acceptodds
Under review as a conference paper at ICLR 2027

Anchor-Conditioned On-Policy Distillation for Concept Erasure in Text-to-Image Generative Models

Abstract

Concept erasure in text-to-image (T2I) generative models aims to suppress specified concepts while preserving non-target generation capabilities. However, existing methods face a persistent trade-off between effective erasure and non-target preservation. Motivated by recent advances in on-policy distillation (OPD), we investigate whether supervision at student-visited states can improve this trade-off. We first apply OPD using existing concept-erased models as frozen teachers and find that it strengthens target suppression while largely preserving non-target generation capabilities across multiple erasure methods. Although effective, this procedure requires a separately trained concept-erased teacher. We propose Anchor-Conditioned On-Policy Distillation (AC-OPD), which instead uses a frozen copy of the original pretrained T2I model conditioned on a non-target anchor prompt. Specifically, the student generates trajectories under a target prompt specifying the concept to be erased and matches the frozen model’s anchor-conditioned transition means at student-visited states within an early generation window. Restricting supervision to this window yields a better erasure–retention trade-off than distillation over the full trajectory. Experiments on nudity and IP-character erasure demonstrate strong target suppression while largely preserving non-target text–image alignment and compositional generation capabilities. Compared with off-policy anchor-conditioned distillation, AC-OPD better preserves generation performance and reduces sensitivity to anchor choice. Code is available at https://anonymous.4open.science/r/AC-OPD-06F5.

Primary area:
generative models
Keywords:
Concept Erasure, On-Policy Distillation, Text-to-Image Generation, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZV5mOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the ''concept erasing'' problem in text-to-image generation models, aiming to suppress specific target concepts while preserving the model's general generation capabilities. The authors introduce On-Policy Distillation (OPD) for concept erasing: a ''teacher'' model is frozen after being trained via existing erasing methods, while a ''student'' model—initialized from the original pre-trained model—generates images using the target prompt. The student then aligns its outputs with the teacher's predictions for the same target prompt at the states it visits. OPD further improves the trade-off between target suppression and generation fidelity across various existing methods. Furthermore, the authors propose Anchor-Conditioned Policy Distillation (AC-OPD), which employs a specially pre-trained teacher. While the student generates images using the target prompt, AC-OPD conditions the teacher using a non-target ''anchor prompt'' and requires the student to match the teacher's anchor-conditioned output at the states it visits; the authors implement this supervision constraint during the early stages of the generation process. Experiments covering nudity suppression and IP-based character erasure demonstrate that AC-OPD effectively suppresses target concepts while better preserving text-image alignment and compositional generation capabilities. the method was validated on FLUX and Z-Image Turbo.

Critical strengths
  1. The paper is generally clear and readable. The overall presentation is well organized, and the figures and tables help communicate the main idea and experimental findings effectively. Despite some terminology that could be clarified, the manuscript is overall easy to follow.

  2. The proposed method is simple and easy to implement, and the paper provides rich ablations. The authors examine several important design choices, including on-policy versus off-policy supervision, anchor selection, supervision horizon, and adapted blocks, which helps clarify the behavior of the proposed method.

Critical weaknesses
  1. The methodological novelty appears limited relative to the closest existing ingredients. The erased-teacher OPD formulation largely overlaps with the objective already present in FlowErase-OPD, while the main additional idea in AC-OPD is to replace the erased-teacher prediction with an anchor-conditioned prediction from the original model at the same student-visited state, which is not suprising. Since target-to-anchor distribution alignment is already the central principle, AC-OPD appears to be a relatively direct combination of anchor-based erasure and on-policy distillation. The paper does not provide sufficient theoretical evidence that this combination. This concern is compounded by the absence of a controlled comparison with FlowErase-OPD (or an appropriately adapted single-concept variant) under a matched experimental setting. As a result, I find it difficult to assess whether the contribution is sufficiently distinct from closely related existing methods.

  2. The preservation evaluation is insufficient to substantiate the claimed erasure–retention trade-off. The paper evaluates retention using global metrics such as COCO text–image alignment and GenEval, but does not systematically evaluate semantically related non-target concepts. This omission is important for AC-OPD because the method explicitly redirects transitions toward anchor-conditioned predictions, which may induce suppression of neighboring concepts or anchor-specific semantic replacement.

Questions for authors

None. My main concerns are already described in the Critical Weaknesses section and do not hinge on clarification of the current manuscript.

4 comments
AuthorsOpenReview ↗

Official Response to Reviewer ZV5m (1/2)

[W1] Limited novelty relative to FlowErase-OPD and anchor-based erasure; missing matched comparison and justification for the combination.


Publication timeline. FlowErase-OPD was first released on arXiv on August 7, 2026, less than two months before the ICLR 2027 submission deadline, and, to our knowledge, has not yet appeared in a peer-reviewed venue. We therefore regard it as concurrent work, and it was already cited and distinguished from AC-OPD in the Related Work section of our submitted manuscript.

Methodological distinction from FlowErase-OPD. FlowErase-OPD focuses on multi-concept erasure by distilling multiple pretrained single-concept erased teachers into a unified LoRA. It additionally employs the frozen original model as an auxiliary retention teacher on general prompts and adaptively balances erasure and retention objectives.

In contrast, our work first investigates OPD as a post-training approach for individual erased models, demonstrating improvements in erasure effectiveness and generation retention across four erasure methods on FLUX.1-dev, and further demonstrating its applicability to Z-Image Turbo (Tables 1, 2, and 7 of our paper). Building on these findings, we introduce AC-OPD, which directly uses the original model's anchor-conditioned transitions as erasure supervision at target-conditioned student states, rather than relying on separately trained erased teachers. Thus, although both approaches employ OPD, they differ in their research objectives and the construction and role of their supervisory signals.

Controlled comparison. No official code or checkpoints for FlowErase-OPD are publicly available, so a matched experimental comparison is not possible at this time. The closest comparison available is already in our paper: the erased-teacher OPD students in Tables 1 and 2 represent single-teacher, single-concept OPD-based erasure without FlowErase-OPD's multi-teacher consolidation and retention teacher, and are trained and evaluated under the same setting as AC-OPD.

Without requiring any erased teacher, AC-OPD achieves comparable or stronger target suppression, including the lowest average Ring-A-Bell ASR (1.76%), UnlearnDiffAtk ASR (19.35%), and average IP detection count (0.60), while maintaining comparable GenEval (Tables 1 and 2). Furthermore, Table 7 demonstrates its applicability beyond FLUX.1-dev, where AC-OPD improves both target suppression and generation retention over Z-Erase on Z-Image Turbo.

Why the combination is not a direct consequence of its ingredients. Although AC-OPD builds on anchor-based erasure and on-policy distillation, their combination does not inherently guarantee effective erasure and generation retention. Our controlled experiments demonstrate that both the source of training states and the supervisory objective are critical to achieving this balance.

As shown in Table 4 and revised Section 5.2, under identical anchor-conditioned supervision, using frozen-model anchor-conditioned trajectories degrades generation retention (GenEval 55.24 vs. 65.04), whereas using frozen-model target-conditioned trajectories yields substantially weaker suppression (Ring-A-Bell ASR 17.63% vs. 2.19%). In contrast, AC-OPD achieves strong suppression and retention by supervising the current student's target-conditioned trajectories.

As shown in the objective ablation in Appendix B.3 of the revised manuscript, even with on-policy sampling, the choice of supervisory objective substantially affects generation retention. Repulsion-only and attraction-plus-repulsion achieve 0.00% Ring-A-Bell ASR but reduce GenEval to 14.48 and 58.30, respectively. In contrast, our attraction-only AC-OPD achieves strong suppression while preserving GenEval (65.86) and COCO (30.72).

Overall, these results demonstrate that effective concept erasure with generation retention does not follow automatically from on-policy sampling or strong suppression objectives alone. Rather, it depends on the combination of student-generated training states and appropriately designed anchor-conditioned supervision.

AuthorsOpenReview ↗

Official Response to Reviewer ZV5m (2/2)

[W2] Insufficient evaluation of related-concept preservation and risk of anchor-specific semantic replacement.


We recognize that general preservation metrics such as COCO and GenEval may not fully capture unintended degradation of concepts semantically related to the erased target. To address this concern, we conduct two additional evaluations of non-target concept preservation.

First, we evaluated cross-IP preservation using four non-target characters after erasing each of Pikachu, Snoopy, and SpongeBob, with 250 images per character (1,000 per erased target). As these characters belong to the semantic category of our anchor prompt ("A photo of a cartoon character"), this evaluation also examines whether anchor-conditioned supervision suppresses neighboring concepts.

Second, we evaluated related-concept preservation, following HUB's pinpoint-ness evaluation protocol, which assesses unintended erasure of non-target concepts close to the erased target in CLIP text-embedding space. We selected three semantically related WordNet concepts for each erased target based on CLIP text-embedding similarity (e.g., minion for Pikachu and sponge for SpongeBob) and generated 100 images per concept. In both evaluations, Qwen3.5-9B assessed concept presence, and retention was defined as the percentage of positive judgments.

For cross-IP preservation, each entry reports the target detection rate (%) / non-target retention rate (%).

Erased targetFLUX.1-devEAEraseFlowEraseFlow + OPDAC-OPD (Ours)
Pikachu—/96.72.7/80.42.6/80.61.0/85.90.0/90.0
Snoopy—/96.10.0/90.10.0/86.40.0/93.10.0/95.0
SpongeBob—/95.70.0/87.30.0/78.30.0/86.00.0/90.5
Average—/96.20.9/85.90.9/81.80.3/88.30.0/91.8

Cross-IP Preservation. As shown in the table above, applying OPD to EraseFlow improves non-target character retention while reducing the target detection rate, demonstrating that OPD also benefits in-domain concept preservation beyond general generation quality. AC-OPD achieves the highest retention among the erasure methods (91.8%, compared with 96.2% for the original model), with zero target detections across all three settings.

For related-concept preservation, we report retention rates (%) and their changes relative to FLUX.1-dev in percentage points.

Erased targetFLUX.1-devEAEraseFlowEraseFlow + OPDAC-OPD (Ours)
Pikachu98.387.0(−11.3)56.7(−41.7)86.0(−12.3)91.7(−6.7)
Snoopy96.773.7(−23.0)73.0(−23.7)87.7(−9.0)92.7(−4.0)
SpongeBob91.339.3(−52.0)53.3(−38.0)68.0(−23.3)76.0(−15.3)
Average95.466.7(−28.8)61.0(−34.4)80.6(−14.9)86.8(−8.7)

Related-Concept Preservation. Consistent with the cross-IP preservation results, applying OPD to EraseFlow substantially improves the retention of semantically related concepts (61.0% → 80.6%), suggesting that OPD also helps mitigate collateral erasure of neighboring concepts. AC-OPD achieves the highest average retention among the erasure methods (86.8%). Some degradation remains relative to the original model (−8.7 pp on average and −15.3 pp for SpongeBob), but it is smaller than for the compared methods in every setting.

Anchor-specific semantic replacement. Regarding the concern that redirecting transitions toward anchor-conditioned predictions may replace outputs with anchor content, Fig. 3 and Figs. 22–23 in Appendix B.2 of our paper show that AC-OPD outputs are not dominated by the anchor content: even with semantically distant object anchors, AC-OPD continues to generate human subjects with limited anchor-related visual features.

We have incorporated the cross-IP and related-concept preservation evaluations and their analysis into Section 5.5 of the revised manuscript.

Fine, I have no further concern. Although I view the innovation in this work as being more incremental in nature, the experiments are undeniably extensive. In my opinion, the paper meets the requirements for acceptance at ICLR, so I have raised my score : )

Dear Reviewer ZV5m,

We sincerely thank the reviewer for their positive feedback and for reconsidering their assessment of our work. We are pleased that our responses have addressed the reviewer’s concerns. We would be happy to answer any further questions.

Best regards,

The Authors

Reviewer_oJXyOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates on-policy distillation (OPD) for concept erasure in text-to-image models, where “on-policy” means that supervision is applied at states visited by the student’s own current generation trajectory. The authors first show that OPD can improve existing concept-erased models by using them as teachers. They then propose AC-OPD, which avoids the need for an erased teacher: the student generates under the target concept, while the original frozen model provides supervision under a non-target anchor prompt at the same student-visited states. Applying this supervision during early generation steps achieves strong concept suppression while largely preserving general generation quality. The paper propose that training on the model tha

Critical strengths

The paper is clearly written and presents a simple and intuitive approach to concept erasure. The empirical results are strong and consistent across several existing erasure methods, showing that on-policy distillation can improve target suppression while maintaining or even recovering generation quality. The evaluation is reasonably comprehensive, covering nudity and IP-character erasure, several adversarial attacks, and two different model architectures. The comparison against off-policy variants and the analysis of anchor sensitivity also provide useful evidence for the proposed design choices.

Critical weaknesses
  • The novelty of the “on-policy” component is not sufficiently established relative to prior concept-erasure methods. The paper presents supervision at student-visited states as a central conceptual contribution and contrasts this with teacher-generated trajectories. However, ESD [1] already generates partially denoised states using the current edited model and queries the frozen original model at those states to construct its erasure target. Under the operational definition used in this paper, this already resembles on-policy erasure. The contribution would therefore be clearer if the paper distinguished the source of training states from the construction of the supervisory signal, rather than treating student-generated trajectories themselves as the main novelty.

  • The comparison to prior erasure objectives is incomplete, particularly for recent rectified-flow methods. AC-OPD combines target-conditioned student trajectories with attraction toward an anchor-conditioned prediction from the frozen original model. This is closely related to Concept Ablation’s anchor-based redirection, while GEM [2] is an especially relevant recent method that combines anchor attraction and target repulsion through early timestep teacher-guided velocity matching in rectified flows (the first 10 timesteps are explicitly identified there as strong choices for effective concept erasure). GEM does not appear to be discussed or cited. This omission makes it difficult to determine whether AC-OPD introduces a fundamentally new erasure principle or rather a different combination of already established ingredients.

  • More generally, the binary distinction between “on-policy” and “off-policy” obscures several independent design choices in concept-erasure methods. Existing approaches differ in (i) which model generates the training states, (ii) whether trajectories are generated under the target, anchor, or another condition, (iii) which model supplies the supervisory prediction, and (iv) how the erasure signal itself is constructed. For example, ESD primarily uses repulsion from the target concept, Concept Ablation uses attraction toward an anchor, and GEM combines attraction and repulsion. A structured analysis along these axes would make the relationship to prior work substantially clearer and would better isolate the actual contribution of AC-OPD.

  • The comparison between AC-OPD and AC-Off-Policy does not appear to isolate trajectory source cleanly. In AC-OPD, states are generated by the current student under the target prompt, whereas the off-policy counterpart generates trajectories using the frozen original model under the anchor prompt. Thus, both the model generating the states and the conditioning of the rollout change simultaneously. Consequently, improvements cannot be attributed uniquely to on-policy sampling. A cleaner comparison would hold the rollout condition fixed and vary only whether states are sampled from the current student or a frozen reference model.

  • Several important ingredients are introduced together, including student-generated trajectories, anchor-conditioned targets, an early supervision window, and early-block adaptation. Although the paper contains useful ablations, these experiments do not fully disentangle whether the gains arise from on-policy state coverage or from the particular form of the supervisory signal. A particularly informative comparison would evaluate attraction-only, repulsion-only, and attraction-plus-repulsion objectives on the same student-generated trajectories.

  • Overall, the empirical results are strong, but the conceptual contribution appears narrower than the current framing suggests. Student-generated trajectories have precedent in concept erasure, while anchor-based redirection is also well established.

[1] Gandikota R, Materzynska J, Fiotto-Kaufman J, Bau D. Erasing concepts from diffusion models. InProceedings of the IEEE/CVF international conference on computer vision 2023 [2] Grebe JH, Braun T, Rohrbach A, Rohrbach M. GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows. InForty-third International Conference on Machine Learning.

Questions for authors
  1. How does AC-OPD compare experimentally with GEM, given its close relevance as a recent concept-erasure method for rectified-flow models? Could the authors also clarify the methodological distinction between the two approaches?

  2. In AC-OPD, the reference model is queried under the anchor condition at states generated by the target-conditioned student. Why should the reference model provide a reliable supervision signal at states that may be outside the trajectories it would naturally visit under the anchor condition?

  3. Could the authors disentangle the benefit of early-step supervision from the normalization of the loss? Since the objective is averaged over the supervised transitions, extending the horizon reduces the relative contribution of the early steps. Would the advantage of early supervision remain if their weighting were held fixed?

  4. How localized is the erasure? In particular, could the authors evaluate whether AC-OPD degrades generation of concepts that are semantically related to the erased concept or to the chosen anchor? Current preservation metrics such as COCO and GenEval may not capture this type of collateral effect.

10 comments
AuthorsOpenReview ↗

Official Response to Reviewer oJXy (1/8)

[W1] On-policy novelty relative to ESD: distinguishing the source of training states from the supervisory signal.


We acknowledge that ESD also generates training states using the evolving edited model. Following the reviewer's suggestion, we distinguish the source of training states from the construction of the supervisory signal.

Source of training states. ESD and OPD share this aspect: both obtain states from the current model under the target prompt. In ESD, however, model-generated states serve as a substitute for training data, thereby eliminating the need for data collection", and their effect on erasure is not examined. Our work treats the state source as a variable and isolates its effect. With the supervision held fixed, sampling states from the current student rather than from a frozen model substantially improves erasure, both with pretrained erased teachers (Table 3) and within AC-OPD, where Ring-A-Bell ASR decreases from 17.63% to 2.19% (response to W4).

Supervisory signal. ESD supervises one partially denoised state per update and constructs a negative-guidance target from the frozen original model's conditional and unconditional predictions. In contrast, OPD performs transition matching across consecutive early steps of each target-conditioned student trajectory, and its target is the transition that a reference model predicts at the same state: an erased teacher in OPD and the anchor-conditioned original model in AC-OPD. ESD is also one of the erased teachers in our experiments, and applying OPD to it further improves both erasure and retention, reducing UnlearnDiffAtk ASR from 70.97% to 22.58% while increasing GenEval from 60.38 to 63.20 (Table 1).

Our contribution lies not in student-generated state sampling alone, but in establishing OPD as an effective post-training framework for concept erasure, extending it to erased-teacher-free AC-OPD, and empirically validating the benefits of on-policy supervision through controlled comparisons.

We have clarified these distinctions in the Related Work section of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer oJXy (2/8)

[W2 & Q1, Part 1/2] Methodological overlap with CA and GEM, and the need for clearer distinctions and comparisons.


We thank the reviewer for highlighting the methodological similarities with Concept Ablation (CA) and GEM.

Distinction from Concept Ablation. Although CA and AC-OPD share the principle of anchor-based attraction, they differ in their training states and supervision objectives. CA matches target- and anchor-conditioned noise predictions at forward-noised anchor-concept images, whereas AC-OPD matches anchor-conditioned transitions from the frozen original model at target-conditioned student-generated states. We discussed these differences in the Related Work section and Section 3.3 of our submission.

Publication timeline. GEM was first released on arXiv on May 29, 2026, and published at ICML 2026 in July, approximately 2.5 months before the ICLR 2027 submission deadline. This is within the four-month window that the ICLR 2027 Area Chair Guidelines regard as contemporaneous. Despite its recency, we acknowledge its methodological relevance and provide detailed comparisons below.

Distinctions between GEM and AC-OPD. Although GEM and AC-OPD both employ anchor-conditioned supervision and early-step guidance, they differ in their starting point, in how the choice of training states is justified, and in whether target repulsion and its coefficient η\eta are required.

Scope of contributions. GEM builds on EraseFlow's GFlowNet-based Trajectory Balance objective to derive a new erasure objective. In contrast, our work first investigates the application of on-policy distillation (OPD) to concept erasure using existing erased models as teachers (Tables 1, 2, and 7). This establishes the effectiveness and generality of OPD across different erasure methods and is independent of GEM's formulation. We then develop AC-OPD, which removes the dependence on pretrained erased teachers by directly distilling anchor-conditioned transitions from the frozen original model.

Trajectory selection and on-policy supervision. GEM derives its velocity-matching objective from anchor-conditioned trajectories (Steps 1–4 of GEM's methodology) but subsequently adopts target-conditioned student trajectories and frozen-teacher contrastive guidance in Step 5. Although GEM provides an intuitive motivation, these changes are not directly derived from the preceding formulation. GEM also identifies a systematic analysis of trajectory selection as a direction for future work.

In contrast, AC-OPD formulates anchor-conditioned transition matching directly on target-conditioned student trajectories within the OPD framework, where supervising states visited by the current student is the defining principle. We investigate the effect of on-policy sampling with pretrained erased teachers (Table 3 of our paper) and further isolate its contribution to AC-OPD through controlled comparisons (response to W4). With the rollout condition and supervision held fixed, sampling from the evolving student reduces Ring-A-Bell ASR from 17.63% to 2.19% while maintaining comparable generation retention.

Role of repulsion and adapted blocks. GEM introduces target repulsion to enhance concept erasure, and its ablation study on the nudity target shows limited suppression with anchor attraction alone (Ring-A-Bell 67.02% at η\eta=0, compared with 63.86% for the original model; GEM, Table 7). It describes the repulsion coefficient η\eta as its main lever for adapting to each erasure scenario and reports η\eta=1 for nudity, 2 for characters, 5 for celebrities, 0.2 on SD3, and 0.3 for a longer nudity prompt. In contrast, AC-OPD uses anchor attraction without target repulsion and therefore has no such coefficient. The experiments below show that this attraction-only supervision, which is insufficient in GEM's nudity ablation, achieves strong erasure while preserving generation capabilities under early single-stream adaptation.

[W2 & Q1, Part 2/2] Methodological overlap with CA and GEM, and the need for clearer distinctions and comparisons.


Experimental comparison with GEM.

We conduct additional experiments using GEM's official implementation on FLUX.1-dev under our evaluation protocol. We compare its released checkpoint and analyze the effects of adapted blocks, repulsion coefficient η\eta, and target–anchor prompts. The GEM prompts are its original target–anchor pair ("nudity" and "fully dressed"), and the Ours prompts follow EraseAnything, as described in Appendix A.3 of our paper.

MethodAdapted blocksηPromptsRing-A-Bell Avg. (%) ↓MMA (%) ↓UnlearnDiffAtk (%) ↓COCO ↑GenEval ↑
GEM (official)Dual-stream1GEM24.212.0035.4830.0254.39
GEMDual-stream1Ours48.4211.2077.4230.5962.45
GEMEarly single-stream1GEM3.513.2022.5827.3324.26
GEMEarly single-stream1Ours0.700.603.2329.4948.08
GEMEarly single-stream0Ours1.402.7019.3530.7265.72
AC-OPDEarly single-stream–Ours1.761.9019.3530.7265.86

Analysis of experimental results. Compared with GEM's official checkpoint, AC-OPD achieves stronger erasure and higher generation retention. Our additional experiments show that applying AC-OPD's early single-stream adaptation strategy to GEM substantially lowers Ring-A-Bell and UnlearnDiffAtk attack success rates but degrades retention under GEM's contrastive objective (GenEval 24.26 and 48.08 with η\eta=1). Removing target repulsion restores retention (GenEval 65.72) while maintaining strong erasure, yielding performance comparable to AC-OPD, which uses the same attraction-only signal. Although repulsion further strengthens erasure against the evaluated attacks, it introduces a trade-off with generation retention.

We also observe that GEM's erasure effectiveness varies across target–anchor prompt pairs under a fixed η\eta. With η\eta=1, changing the prompt pair raises Ring-A-Bell ASR from 24.21 to 48.42 under dual-stream adaptation but lowers it from 3.51 to 0.70 under early single-stream adaptation. GEM's behavior at a fixed η\eta therefore depends on both the prompt pair and the adapted blocks, suggesting that η\eta requires configuration-dependent tuning. This is consistent with GEM's Appendix D.2, where a different η\eta is preferred for a longer target prompt. In contrast, AC-OPD eliminates the need to tune this coefficient by omitting the repulsion term.

Overall, these findings highlight the importance of adaptation configuration and demonstrate that effective anchor-conditioned erasure does not necessarily require explicit target repulsion.

We have incorporated the comparison with GEM and the additional experimental results into Appendix B.1 of the revised manuscript.

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

This paper focuses on the on-policy distillation (OPD) for concept erasure in T2I generative models. They propose the Anchor-Conditioned On-Policy Distillation (AC-OPD), which matches anchor-conditioned transitions from the frozen original T2I model at target-conditioned student states within an early generation window, without requiring a separately trained erased teacher. Experiments on nudity and IP character erasure demonstrate effective target suppression while largely preserving non-target text–image alignment and compositional generation capabilities of their method.

Critical strengths

1.​ OPD consistently improves target suppression and compositional generation across multiple concept-erased models. 2. The formulation is simple and well-motivated. The anchor-conditioned predictions from the original model supervise target-conditioned student trajectories, eliminating the need for an erased-teacher training stage. The analysis of on-policy v.s. off-policy supervision, anchor sensitivity, adapted blocks, and supervision horizons is useful. 3. This paper is clear and easy to understand.

Critical weaknesses
  1. The novelty of the proposed method is limited. GEM [1] already adopts a frozen pretrained model to provide anchor-conditioned velocity guidance for concept erasure in rectified-flow models, including supervision at early generation stages, which substantially overlaps with the core formulation of AC-OPD. Moreover, FlowErase-OPD [2] uses OPD to concept erasure in flow-matching models, showing that student-trajectory-based distillation for concept erasure itself is not new. Although AC-OPD introduces different combinations of anchoring conditional supervision and policy sampling without the need for separately trained erasing teachers, the fundamental methodological progress compared with these closely related methods has still not been fully established. Therefore, it is very difficult to assess the actual novelty of the proposed method. To verify the novelty of this paper, it is necessary to conduct a direct comparison with these methods. [1] GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows, ICML 2026. [2] FlowErase-OPD: Multi-Concept Erasure via Anchored On-Policy Distillation in Flow Matching Models, arXiv 2026.
  2. The on-policy contribution is not cleanly isolated.​ In Sec.5.2, AC-Off-Policy samples trajectories from the frozen original model under the anchor prompt, whereas AC-OPD samples from the current student under the target prompt. This changes both the sampling policy and the conditioning distribution simultaneously, so the preservation gap in Tab 4 may reflect different rollout conditions rather than on-policy adaptation itself. A baseline, i.e., off-policy trajectories from the frozen original model under the target prompt, with identical anchor-conditioned reference predictions and optimization settings, would directly isolate the benefit of sampling from the evolving student distribution.
  3. Non-target preservation metrics may miss semantic collateral damage.​ The main claim of AC-OPD is effective concept erasure with retained non-target generation. However, the preservation is primarily measured by COCO CLIP scores and GenEval. These benchmarks capture general text–image alignment and compositional generation but may not reveal degradation of concepts semantically adjacent to the erased target, e.g., visually similar IP characters or benign human attributes under nudity suppression, making the claimed erasure–retention trade-off under-supported.
  4. IP-character evaluation relies on a single multimodal judge.​ With uncertain judgments treated as negative detections and several concepts reporting near-zero detection counts, it remains unclear whether the results reflect successful erasure or evaluator uncertainty. Additional validation (e.g., human evaluation or prompt-level breakdown) would strengthen confidence in the IP erasure results.
Questions for authors

Please refer to the weakness.

6 comments
AuthorsOpenReview ↗

Official Response to Reviewer jcaa (1/4)

[W1, Part 1/2] Novelty and Direct Comparison with GEM and FlowErase-OPD


We thank the reviewer for highlighting these closely related works.

Publication timeline. GEM was first released on arXiv on May 29, 2026, and published at ICML 2026 in July, approximately 2.5 months before the ICLR 2027 submission deadline. This is within the four-month window that the ICLR 2027 Area Chair Guidelines regard as contemporaneous. Despite its recency, we acknowledge its methodological relevance and provide detailed comparisons below. FlowErase-OPD was released on arXiv on August 7, 2026, and, to our knowledge, has not yet appeared in a peer-reviewed venue. As concurrent work, it was already cited and distinguished from AC-OPD in the Related Work section (Section 2) of our submitted manuscript. We briefly reiterate these distinctions below.

Distinctions between GEM and AC-OPD. Although GEM and AC-OPD both employ anchor-conditioned supervision and early-step guidance, they differ in their starting point, in how the choice of training states is justified, and in whether target repulsion and its coefficient η\eta are required.

Scope of contributions. GEM builds on EraseFlow's GFlowNet-based Trajectory Balance objective to derive a new erasure objective. In contrast, our work first investigates the application of on-policy distillation (OPD) to concept erasure using existing erased models as teachers (Tables 1, 2, and 7). This establishes the effectiveness and generality of OPD across different erasure methods and is independent of GEM's formulation. We then develop AC-OPD, which removes the dependence on pretrained erased teachers by directly distilling anchor-conditioned transitions from the frozen original model.

Trajectory selection and on-policy supervision. GEM derives its velocity-matching objective from anchor-conditioned trajectories (Steps 1–4 of GEM's methodology) but subsequently adopts target-conditioned student trajectories and frozen-teacher contrastive guidance in Step 5. Although GEM provides an intuitive motivation, these changes are not directly derived from the preceding formulation. GEM also identifies a systematic analysis of trajectory selection as a direction for future work.

In contrast, AC-OPD formulates anchor-conditioned transition matching directly on target-conditioned student trajectories within the OPD framework, where supervising states visited by the current student is the defining principle. We investigate the effect of on-policy sampling with pretrained erased teachers (Table 3 of our paper) and further isolate its contribution to AC-OPD through controlled comparisons (response to W2). With the rollout condition and supervision held fixed, sampling from the evolving student reduces Ring-A-Bell ASR from 17.63% to 2.19% while maintaining comparable generation retention.

Role of repulsion and adapted blocks. GEM introduces target repulsion to enhance concept erasure, and its ablation study on the nudity target shows limited suppression with anchor attraction alone (Ring-A-Bell 67.02% at η\eta=0, compared with 63.86% for the original model; GEM, Table 7). It describes the repulsion coefficient η\eta as its main lever for adapting to each erasure scenario and reports η\eta=1 for nudity, 2 for characters, 5 for celebrities, 0.2 on SD3, and 0.3 for a longer nudity prompt. In contrast, AC-OPD uses anchor attraction without target repulsion and therefore has no such coefficient. The experiments below show that this attraction-only supervision, which is insufficient in GEM's nudity ablation, achieves strong erasure while preserving generation capabilities under early single-stream adaptation.

[W1, Part 2/2] Novelty and Direct Comparison with GEM and FlowErase-OPD


Experimental comparison with GEM.

We conduct additional experiments using GEM's official implementation on FLUX.1-dev under our evaluation protocol. We compare its released checkpoint and analyze the effects of adapted blocks, repulsion coefficient η\eta, and target–anchor prompts. The GEM prompts are its original target–anchor pair ("nudity" and "fully dressed"), and the Ours prompts follow EraseAnything, as described in Appendix A.3 of our paper.

MethodAdapted blocksηPromptsRing-A-Bell Avg. (%) ↓MMA (%) ↓UnlearnDiffAtk (%) ↓COCO ↑GenEval ↑
GEM (official)Dual-stream1GEM24.212.0035.4830.0254.39
GEMDual-stream1Ours48.4211.2077.4230.5962.45
GEMEarly single-stream1GEM3.513.2022.5827.3324.26
GEMEarly single-stream1Ours0.700.603.2329.4948.08
GEMEarly single-stream0Ours1.402.7019.3530.7265.72
AC-OPDEarly single-stream–Ours1.761.9019.3530.7265.86

Analysis of experimental results. Compared with GEM's official checkpoint, AC-OPD achieves stronger erasure and higher generation retention. Our additional experiments show that applying AC-OPD's early single-stream adaptation strategy to GEM substantially lowers Ring-A-Bell and UnlearnDiffAtk attack success rates but degrades retention under GEM's contrastive objective (GenEval 24.26 and 48.08 with η\eta=1). Removing target repulsion restores retention (GenEval 65.72) while maintaining strong erasure, yielding performance comparable to AC-OPD, which uses the same attraction-only signal. Although repulsion further strengthens erasure against the evaluated attacks, it introduces a trade-off with generation retention.

We also observe that GEM's erasure effectiveness varies across target–anchor prompt pairs under a fixed η\eta. With η\eta=1, changing the prompt pair raises Ring-A-Bell ASR from 24.21 to 48.42 under dual-stream adaptation but lowers it from 3.51 to 0.70 under early single-stream adaptation. Therefore, GEM's behavior at a fixed η\eta depends on both the prompt pair and the adapted blocks, suggesting that η\eta requires configuration-dependent tuning. This is consistent with GEM's Appendix D.2, where a different η\eta is preferred for a longer target prompt. In contrast, AC-OPD eliminates the need to tune this coefficient by omitting the repulsion term.

Overall, these findings highlight the importance of adaptation configuration and demonstrate that effective anchor-conditioned erasure does not necessarily require explicit target repulsion.

Relation to FlowErase-OPD. This distinction is already part of our submitted manuscript. As discussed in our Related Work section, FlowErase-OPD is concurrent work that distills multiple pretrained erased teachers into a unified LoRA, using the original model for auxiliary retention supervision. In contrast, our work investigates OPD as a post-training approach for individual erased teachers and extends it to AC-OPD, which requires no separately trained erased teachers and instead uses the original model for anchor-conditioned erasure supervision. No official code or checkpoints for FlowErase-OPD are publicly available, so a matched experimental comparison is not possible at this time.

We have included a discussion of GEM in the Related Work section and incorporated the comparison with GEM and additional experimental results into Appendix B.1 of the revised manuscript.

AuthorsOpenReview ↗

Official Response to Reviewer jcaa (2/4)

[W2] Isolating the Effect of On-policy Sampling


We thank the reviewer for identifying this confounding factor. We acknowledge that the original AC-Off-Policy (anchor) baseline differs from AC-OPD in both the rollout policy and conditioning prompt, preventing a direct assessment of the on-policy contribution. To address this concern, we introduce AC-Off-Policy (target), the baseline proposed by the reviewer, which generates trajectories using the frozen original model under the target prompt while retaining the same anchor-conditioned reference model and optimization settings. It differs from AC-OPD only in whether training states are sampled from the frozen original model or the current student.

MethodRing-A-Bell (%) ↓MMA ASR (%) ↓COCO ↑GenEval ↑
AC-Off-Policy (anchor)2.46 ± 4.050.78 ± 1.1130.41 ± 0.3555.24 ± 9.73
AC-Off-Policy (target)17.63 ± 9.086.33 ± 3.5730.76 ± 0.0365.93 ± 0.71
AC-OPD (ours)2.19 ± 1.630.73 ± 0.7930.75 ± 0.0465.04 ± 1.16

Values are mean ± standard deviation across the four anchor prompts, as in Table 4.

Compared with AC-Off-Policy (target), AC-OPD substantially reduces attack success rates on both Ring-A-Bell (from 17.63% to 2.19%) and MMA-Diffusion (from 6.33% to 0.73%) while maintaining comparable COCO and GenEval performance. This suggests that on-policy supervision improves erasure effectiveness without compromising general generation capabilities. A similar trend is observed with existing erased teachers in Table 3 of our paper, where the off-policy baseline also uses target-conditioned trajectories (CA: Ring-A-Bell ASR of 14.03% versus 3.51%, with comparable GenEval performance).

AC-Off-Policy (target) also preserves GenEval (65.93), unlike AC-Off-Policy (anchor) (55.24). The preservation gap in the original Table 4 is associated with the anchor-conditioned rollout rather than with on-policy sampling. We have revised the corresponding statements in the abstract, Section 1, and Section 5.2 accordingly.

AC-OPD further exhibits reduced performance variability across the four anchor prompts, with the standard deviation decreasing from 9.08 to 1.63 for Ring-A-Bell ASR and from 3.57 to 0.79 for MMA ASR. This indicates that on-policy supervision yields more consistent erasure effectiveness across different anchor conditions.

Overall, these controlled comparisons isolate the contribution of on-policy state sampling and provide empirical evidence for its effectiveness in improving concept erasure while preserving general generation capabilities.

We have incorporated the controlled comparison and its analysis into Section 5.2 of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.