insufficiently established. A middle-trajectory causal-patching peak may reflect a general property of image editing rather than a mechanism specific to unsafe outcomes. The current controls therefore support temporal causal influence more directly than a safety-specific commitment mechanism. 2.Cross-model replication does not yet establish cross-model adaptive robustness. Although the temporal analysis is independently reproduced on three editors, defense-aware attacks are evaluated primarily on FLUX. The remaining uncertainty is whether robustness and its degradation with attack budget depend strongly on the underlying editor.

Questions for authors

1.Do scope- and magnitude-matched benign edit pairs exhibit similar CCC peaks? If so, what evidence distinguishes safety-related commitment from the formation of ordinary edit semantics? 2.Can the same defense-aware attacks be evaluated on at least one additional editor, with ASR across query budgets and paired differences against matched baselines? This would clarify whether the robustness findings generalize across editors.

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

This paper investigates when and how unsafe content emerges during the image-editing trajectories of diffusion and flow models. The authors propose TRACE-Edit, which combines three measurements—Safety Emergence Curve (SEC), Causal Commitment Curve (CCC), and Visual Evidence Curve (VEC)—to identify an Unsafe Commitment Window (UCW). Based on this window, they develop TrajGuard, a lightweight defense that monitors intermediate representations to detect and block potentially harmful edits.

Despite the meaningful research question, I recommend rejecting this paper because (1) the manuscript is poorly organized and difficult to follow, with unclear explanations of the proposed architecture and methodology, (2) the methodological design lacks sufficient analysis and justification, particularly regarding the safety objective and UCW selection, and (3) the experimental evaluation is not sufficiently thorough, lacking important component-wise ablations and detailed analysis to support the claimed contributions.

Critical strengths
  • The paper raises a meaningful research question by investigating when and how unsafe content emerges during the image-editing trajectories of diffusion/flow models. By challenging the idea that harmful content should only be detected at the beginning or at the end of the process, the paper opens an interesting direction for better understanding and preventing harmful content generation.

  • The paper proposes TRACE-Edit, an analytical framework that combines three curves (SEC, CCC, and VEC) to identify the Unsafe Commitment Window (UCW), where harmful images can be predicted from internal features before becoming visually detectable. This window then guides TrajGuard in detecting potentially harmful edits early in generation.

Critical weaknesses

The presentation and readability of the manuscript need significant improvement:

  • Unclear architecture overview: Figure 2 fails to clearly illustrate the mechanisms and roles of SEC, CCC, and VEC, why they are needed, and how they together identify the UCW. The connection between TRACE-Edit's offline analysis and TrajGuard's online defense is also poorly illustrated.

  • Ambiguous and complicated terminology: Terms such as “natural donor patching” (line 234) and “rank-8 donor-difference basis” (line 240) are introduced without clear definitions or explanations, making it difficult to understand what these operations actually mean or how they are performed.

  • Poor organization of the methodology: The authors frequently mix methodological explanations with implementation details (e.g., lines 213–215, 252–254, and 266–267), making it difficult to follow the main ideas and the reasoning behind the proposed approach.

The methodological principles, hyperparameter setting and architectural details need clearer explanations and more thorough analysis:

  • Unclear architectural details: Some description about the details, such as “image-token states after the joint image/text conditioning path” (lines 218–219) is difficult to interpret without further explanation of what it refers to in the model architecture. Similarly, the transformer block ℓ\ell (Equation 4) used in CCC is not sufficiently explained, making the implementation difficult to understand and reproduce.

  • Questionable safety objective: The paper defines attack success as UVE=Pr⁡(U=1∧V=1)UVE=\Pr(U=1\land V=1) (line 148), requiring the output to be both unsafe and instruction-valid. However, existing safety-generation methods commonly evaluate harmful content separately from prompt adherence. This narrow objective may overlook unsafe outputs that fail to follow the instruction.

  • Insufficient justification of UCW thresholds: The thresholds d0=0.75d_0=0.75, c0=0.20c_0=0.20, and g0=0.70g_0=0.70 are not thoroughly analyzed (Section 4.3, lines 251–255). Although the authors state that these values were selected on the development set, they provide no sensitivity analysis showing how different thresholds affect the selected UCW or defense performance.

The experimental evaluation needs more thorough analysis of the proposed method and clearer explanations of the experimental settings:

  • Insufficient ablation studies: Appendix Table 7 lacks clear explanations of all configurations, particularly “random three checkpoints” and “shuffled checkpoints,” making the results difficult to assess. More importantly, the authors do not isolate the contributions of SEC, CCC, and VEC to UCW selection.

  • Unclear "utility" evaluation: The Utility (%) metric in Table 3 is not clearly defined. Since TrajGuard primarily monitors and aborts generation without modifying the editing process, the authors should explain how utility is measured and what is the meaning of the reported values.

  • Questionable adaptive-attack comparison: Figure 5 and Appendix Table 8 report an “oracle-score (Q=64Q=64)” attack across multiple defenses, but S(x)S(x) is defined only for TrajGuard. The feedback signals for other baselines remain unspecified, raising concerns about fairness of the comparison.

Questions for authors
  • From the paper, the test set is relatively small, with 144 held-out pairs in SafetyPairs-Traj (Section 4.1) and 208 external pairs in SafeEditBench-Traj (Section 6.4). I wonder whether this evaluation is sufficient to demonstrate generalization. How well would TrajGuard perform on larger and more diverse datasets?

  • Other questions, see weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.