acceptodds
Under review as a conference paper at ICLR 2027

Beyond Text-to-Image: Unified Erasure of Sexually Explicit Content in Diffusion Models

Abstract

Diffusion models are increasingly weaponized for sexual abuse, enabling large-scale generation of illicit pornographic content and targeted harassment via fabricated sexualized media depicting real individuals. These threats span four distinct task categories, covering both image and video generation under text-only and image-conditioned inputs. Yet existing defenses and benchmarks are narrowly scoped to text-to-image pipelines, leaving the remaining attack surface exposed. We introduce UniErase-Bench, the first benchmark to systematically evaluate sexually explicit content generation across the full attack surface, and then present UniErase, a unified training-free framework that erases such concepts from the underlying architectures across all task categories. Our key insight is that the feed-forward network (FFN) layers in modern diffusion architectures function as key-value memory units, where sexually explicit concepts are encoded in a subset of channels. Leveraging this property, UniErase first localizes the channels encoding these concepts and then removes them via closed-form weight projection. Extensive experiments on UniErase-Bench demonstrate that UniErase consistently suppresses explicit content generation across all evaluated task categories while preserving the generation quality of benign content, achieving state-of-the-art performance on text-to-image generation and establishing the first unified defense covering the full attack surface. Our code is available at https://anonymous.4open.science/r/UniErase.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
text to image, concept erasure

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_JMY6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniErase, a training-free concept erasure method for suppressing sexually explicit content generation in diffusion models, and UniErase-Bench, a benchmark for evaluating such methods. UniErase identifies concept-related channels in generation-side FFNs and applies a closed-form weight projection to remove the corresponding concept subspace from their outputs. Since the weights are modified offline, the method introduces no additional computation at inference time. UniErase-Bench covers four generation tasks spanning text-only and image-conditioned inputs, with image or video outputs.

Critical strengths
  • By using FFNs as an editing target, UniErase demonstrates applicability across diffusion models based on U-Nets and transformers, as well as across all four generation tasks.

  • In Table 1, UniErase achieves the highest average ER among the evaluated methods at 94.48%, including an ER of 82.70% against the white-box attack P4D. Its CLIP score and FID on benign generation also remain close to those of the original model, suggesting a favorable balance between concept erasure and preservation of generation quality.

Critical weaknesses
  • The benchmark’s novelty is limited, and its coverage of attack strategies falls short of the claimed “full attack surface.” Its text-only subsets reuse I2P for T2I and T2VSafetyBench for T2V, while MMA-Diffusion [4] already provides both text- and image-based attacks and associated evaluation data. Although UniErase-Bench extends evaluation across four generation tasks, it does not include systematic evaluation tracks for black-box, white-box, or adaptive attacks, particularly optimized adversarial prompts, images and joint text–image attacks. The adversarial experiments in Table 1 are separate T2I evaluations and do not extend the scope of the benchmark itself. The paper should more clearly demonstrate what additional evaluation value the newly constructed image-conditioned subsets and unified four-task evaluation provide over existing benchmarks.

  • The algorithmic distinction from prior work requires clearer explanation and validation. Training-free direct weight editing has already been explored in UCE [1]. Bae et al. [2] analyze cross-attention OV subspaces and remove concept-related spectral components, while CAD [3] selectively removes parameters based on concept attribution. The paper should more clearly establish how UniErase’s FFN channel localization and concept-subspace projection differ algorithmically from these approaches, and what additional benefits these design choices provide.

Questions for authors
  • Why does UniErase-Bench not include evaluations for black-box and white-box adversarial attacks?

  • Have you analyzed the roles that the FFN channels identified by UniErase and their corresponding weights play during generation? Do these components contribute specifically to sexually explicit content, or do they also support the generation of other concepts or visual attributes?

  • Have the authors conducted a failure case analysis? What are the main failure modes? Figure 7 also shows suppression of edits involving swimwear and shirtless men. What criteria are used to distinguish sexually explicit content from benign content? Has over-erasure been evaluated for benign concepts that are semantically or visually related to the target concept?

4 comments
AuthorsOpenReview ↗

Response to Reviewer JMY6 1/3

W1 & Q1 Benchmark Scope and Novelty

We thank the reviewer for raising this concern and clarify that the "full attack surface" in our work refers to the four generation task categories defined by input conditioning and output modality, rather than exhaustive coverage of adversarial attack strategies.

Research scope. As also recognized by Reviewer WmXb, real-world misuse such as nudify applications and sextortion highlights the importance of image-conditioned and video generation. Motivated by the growing gap between this expanding threat surface and the T2I-centric focus of existing concept-erasure benchmarks, we introduce UniErase-Bench to provide a unified evaluation foundation for developing and validating our erasure framework. While MMA-Diffusion focuses on adversarial image perturbations in text-guided image editing to bypass post-hoc safety checkers, UniErase-Bench systematically covers four generation tasks and three complementary cross-modal attack scenarios, enabling comprehensive evaluation of risks arising from textual, visual, and joint conditioning.

Adversarial attacks. We agree that adversarial optimization of textual, visual, and joint inputs could yield stronger attacks and represents a promising research direction. However, developing systematic adversarial attacks across T→V, T+I→I, and T+I→V involves substantial task-specific challenges and warrants dedicated investigation. Therefore, in this work, we adopt existing black-box and white-box adversarial attacks to evaluate the robustness of UniErase. Building upon UniErase-Bench to explore adversarial attacks across these generation tasks is an important direction for our future work.

W2 Algorithmic Distinctions from Prior Work

We thank the reviewer for highlighting these relevant works. We acknowledge that our method draws inspiration from established techniques, including closed-form weight editing, concept attribution, and subspace-based concept removal. However, directly applying these techniques across heterogeneous diffusion architectures and generation tasks remains challenging. Our contribution lies not in simply combining existing techniques, but in carefully adapting and integrating them into a unified, selective concept-erasure framework. The key technical distinctions are as follows:

(1) Unified intervention space across architectures. UCE edits text-conditioned cross-attention weights, limiting its ability to address harmful concepts introduced through image conditioning, while Bae et al. focus on spectral analysis of cross-attention OV subspaces. In contrast, UniErase operates on generation-side FFNs, which process representations after conditioning information has been integrated, providing a shared intervention space across heterogeneous architectures and input modalities.

(2) Selective concept subspace editing. Unlike CAD, which directly ablates concept-attributed parameters, UniErase aims to preserve benign information entangled within concept-relevant FFN channels. Since FFNs lack directly accessible text-aligned concept directions, we develop contrastive subspace estimation to isolate target-specific directions from shared benign representations, enabling precise and selective weight editing.

Component-level ablation. To further demonstrate the significance of our key design choices, we conduct a component-level ablation on both SD-v1.4 and FLUX.1-dev using the adversarial RAB dataset. The results demonstrate that selective projection and contrastive refinement improve benign content preservation while maintaining strong erasure effectiveness.

Table 1: Component ablation on SD-v1.4 and FLUX.1-dev.

SD-v1.4FLUX.1-dev
VariantProjectionContrastive RefinementER ↑FID ↓ER ↑FID ↓
(a) Channel zeroing✗✗98.7426.1895.5436.55
(b) Target-only PCA✓✗97.6625.5696.5236.01
(c) UniErase (Full)✓✓98.8324.9596.4134.85
AuthorsOpenReview ↗

Response to Reviewer JMY6 2/3

Q2 Roles of Concept-Relevant FFN Channels

As discussed in Section 4.3, the identified FFN channels are concept-relevant rather than concept-exclusive, as they may also encode benign visual attributes such as skin texture and human anatomy. Directly zeroing these channels would therefore risk removing valuable benign information. To address this, UniErase employs contrastive subspace refinement to selectively remove target-specific directions while preserving other semantic information within the same channels. Our component-level ablation in Table 1 further supports this design, showing that UniErase achieves better benign content preservation than direct channel zeroing while maintaining strong erasure effectiveness across both U-Net and DiT architectures.

Q3 Failure Cases and Adjacent Benign Concepts.

We thank the reviewer for raising the important concern and further clarify the scope of our concept-erasure criteria below.

(1) Failure case analysis. Our failure case analysis shows that while UniErase effectively suppresses text-induced explicit generation, residual failures are more concentrated in S3, where both textual and visual conditions introduce harmful concepts. Such dual-modal reinforcement makes concept erasure more challenging and requires broader intervention to achieve stronger suppression. As shown in Table 2, increasing the channel budget K substantially reduces residual ASR with minimal impact on reference-image similarity, whereas a larger λ achieves stronger suppression at a greater preservation cost.

Table 2: Sensitivity analysis of K and λ on FLUX.1-Kontext (S3).

KλASR (%) ↓CLIP-I ↑
50535.099.21
75527.799.04
100522.699.22
50725.098.53
501012.395.37

(2) Erasure criteria. Our current implementation treats substantial exposure of sensitive anatomical regions, including the chest, genitals, buttocks, and other intimate body areas, as nudity-related explicit content. This operational definition follows the broader evaluation settings adopted in prior concept-erasure research [1–4], ensuring consistency with existing studies.

(3) Adjacent benign concepts. UniErase employs contrastive subspace refinement to preserve benign visual attributes, such as skin texture and everyday clothing, including conventional swimwear and non-explicit underwear.

However, borderline cases such as the "revealing bikini" in Figure 7(b) may expose substantial portions of sensitive body areas, potentially causing distress and reputational harm to the depicted individual, thus motivating our intervention. Similarly, while classical nude sculptures and men with exposed upper bodies may carry different meanings across cultural contexts, our current implementation follows the adopted nudity-related criteria[3, 4] in treating such representations as targets for erasure.

[1] Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models, CVPR 2023

[2] Erasing Concepts from Diffusion Models, ICCV 2023

[3] Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models, ECCV 2024

[4] Sculpting Memory: Multi-Concept Forgetting in Diffusion Models via Dynamic Mask and Concept-Aware Optimization, ICCV 2025

AuthorsOpenReview ↗

Response to Reviewer JMY6 3/3

Our Question

We noticed that the review uses numbered citations [1]–[4], but the corresponding reference list is missing. We have identified the relevant works based on their descriptions and hope that our interpretations are correct. If any references have been misidentified or our understanding differs from the reviewer's intent, we would appreciate further clarification and will revise our discussion accordingly.

Area_Chair_GdDzOpenReview ↗

Reviewer interaction needed

Could the reviewer please react to the author's question and to their rebuttal? Kind regards, The AC

Reviewer_9R82OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces UniErase-Bench, the first benchmark covering four categories of sexually explicit content generation tasks involving text- and image-conditioned image and video generation. It also proposes UniErase, a unified training-free defense framework that identifies and removes channels encoding sexually explicit concepts in diffusion model FFN layers through closed-form weight projection

Critical strengths
  1. Introduces UniErase-Bench, a comprehensive benchmark for systematically evaluating sexually explicit content generation across four distinct attack surfaces involving image and video generation.
  2. Proposes UniErase, a unified training-free framework that effectively suppresses sexually explicit content generation across all four attack surfaces while preserving benign content quality.
  3. Exploits the FFN layers to identify concept-associated channels and suppress targeted concepts through closed-form weight projection, providing an efficient mechanism for concept erasure.
Critical weaknesses
  1. Although the authors evaluate UniErase across several representative diffusion architectures, the main comparison with existing methods is conducted on the relatively outdated SD-v1.4 model. The authors are encouraged to include more recent generative models to further validate the effectiveness and generalizability of UniErase.
  2. Although the authors demonstrate negligible inference overhead, the FFN channel localization stage requires gradient-based computations across multiple condition pairs, random seeds, and denoising timesteps, potentially introducing substantial offline computational costs. The authors are encouraged to report the computational overhead of the entire model editing process, including channel localization and subspace refinement, to provide a more comprehensive evaluation of the method's efficiency.
  3. The current ablation studies primarily focus on hyperparameter sensitivity. A component-level ablation would be valuable. For example, comparing direct channel zeroing with subspace projection under identical channel selections would help isolate and validate the effectiveness of the proposed editing strategy.
  4. Although the paper provides descriptions of the compared methods in the appendix, some baselines lack clear references to their original publications. The authors are encouraged to include the publication years, venues, and links to the corresponding papers for all compared methods to improve clarity and facilitate the assessment of their recency and relevance.
Questions for authors

Please see the Critical Weaknesses for details.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 9R82

W1 Comparisons on Recent Diffusion Architectures

We thank the reviewer for this suggestion and agree that quantitative comparisons with existing methods on newer models would further strengthen our evaluation.

We therefore extend our evaluation on FLUX.1-dev to include UCE[1], ConceptPrune[2], EraseAnything[3], DVE[4], and GEM[5]. As shown in Table 1, UniErase achieves the highest ER and CLIP score, while maintaining competitive FID. These results further validate the effectiveness of UniErase on modern diffusion transformer architectures.

Table 1: Quantitative comparison of concept erasure methods on FLUX.1-dev.

MethodER ↑CLIP ↑FID ↓
UCE52.9330.9235.92
ConceptPrune92.9429.9950.23
EraseAnything30.7130.7934.18
DVE91.2929.9533.49
GEM80.4230.0241.35
UniErase (Ours)96.4130.9434.85

We have included the additional experimental results and detailed comparisons in Appendix F of the revised manuscript.

[1] Unified Concept Editing in Diffusion Models, WACV 2024

[2] ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning, ICLR 2025

[3] EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers, ICML 2025

[4] Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models, arXiv 2026

[5] GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows, ICML 2026

W2 Computational Overhead of the Editing Process

We thank the reviewer for this valuable suggestion. We agree that reporting the complete offline editing cost would provide a more comprehensive assessment of UniErase's computational efficiency.

Stage-wise computational analysis. (1) Channel localization requires forward and backward passes across multiple condition pairs and random seed, making it the primary gradient-dependent stage. (2) Subspace refinement requires forward passes to collect activation-weighted FFN outputs from N condition samples, followed by limited number of matrix operations on covariance computation. (3) Weight projection involves only a limited number of matrix operations on the selected FFN weights.

Offline editing cost. We provide a detailed breakdown of the computational costs associated with the three stages of UniErase. (1) Channel localization. We first conduct a stability analysis on SD-v1.4 using ten mutually disjoint subsets, each containing 10 prompt pairs. The five most frequently selected channels appear in 10, 10, 10, 9, and 9 subsets, respectively, demonstrating high localization stability. This high consistency motivates our use of only 10 target–safe pairs with 10 valid random seeds for efficient channel localization. (2) Subspace refinement. To obtain more representative activation statistics for reliable concept subspace estimation, we use 250 target–safe pairs with 2 random seeds, corresponding to 1,000 condition-specific forward, without backward propagation.

Importantly, all three stages are performed only once offline, introducing no additional inference-time computation.

W3 Component-level ablation

We thank the reviewer for this valuable suggestion. To validate the significance of our key design choices, we conduct a component-level ablation on SD-v1.4 and FLUX.1-dev using the adversarial RAB dataset.

As shown in Table 3, UniErase achieves higher ER and lower FID than direct channel zeroing on both architectures. Moreover, contrastive refinement further improves benign content preservation over target-only PCA while maintaining comparable erasure effectiveness, demonstrating the benefits of our selective editing design.

Table 3: Component ablation on SD-v1.4 and FLUX.1-dev.

SD-v1.4FLUX.1-dev
VariantProjectionContrastive RefinementER ↑FID ↓ER ↑FID ↓
(a) Channel zeroing✗✗98.7426.1895.5436.55
(b) Target-only PCA✓✗97.6625.5696.5236.01
(c) UniErase (Full)✓✓98.8324.9596.4134.85

We have included the additional experimental results and detailed comparisons in Appendix E of the revised manuscript.

W4 Baseline references

We thank the reviewer for pointing out the incomplete baseline references in our original submission. We have now added the corresponding citations to all baseline descriptions in the appendix. In the camera-ready version, we will further include the publication years, venues, and paper links directly in the main comparison table, as suggested, to improve clarity and facilitate comparison.

Area_Chair_GdDzOpenReview ↗

Reviewer interaction needed.

Dear reviewer, The authors have replied to your comments and the discussion is close to the end. We encourage you to participate in the discussion to see whether your questions and concerns have been answered by the authors. Thank you! Best, Your AC

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

This paper studies concept erasure of sexually explicit content in diffusion models. The authors first observe that existing defenses and benchmarks are almost exclusively scoped to text-to-image (T2I) pipelines, while the actual threat surface spans four task categories: text-to-image (T→I), image-conditioned editing (T+I→I), text-to-video (T→V), and image-to-video (T+I→V). Addressing this gap, the paper makes two contributions: UniErase-Bench: the first benchmark covering explicit content generation across the full attack surface. Its image-conditioned subsets are constructed from scratch under a unified LLM-assisted protocol with three cross-modal attack scenarios (text-led, visual jailbreak, and dual reinforcement), 300 instances per scenario per task; UniErase: a unified training-free erasure framework. The key insight is that FFN layers in diffusion models act as key-value memories, with the target concept encoded in a subset of channels. The method proceeds in three stages: (a) localize concept-relevant channels via target–safe contrastive condition pairs with first-order attribution; (b) extract the concept subspace from the selected channels' outputs via contrastive covariance analysis (Σ⁺ − λΣ⁻) followed by eigendecomposition; (c) remove the concept component via a closed-form orthogonal projection v′ = v(I − UUᵀ). Experiments cover six models: SD-v1.4, SDXL, SD3, FLUX.1-dev, FLUX.1-Kontext, and HunyuanVideo-1.5. On T2I, UniErase achieves an average erasure rate of 94.48% across six benchmarks, surpassing 12 baselines; under the white-box attack P4D it attains the best ER of 82.70%; benign generation quality (CLIP/FID) remains nearly identical to the unedited model, with no inference overhead.

Critical strengths
  1. Timely and important problem definition, with a compelling threat-surface characterization. The paper's argument for the gap between an expanding attack surface and narrow defense coverage is clear: existing methods are tied to U-Net T2I pipelines, while real-world abuse (nudify applications, sextortion) predominantly occurs in image-conditioned and video settings.
  2. The benchmark is an independent, significant contribution. The construction protocol of UniErase-Bench's image-conditioned subsets is complete, covering two models, three scenarios, and 3,600 image-conditioned instances in total.
  3. Cross-architecture / cross-task unification is the paper's main contribution, with a clear mechanistic rationale. Choosing FFNs as the intervention point is the paper's most elegant observation and the technical foundation on which "unification" rests. Compared with ConceptPrune's direct zeroing of channels, the present design is supported by both ablations and baseline data.
  4. Comprehensive and fairly compared experiments. Baselines span external filtering (3), fine-tuning-based (5), inference-time intervention (2), and weight editing (2); evaluation covers handcrafted prompts, four black-box attacks, and one white-box attack. Benign-cost metrics are reported, avoiding the common "erasure-rate-only" pitfall of safety papers.
  5. Strong practicality. One-off offline editing, zero inference overhead, training-free. Code and benchmark are released, supporting reproducibility.
Critical weaknesses
  1. Novelty limitation: no single technical component is first-proposed; the method is a combination of two existing lines of work. The localization stage (FFN channel attribution + selection) shares its lineage with ConceptPrune's skilled-neuron pruning (ICLR 2025); the editing stage (closed-form orthogonal-complement projection to remove the concept subspace) matches the core idea of AdaVD (CVPR 2025), differing only in the intervention site (FFN value vectors vs. cross-attention value representations). The "FFN as key-value memory" premise is likewise the established conclusion of Geva et al. (2021). The novelty claim should therefore be understood at the system/composition level — placing the intervention in FFNs to gain cross-architecture/cross-task capability — rather than at the algorithmic level.
  2. No quantitative baseline comparison on the new architectures (SD3/FLUX/Kontext/video). The authors provide a full comparison against ConceptPrune, UCE, and others on SD-v1.4 (Table 1), but the cross-architecture results in Figure 3 only report UniErase's own before/after ASR; ConceptPrune appears only qualitatively in Figure 1 ("zeroing too many channels causes video generation failure") and never in a formal table.
  3. Residual ASR in the most realistic scenario is high, inconsistent with the strength of the paper's claims. After editing, ASR in S3 of T+I→I remains at 35.0%, and T→V retains 12.1%. The abstract's statement of "consistently suppresses explicit content generation across all evaluated task categories" is too strong given a 35% residual attack success rate. The authors should more honestly differentiate the degree of achievement across scenarios and analyze why S3 is markedly harder.
  4. Single-point dependence on NudeNet with a methodological circularity risk. NudeNet plays three roles simultaneously: filtering the contrastive condition pairs in the localization stage, validating benchmark construction, and scoring the final evaluation. The baselines' supervision signals have never touched NudeNet, whereas UniErase's localization stage uses NudeNet-verified pairs and the final scoring is again NudeNet — a "evaluator-alignment" circularity risk. NudeNet's false-positive/false-negative characteristics propagate simultaneously into both the benchmark and the method evaluation. The paper reports no human-annotated validation or cross-check with a second detector — a clear shortcoming for a work whose selling point is a benchmark.
  5. No adaptive-attack analysis targeting the proposed method. The white-box evaluation uses P4D, a prompt-level attack, whereas UniErase directly modifies weights and the edit is closed-form and public: an attacker aware of the defense could attempt (a) rediscovering equivalent concept directions in the remaining subspace, (b) recovering the concept via light fine-tuning of the edited model (concept recovery), or (c) constructing evasion directions from the orthogonal complement of U. Without an "attacker knows the defense" evaluation, the robustness conclusions of erasure methods are incomplete. The paper does not discuss this at all.
  6. No fine-grained over-erasure analysis of adjacent benign concepts. Only global FID/CLIP on COCO are reported, which cannot answer whether adjacent concepts — swimwear, underwear, figure art, medical anatomy, classical paintings — are collaterally damaged after erasure. This is the most practical concern for deployers: a model that cannot generate "a person in a swimsuit" is equally unusable. A per-concept generation-quality probe (e.g., CLIP alignment / detection rates for benign human-related concepts) is recommended.
Questions for authors

1.How exactly are the safe counterpart conditions c⁻ constructed? How sensitive is the localization result to different construction choices? 2.Please provide quantitative comparisons of UCE/ConceptPrune on SD3, FLUX.1-dev, and the video models (the qualitative display in Figure 1 suggests the authors already have this data). If transfer is genuinely infeasible, please explain the failure mode rather than showing only a qualitative figure. 3.Have fine-tuning-based recovery experiments been performed on the edited model (e.g., LoRA fine-tuning with a small set of explicit samples to observe concept-recovery difficulty)? This directly determines the method's safety boundary in realistic adversarial settings. 4.Is there any human spot-check annotation for UniErase-Bench? What is the agreement rate between NudeNet and a second detector (or human annotators)? 5.For the failure cases in S3 with residual ASR of 35%, do failures predominantly originate from the text side or the visual side? Can increasing K/λ suppress S3, and at what cost? 6.What is the impact of erasure on adjacent benign concepts (swimwear, underwear, nude sculptures/paintings, medical imagery)? Is there any concept-level evaluation?

7 comments
AuthorsOpenReview ↗

Response to Reviewer WmXb 1/4

W1 Novelty and Technical Contributions

We thank the reviewer for the constructive feedback. Our method indeed draws inspiration from several established lines of research, including FFNs as key-value memories, channel-level concept localization, and orthogonal projection. However, existing approaches are often constrained by specific architectures or individual generation tasks, limiting their generalizability to diverse real-world scenarios.

The central challenge of our work is to develop a highly generalizable concept-erasure mechanism that can seamlessly adapt to heterogeneous diffusion architectures and diverse generation tasks. Through extensive experiments and mechanistic analysis, we find that directly applying existing methods or individual components is insufficient to achieve this goal. Therefore, our novelty lies not in simply combining existing techniques, but in carefully adapting and refining them to the representation characteristics of diffusion models, enabling their effective integration into a unified framework. The key technical distinctions from prior work are outlined below.

(1) Unified intervention space across architectures. Unlike AdaVD's inference-time intervention on cross-attention representations, UniErase edits generation-side FFN weights, which operate on representations after conditioning information has been integrated. This provides a shared intervention space across heterogeneous architectures and generation tasks, without relying on architecture-specific attention mechanisms.

(2) Adapting orthogonal projection to FFN weight editing. AdaVD performs inference-time orthogonal projection on cross-attention values, where target concept directions can be directly derived from text embeddings. However, FFNs lack such explicit semantic alignment, making it difficult to directly identify precise concept directions. To address this challenge, we introduce contrastive subspace estimation to identify concept-specific directions, enabling precise and selective FFN weight editing.

Component-level ablation. To further demonstrate the significance of our key design choices, we conduct a component-level ablation on both SD-v1.4 and FLUX.1-dev using the adversarial RAB dataset. The results demonstrate that selective projection and contrastive refinement improve benign content preservation while maintaining strong erasure effectiveness.

Table 1: Component ablation on SD-v1.4 and FLUX.1-dev.

SD-v1.4FLUX.1-dev
VariantProjectionContrastive RefinementER ↑FID ↓ER ↑FID ↓
(a) Channel zeroing✗✗98.7426.1895.5436.55
(b) Target-only PCA✓✗97.6625.5696.5236.01
(c) UniErase (Full)✓✓98.8324.9596.4134.85
AuthorsOpenReview ↗

Response to Reviewer WmXb 2/4

W2 & Q2 Cross-Architecture Baseline Comparisons

Thank you for this valuable suggestion. We agree that extending quantitative baseline comparisons to newer architectures would provide stronger evidence for the generalizability and practical advantages of UniErase.

(1) Comparisons on FLUX.1-dev. We extend our evaluation to include UCE[1], ConceptPrune[2], EraseAnything[3], DVE[4], and GEM[5], covering representative editing-based, training-based, and inference-time erasure methods. We compare erasure effectiveness on the adversarial RAB dataset alongside benign generation quality, providing a more comprehensive assessment of the erasure–utility trade-off on DiT architectures.

Table 2: Quantitative comparison of concept erasure methods on FLUX.1-dev.

MethodER ↑CLIP ↑FID ↓
UCE52.9330.9235.92
ConceptPrune92.9429.9950.23
EraseAnything30.7130.7934.18
DVE91.2929.9533.49
GEM80.4230.0241.35
UniErase (Ours)96.4130.9434.85

We have included the additional experimental results and detailed comparisons in Appendix F of the revised manuscript.

(2) Comparisons on image-conditioned and video generation. We additionally provide quantitative comparisons with ConceptPrune on FLUX.1-Kontext for image-conditioned image generation and HunyuanVideo-1.5 for text-to-video generation. Additionally, we evaluate UCE under the text-led S1 scenario for image-conditioned generation, as its original formulation targets text-conditioned attention mappings rather than visual conditioning.

Table 3: Quantitative baseline comparisons on image-conditioned and video generation.

(a) FLUX.1-Kontext (T+I→I)

MethodS1 ASR ↓S2 ASR ↓S3 ASR ↓Overall ASR ↓CLIP-I ↑
UCE49.7————
ConceptPrune42.559.689.763.993.02
UniErase (Ours)10.725.035.023.699.21

(b) HunyuanVideo-1.5 (T→V)

MethodASR ↓TC ↑Motion ↑
Original50.60.96707.16
ConceptPrune0.00.207793.06
UniErase (Ours)12.10.95508.19

UniErase consistently outperforms ConceptPrune on FLUX.1-Kontext, achieving stronger erasure while better preserving reference-image similarity. For video generation, ConceptPrune's lower ASR primarily results from severe generation failures, including distorted visual content and temporal instability, rather than effective concept erasure. In contrast, UniErase suppresses explicit content while preserving coherent video generation and natural motion.

[1] Unified Concept Editing in Diffusion Models, WACV 2024

[2] ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning, ICLR 2025

[3] EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers, ICML 2025

[4] Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models, arXiv 2026

[5] GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows, ICML 2026

W3 & Q5 S3 Residual ASR

We thank the reviewer for highlighting this concern. We will revise the abstract to more accurately state that UniErase substantially suppresses explicit content generation across the evaluated task categories, and provide further analysis of challenging scenarios.

(1) Erasure effectiveness in S3. Although the residual ASR of 35.0% leaves room for further improvement, UniErase substantially reduces the original ASR from 96.7% to 35.0%, achieving a 61.7 percentage-point reduction and significantly outperforming ConceptPrune (89.7%). This demonstrates the effectiveness of UniErase even under challenging dual-conditioning scenarios.

(2) Analysis of residual failures. Unlike S1 and S2, S3 simultaneously introduces explicit concepts through textual instructions and reference images, creating stronger cross-modal reinforcement. Such dual conditioning may lead to more distributed and entangled concept representations within the model, making selective concept erasure particularly challenging.

(3) Sensitivity to K and λ. To investigate whether stronger editing can further reduce residual ASR, we evaluate different channel budgets K and contrastive coefficients λ on FLUX.1-Kontext (Table 4). Increasing K from 50 to 100 reduces ASR from 35.0% to 22.6% with nearly unchanged CLIP-I, suggesting that expanding the coverage of concept-relevant channels is an effective strategy while preserving benign information through selective projection. Increasing λ from 5 to 10 further reduces ASR to 12.3%, but leads to a greater decrease in CLIP-I, indicating that stronger subspace modification may affect internal visual representations.

Table 4: Sensitivity analysis of K and λ on FLUX.1-Kontext (S3).

KλASR (%) ↓CLIP-I ↑
50535.099.21
75527.799.04
100522.699.22
50725.098.53
501012.395.37
AuthorsOpenReview ↗

Response to Reviewer WmXb 3/4

W4 & Q4 NudeNet Dependence

We thank the reviewer for raising this important concern. We clarify this issue from two perspectives: rigorous benchmark validation and independent cross-evaluator verification.

(1) Two-stage benchmark validation. Since reference images synthesized by FLUX.1-dev may not always satisfy the intended scenario requirements, we employ a rigorous two-stage quality control process. First, NudeNet automatically filters generated images, with non-compliant samples regenerated until the requirements are met. More importantly, all cases included in the final benchmark undergo double-blind human review to ensure that both textual instructions and reference images conform to their respective scenario definitions. Thus, the benchmark's validity is not solely dependent on NudeNet's automated predictions.

(2) Independent evaluation with Qwen3-VL. To address potential evaluator-alignment bias, we additionally employ Qwen3-VL as an independent evaluator, performing binary classification of the same generated images into explicit and benign categories. As shown in Table 5, Qwen3-VL and NudeNet yield consistent performance rankings across all evaluated methods, with UniErase achieving the highest erasure rate under both evaluators. These results further support the reliability of our evaluation and demonstrate that UniErase's comparative advantage is not specific to NudeNet.

Table 5: Cross-evaluator comparison on FLUX.1-dev using RAB.

MethodNudeNet ER ↑Qwen3-VL ER (Sample-level)↑
UCE52.9374.44
ConceptPrune92.9495.44
EraseAnything30.7158.15
DVE91.2994.24
GEM80.4292.98
UniErase (Ours)96.4198.16

W5 & Q3 Adaptive-attack

We thank the reviewer for highlighting the importance of fine-tuning-based concept recovery. Recent works[1,2] have investigated resistance to concept relearning, aiming to increase the difficulty and cost of recovering erased concepts through fine-tuning. While our primary contribution focuses on unified concept erasure across architectures and generation tasks, we further conduct comparative experiments to evaluate UniErase's robustness against LoRA-based concept recovery.

LoRA-based concept recovery. To assess this robustness dimension, we conduct additional experiments on FLUX.1-dev, comparing UniErase with representative baselines under identical LoRA fine-tuning settings. We evaluate the ASR on RAB at different fine-tuning stages to measure how readily erased concepts can be recovered.

As shown in Table 6, UniErase consistently achieves the lowest ASR across all LoRA fine-tuning stages, demonstrating strong resistance to concept recovery. Interestingly, both ConceptPrune and UniErase exhibit a slight initial decrease in ASR before gradually increasing, which may reflect early perturbations to generation behavior followed by progressive recovery of concept-related representations. These findings further support the robustness of UniErase against fine-tuning-based adaptive attacks.

Table 6: Robustness against LoRA-based concept recovery on FLUX.1-dev.

MethodBefore Recovery ↓50 Steps ↓100 Steps ↓150 Steps ↓200 Steps ↓
Original85.39————
ConceptPrune10.159.379.5610.4412.83
GEM25.7825.9927.5042.8646.40
UniErase (Ours)3.632.152.114.096.52

[1] The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models, CVPR 2025

[2] Meta-Unlearning on Diffusion Models: Preventing Relearning Unlearned Concepts, NeurlPS 2025

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Major Contributions, Strengths, and Responses to Concerns

Dear AC, SAC, and PCs,

We sincerely appreciate your time and efforts in managing the review process. Our submission initially received review scores of 3, 3, and 2. As we have not received further responses from the reviewers following the AC's reminder, we summarize the reviewers' concerns and our rebuttal below. We hope this summary will facilitate the final assessment and help reduce your review workload.

1. Core Contributions

Our work addresses a critical gap between the expanding threat surface of diffusion models and existing T2I-centric concept erasure research.

  • UniErase-Bench, a unified benchmark covering four generation tasks, with three cross-modal attack scenarios for image-conditioned generation.
  • UniErase, a training-free framework that selectively edits generation-side FFN weights, enabling concept erasure across heterogeneous U-Net and DiT architectures and diverse generation tasks without additional inference overhead.

2. Acknowledged Strengths

  • Timely motivation and benchmark significance. WmXb explicitly described UniErase-Bench as an "independent, significant contribution," while 9R82 highlighted its comprehensive coverage of four generation tasks.
  • Cross-architecture and cross-task unification. In particular, WmXb identified cross-architecture and cross-task unification as our main contribution, while JMY6 highlighted its applicability across U-Net and Transformer architectures and all four tasks.
  • Comprehensive experiments and strong empirical performance. WmXb highlighted the comprehensive baselines and adversarial evaluations, while 9R82 and JMY6 acknowledged UniErase's effective erasure and benign content preservation.
  • Efficiency and practicality. WmXb explicitly highlighted the training-free, one-time offline editing mechanism with zero inference overhead and the release of code and benchmark resources. Similarly, 9R82 recognized the efficiency of our closed-form FFN weight editing approach.

3. Summary of Rebuttal

In response to the reviewers' feedback, we have provided substantial additional experiments and clarifications.

ReviewerRequestOur Response
WmXbAlgorithmic noveltyClarified two key technical innovations, further validated by component-level ablations.
Cross-architecture comparisonsValidated UniErase's effectiveness and generalizability across modern diffusion architectures and generation tasks.
Residual ASR in S3Analyzed S3 failures and demonstrated improved erasure through stronger editing.
NudeNet dependenceClarified benchmark validation through human review and independently confirmed consistent method rankings using Qwen3-VL.
Adaptive attack robustnessDemonstrated the lowest ASR under 50–200 steps of LoRA-based concept recovery.
Adjacent benign conceptsClarified benign preservation mechanisms, erasure criteria, and borderline cases.
Safe counterpart constructionDetailed condition-pair construction and verified high localization stability across disjoint subsets.
9R82Comparisons on recent architecturesvalidate against five baselines on FLUX.1-dev.
Offline computational overheadClarified stage-wise computational requirements and sample budgets.
Component-level ablationValidated the benefits of selective projection and contrastive refinement across architectures.
Missing baseline referencesAdded missing citations and clarified baseline publication information.
JMY6Benchmark novelty and coverageClarified the benchmark's four-task, three-scenario coverage and distinctions from prior work.
Algorithmic distinctionsDetailed the technical distinctions and improvements over prior methods.
Roles of concept-relevant FFN channelsDemonstrated the benefits of selective editing over direct channel zeroing for benign preservation.
Failure case analysisAnalyzed dual-conditioning failures and further reduced S3 ASR through stronger editing.
Adjacent benign concepts and over-erasureClarified nudity-related erasure criteria, benign preservation mechanisms, and the treatment of borderline cases.

We have incorporated the most important additional experiments and analyses into the revised manuscript and will further include the remaining details in the final version.

We also note that Reviewer JMY6's citations lacked a corresponding reference list, so we addressed the likely intended works based on their descriptions.

4. Concluding Remarks

We believe our work identifies and addresses a critical gap between the expanding threat surface of diffusion models and the limited coverage of existing concept erasure methods through a unified benchmark and editing framework. The additional experiments and clarifications further substantiate our central claims and strengthen our contribution.

Thank you again for your time and consideration. We hope the additional evidence will support a positive final assessment.

Sincerely,

The authors of UniErase

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.