acceptodds
Under review as a conference paper at ICLR 2027

Erasing Unsafe Image and Text Concepts in Image-to-Video Generation

Abstract

Video generation models have been rapidly improving, but they also amplify the risk of propagating undesired concepts (e.g., unsafe or copyrighted content), motivating reliable concept erasing. Prior video concept erasing largely assumes text-only conditioning, overlooking Image-to-Video settings where the input image provides strong evidence that persists across frames. As a result, existing text-focused erasing methods become ineffective in I2V generation, where the input image may preserve the undesired concept across frames. To address this gap, we propose Image-to-Video Concept Erasing (I2V-CE), a new task that removes an unsafe concept in I2V generation while preserving non-target content and temporal coherence. To achieve this, we introduce UNsafe Image and Text Erasure (UNITE), a training-free method that jointly handles unsafe evidence from image and text conditions: First-frame Masking and Blending (FMB) regenerates concept-related regions in the first-frame latent to suppress unsafe visual evidence, and Token-wise Safe Projection (TSP) projects concept-related token embeddings onto a safe subspace. Extensive experiments show that UNITE enables reliable concept removal in I2V generation across diverse scenarios.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Concept Erasing, Image-Text-to-Video Generation, Machine Unlearning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_yTuZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the earsure of unsafe concepts in image-to-video generation, focusing on the issue that both reference images and text prompts may introduce unsafe content. The paper propose the training-free UNITE framework, which consists of two complementary modules: FMB, which relaxes the first-frame constraints on target regions in reference images, and TSP, which identifies text tokens related to the target concept and projects them into a predefined safe subspace. The paper conducts experiments across three scenarios—where unsafe content originates from text, images, or both—and evaluates the effectiveness of erasing nudity and specific character concepts.

Critical strengths
  1. Clear practical significance. This paper addresses the issue of continuously injecting unsafe information into the video generation process from both reference images and text conditions. The clear scene segmentation in S1–S3 helps analyze the impact of different condition sources on the erasure effect.

  2. FMB and TSP operate on image and text data, respectively, with a clear division of labor. Ablation experiments support the complementary nature of the two modules; in particular, when the reference image contains the target concept, joint processing significantly outperforms using only the text module.

  3. The main experiments demonstrated strong erasure performance. UNITE outperformed the compared methods on the nudity detection metrics across all three scenarios, with particularly significant improvements observed in S2 and S3.

Critical weaknesses
  1. Lack of robustness evaluation against multimodal adversarial attacks. UNITE was designed to defend against multimodal injection of unsafe information during video generation. Whether it is sufficiently robust against multimodal jailbreaking attacks. Could you evaluate UNITE on the benchmarks like RunawayEvil and MMA.

  2. The cross-model generalizability. Most experiments have focused on Wan2.2-TI2V-5B. Although validated TSP on the CogVideoX and Wan T2V settings, these results only support the transferability of the text module and are insufficient to demonstrate that the full UNITE framework is equally effective on other I2V models. Differences in image condition injection methods, text representations, and generation mechanisms across models may affect the performance of the two modules.

Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

W1. Evaluation against multimodal adversarial attacks.

A1. We evaluate Base and UNITE on identical attack inputs.

(i) MMA. We transfer all ten final textual attacks from the public example logs to paired safe and unsafe references, without outcome-based filtering.

Table R1. NudeNet detection ratio under transferred MMA textual attacks.

Reference conditionBaseUNITE
Safe reference4.94% (40/810)0% (0/810)
Unsafe reference97.16% (787/810)0% (0/810)

UNITE reduces NudeNet detections under both reference conditions.

(ii) RunawayEvil-based attack transfer. Starting from 20 clothed reference images, we use a RunawayEvil-based pipeline to generate edited images and video prompts. For each policy, all 20 retained image–prompt pairs, including unsuccessful attack attempts, are applied to Base and UNITE under identical generation settings.

Table R2. VLM-flagged video ratio (↓\downarrow).

Attack policyBaseUNITE
Partial Stage 125% (5/20)5% (1/20)
Stage 2, 256 steps20% (4/20)0% (0/20)

UNITE reduces the flagged-video ratio by 20 percentage points under both attack policies.

W2. The cross-model generalizability.

A2. We further evaluate the full UNITE framework, including both FMB and TSP, on CogVideoX1.5-5B-I2V, a different I2V architecture. We adapt FMB to its concatenated image-conditioning channels, with first-frame latent blending at the scheduler's noise level. TSP operates in the model's native text-embedding space during the first 13 denoising steps, together with concept-negative conditioning.

Table R3. Full-framework CogVideoX development evaluation.

Scenario / methodNudeNet detection ratio ↓Imaging Quality ↑Subject Consistency ↑
S2 / Base48.64%0.58600.8846
S2 / UNITE0%0.60220.8468
S3 / Base74.20%0.70870.9830
S3 / UNITE0%0.63610.8846

As shown in Table R3, UNITE reduces NudeNet detections to 0% in both scenarios, demonstrating that the combined image- and text-side interventions can be extended to another I2V architecture. However, some regenerated-region artifacts remain. These results support cross-model applicability for concept erasure, while further improving generation quality across architectures remains future work.

Thank you for the detailed response. The clarifications have resolved my questions and I will maintain my current score.

Thank you for your feedback. We are glad that our responses have successfully addressed your questions. If you need any further clarification or have additional questions, please feel free to let us know. We would be happy to provide further details.

Reviewer_yduDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Image-to-Video Concept Erasing (I2V-CE), a new task that remove undesirable concepts from image-to-video generation when unsafe information originates from the text prompt, the reference image, or both. To address this issue, they propose the UNITE, a training-free framework consisting of First-frame Masking and Blending (FMB) and Token-wise Safe Projection (TSP). Specifically, FMB selectively releases detected unsafe regions from first-frame conditioning and regenerates them during sampling. TSP identifies concept-related text tokens and projects their embeddings onto a predefined safe subspace. They also conduct a theoretical analysis of potential concept leakage caused by global reprojection. Experiments on Wan2.2-TI2V-5B show substantial improvements over existing training-free concept-erasing baselines, with additional results on object erasing, detector variants, and I2V generation.

Critical strengths
  1. This paper identifies a critical gap in current video concept-erasing methods, i.e., approaches designed for text conditioning fail to remove unsafe concepts strongly anchored in I2V reference images.
  2. The proposed UNITE requires no model retraining and can be easily merged into existing inference pipelines. Its effectiveness is demonstrated by comprehensive component ablations and comparisons with FMB-augmented baselines.
  3. This paper is well-motivated, and the presentation is clear.
Critical weaknesses
  1. The proposed FMB relies on a pre-trained detector to localize unsafe concepts in the reference image. This introduces an external dependency and creates a single point of failure, i.e., detection errors or out-of-distribution concepts can compromise the entire erasure process. However, the authors do not explore the robustness of UNITE when the detector fails.
  2. The visual baseline comparisons may overstate the novelty. The main visual baseline seems to be the plain text erasure method, which is structurally disadvantageous in S2 and S3 scenes. Although the comparison with the enhanced version of FMB is helpful, it is not clear whether the visual gain comes from the proposed hybrid strategy itself or merely from applying any masking mechanism to the first frame.
  3. Regenerating masked regions in the first frame may introduce new unsafe content, and token projection may cause over-erasure of benign concepts that share linguistic features with unsafe ones. This paper lacks a detailed analysis of these effects.
  4. The evaluations are limited to the I2V setup. It is unclear whether the conclusion is applicable to other architectures or more diverse video datasets.
  5. The proposed method largely combines existing techniques, e.g., mask-based latent regeneration and token-level embedding projection. Besides, ConceptGuard [1] has already explored safety in image-and-text-conditioned video generation through visual editing and text intervention. The authors should better clarify the novelty of UNITE and its differences from related methods. [1] ConceptGuard: Proactive Safety in Text-and-Image-to-Video Generation through Multimodal Risk Detection, Arxiv 2025.
Questions for authors

Please refer to the weakness.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We appreciate Reviewer yduD for your constructive feedback. We hope our response effectively addresses your concerns and questions.

W1. Dependency to the external detector.

A1. We acknowledge that detection failure limits FMB's effectiveness, as missed regions cannot be regenerated. However, FMB is not tied to any specific detector or architecture. Its detector-agnostic design allows us to adopt the strongest available detector for each target concept and readily incorporate future improvements in localization. As demonstrated in Table 6, UNITE achieves effective erasure with multiple off-the-shelf detectors (NudeNet, EraX, and EraX-s) without modifying the erasure pipeline. Furthermore, our mask-coverage experiments confirm that more complete localization leads to more reliable erasure (Table R1). Therefore, while detection dependency remains a limitation, FMB's plug-and-play design enables UNITE to continuously benefit from advances in detection models without additional training.

Table R1. Erasing performance depends on the mask coverage.

Mask interventionMask coverageNudeNet detection ratio ↓
Dilated mask100.00%0% (0/1,620)
Right-shifted mask71.06%11.91% (193/1,620)
Partial mask49.92%89.14% (1,444/1,620)

W2. What is the main novelty compared to the baselines?

A2. UNITE's contribution lies in both FMB's in-model regeneration and TSP's improved text-side intervention, rather than simply applying first-frame masking.

First, Figure A3 directly compares FMB with conventional binary masking. Although hard masking removes unsafe regions, it introduces visible boundary artifacts, whereas FMB regenerates these regions with smooth transitions to the preserved content Second, as the reviewer noted, the original baselines primarily target text conditioning. We therefore combine them with the same FMB module to isolate the differences in text-side erasure. As shown in Table 1, FMB+SAFREE and FMB+VideoEraser still yield Nude/GPT scores of 70.25/84.25 and 79.07/100.00, respectively, whereas FMB+TSP achieves 0.00/24.50.

These results demonstrate that first-frame masking alone does not explain UNITE's performance. FMB improves visual regeneration, while the TSP-based text-side intervention provides substantially stronger concept suppression than the evaluated alternatives.

W3. Unsafe regeneration and unintended changes to benign concepts.

A3. We acknowledge that regenerating masked regions may introduce unsafe content if unsafe signals remain in the conditioning. This motivates combining FMB with TSP-based text-side intervention, which suppresses unsafe conditioning signals during generation. As shown in Table 5, adding TSP-based intervention to FMB reduces the Nude Score from 91.69% to 1.75% in S3.

Regarding unintended erasure, we additionally evaluate cross-concept selectivity using cat, dog, and car. As shown in Table R2, UNITE preserves 100% recognition of non-target categories across all tested erasing conditions. Furthermore, VBench-I2V results (Table 4) show that UNITE maintains general generation quality close to the original model (89.12 vs. 89.28). These results support effective unsafe-concept suppression while preserving unrelated concepts and general generation capability in the evaluated settings.

Table R2. Selective concept removal across three object categories. Each slots mean detection ratio by Qwen3-VL-8B.

Erasing conditionCat input (cat)Dog input (dog)Car input (car)
Base100.00%100.00%100.00%
Erase cat0.00%100.00%100.00%
Erase dog100.00%4.44%100.00%
Erase car100.00%100.00%20.00%

W4. Limited evaluation to the I2V setup.

A4. We already evaluate the generalizability of TSP beyond I2V generation in Section 4.4 and Table 7, using two different generation backbones: CogVideoX-5B and Wan2.2-T2V-5B. TSP achieves the lowest Nude Scores on both models (3.78% and 0.00%), as well as the lowest GPT Score on Wan2.2-T2V-5B (16.80%). On CogVideoX-5B, its GPT Score is comparable to SAFREE (71.70 vs. 68.60). These results demonstrate that TSP's effectiveness extends across different video-generation architectures and T2V settings.

AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

W5. Novelty and Differences from ConceptGuard.

A5. We thank the reviewer for highlighting this relevant prior work and will incorporate it into the final version. ConceptGuard primarily focuses on multimodal risk identification, followed by external image editing (FLUX.1 Kontext) and text embedding intervention. In contrast, FMB releases detected unsafe regions from first-frame conditioning and regenerates them through the video generator itself, using soft blending to preserve surrounding content. For text-side erasure, TSP selects tokens based on their unsafe-subspace components and directly projects them onto a safe subspace, unlike ConceptGuard's orthogonal-complement-based intervention.

We further validate these differences experimentally in Table R3, R4, and R5. Table R3 shows that TSP achieves higher unsafe-token identification accuracy than the existing methods. Table R4 demonstrates that UNITE achieves stronger target-concept erasure than FLUX.1 Kontext across both concepts and scenarios. Finally, Table R5 highlights the computational advantage of FMB, which requires no separate generative image-editing pass, unlike FLUX.1 Kontext.

Table R3. Unsafe token identification.

MethodAUC ↑F1 ↑
SAFREE0.4790.251
VideoEraser0.4540.202
ConceptGuard†0.3630.141
TSP0.7360.413

†In the official repo, they only released the weights without executable inference/forward code, so we reproduce ConceptGuard based on the paper.

Table R4. Comparisons FLUX.1 Kontext and FMB.

MethodMickey Mouse: S2Mickey Mouse: S3Hello Kitty: S2Hello Kitty: S3
CLIP ↓LPIPS ↑CLIP ↓LPIPS ↑CLIP ↓LPIPS ↑CLIP ↓LPIPS ↑
FLUX + TSP0.98790.15650.98590.19800.92530.16130.93480.2390
UNITE0.64460.64840.64520.63020.57500.56940.61870.5177

Table R5. Computational cost.

MethodExternal image-editing time ↓
FLUX + TSP10.68 s
UNITE0 s — no separate editor

The author solved most of my problems. I hope the author can add the additional experimental results to the main paper. After doing this, I can consider improving my score.

Reviewer_TL92OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies Image-to-Video Concept Erasing (I2V-CE) across three scenarios where unsafe concepts originate from text prompts, reference images, or both. It highlights the limitations of existing text-focused erasure methods under strong visual conditioning in image-to-video generation.The authors propose UNITE, a training-free framework combining First-frame Masking and Blending (FMB) and Token-wise Safe Projection (TSP), which intervene on image and textual conditioning, respectively. The paper also provides a theoretical analysis of potential unsafe-component leakage through global reprojection and derives a bound on the residual unsafe components after TSP projection.Experiments on Wan2.2-TI2V-5B demonstrate substantial improvements over baselines such as SAFREE and VideoEraser, particularly under unsafe image conditioning. Overall, the work addresses a relevant problem with promising empirical results, although the selectivity of the proposed erasure mechanism and its applicability boundaries require further validation.

Critical strengths

a)Well-motivated problem: Existing text-focused concept erasure methods struggle with unsafe reference-image conditioning in I2V generation. The S1–S3 taxonomy provides a clear framework for different sources of unsafe input conditions, and UNITE achieves substantial improvements, particularly in S2 and S3. b)Practical training-free design: FMB and TSP intervene on image and text conditioning, respectively, with complementary effects supported by the ablation results. The training-free framework is practical for deployment with large video generation models. Additional evaluations across detectors, copyrighted characters, visual domains, and T2V settings provide encouraging evidence of its applicability, while aggregate video-quality scores remain generally comparable to the original model.

Critical weaknesses

a) Insufficient evidence for selective concept removal and non-target preservation.Full-image CLIP similarity and LPIPS do not directly verify target identity erasure, while VBench does not establish preservation of non-target content within regenerated regions. FMB's potential off-target effects also remain insufficiently quantified. The safety evaluation raises additional concerns, as NudeNet is used for both FMB localization and evaluation, albeit with different thresholds. Although Appendix A.6 discusses discrepancies between NudeNet and GPT-based scores, the residual GPT unsafe score in S2 remains difficult to interpret without independent calibration or human evaluation. These limitations weaken the evidence supporting selective erasure and preservation. b)Insufficient empirical validation of the TSP mechanism and its generalization boundaries.The theoretical analysis is not accompanied by measurements of safe–unsafe subspace overlap, residual unsafe components after TSP projection, or actual reprojection leakage on evaluation prompts. TSP's substantial improvement in S2, despite nominally safe text prompts, also remains insufficiently explained without comparisons against simpler conditioning perturbations. Furthermore, token-selection robustness under semantic distribution shifts and imperfect localization is not systematically evaluated. The S4 scenario (safe image and text producing unsafe output) also represents an applicability boundary that deserves clarification, although addressing it is beyond the paper's stated task. Together, these issues limit confidence in the proposed mechanistic explanation and the generality of selective suppression.

Questions for authors

Q1. Table 5 shows that TSP substantially improves safety in S2, reducing the Nude score from 18.97 to 0.04 despite nominally safe text prompts. Please identify the selected tokens and their embedding changes after projection. Could the authors compare TSP against simpler interventions, including global embedding attenuation, norm-matched random projection, and token-wise attenuation without the safe subspace, under matched intervention strength? These comparisons would help establish whether the gains arise from concept-specific projection rather than general conditioning perturbations. Q2. Full-image CLIP similarity and LPIPS cannot directly confirm target identity erasure. Could the authors provide target-specific identity recognition, non-target region preservation measurements, and temporal consistency evaluation for regenerated content? Although Appendix A.6 discusses the discrepancy between NudeNet and GPT-based scores, independent human evaluation or evaluator calibration would further strengthen the safety assessment. Q3. Could the authors report the actual safe–unsafe subspace overlap, residual unsafe components after TSP projection, and reprojection leakage of SAFREE on real evaluation prompts? Does prompt-dependent leakage correlate with unsafe generation outcomes? Such measurements would help establish whether the theoretical analysis explains the observed performance improvements. Q4. Please clarify UNITE's applicability boundary for an additional scenario, S4 (safe image + safe text → unsafe output), where unsafe content may emerge from the model's generative distribution rather than the conditioning inputs. While solving S4 is not required within the paper's stated scope, a preliminary comparison of unsafe generation rates under benign inputs would help clarify this boundary. Q5. How robust are UNITE's inference-time interventions under distribution shifts, including unseen aliases, semantic paraphrases, multiple target objects, and incomplete localization masks? Could the authors distinguish missed interventions from incorrect modifications to unrelated tokens or regions? This would help clarify the framework's generalization and selectivity limits.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/3)

We appreciate Reviewer TL92 for your constructive feedback. We hope our response effectively addresses your concerns and questions.

W(a)-1. Selective concept removal.

A1. In the main paper, we evaluate concept erasure and general generation quality using VBench-I2V. To further verify selective concept removal, we conduct an additional cross-concept evaluation using three distinct concepts (cat, dog, and car), with five input images per concept. We evaluate whether each input concept remains recognizable using Qwen3-VL after applying different concept-erasing conditions.

Table R1. Selective concept removal across three object categories.

Erasing conditionCat input (cat)Dog input (dog)Car input (car)
Base100.00%100.00%100.00%
Erase cat0.00%100.00%100.00%
Erase dog100.00%4.44%100.00%
Erase car100.00%100.00%20.00%

As shown in Table R1, UNITE effectively suppresses each target concept while preserving all non-target concepts across the evaluated samples. These results provide direct evidence that UNITE can selectively erase targeted concepts without affecting unrelated object categories.

W(a)-2 and Q2-1. Target-specific identity recognition beyond on the CLIP and LPIPS.

A2. As requested, we introduce a VLM-based evaluation to directly assess target-character recognition beyond CLIP and LPIPS. We use Qwen3-VL-8B to determine whether Mickey Mouse or Hello Kitty remains recognizable across nine fixed frames per generated video. To validate this evaluation, we first test all 40 original reference images, all of which are correctly identified.

Table R2. Target-character detection ratio ↓.

MethodMickey S2Mickey S3Hello Kitty S2Hello Kitty S3
SAFREE100.00%100.00%100.00%100.00%
VideoEraser100.00%100.00%100.00%100.00%
UNITE (Ours)0.56%0.56%2.78%1.67%

As shown in Table R2, UNITE reduces target-character recognition to 0.56% for Mickey Mouse and 1.67–2.78% for Hello Kitty, substantially outperforming the baselines. These results are consistent with the CLIP and LPIPS findings and further demonstrate UNITE's effectiveness in removing target-character identities.

W(a)-3 and Q2-2. Non-target region preservation measurements and temporal consistency evaluation for regenerated content.

A3. To evaluate this, we intentionally apply false-positive masks to benign regions in ten safe reference images, covering 13.39% of the image area. We then evaluate the corresponding cropped regions across all 81 generated frames.

Table R3. Non-target region preservation and temporal consistency for regenerated content.

MethodMotion Smoothness ↑Subject Consistency ↑Background Consistency ↑
Base99.3998.4496.76
UNITE99.4198.4096.87
Forced-mask UNITE99.4097.7995.88

Even with forced masking, Motion Smoothness remains at 99.40, while Subject and Background Consistency remain within one point of the base model. These results support temporal consistency within the regenerated regions. Figure A3 further illustrates the preservation of surrounding non-target content and the visual consistency of regenerated regions.

W(a)-4 and Q2-3. Human evaluation about discrepancy between NudeNet and GPT score.

A4. To investigate the discrepancy between NudeNet and GPT scores, we conducted a human evaluation with seven participants on 38 UNITE-generated videos showing GPT-positive/NudeNet-negative disagreement (with no cases of the opposite disagreement). Each participant independently assessed the presence of nudity and explicit sexual activity using three labels: present, absent, or uncertain.

Table R4. Human evaluation of unsafe-content assessment.

Human criterionPresentAbsentUncertain
Nudity69 (25.94%)184 (69.17%)13 (4.89%)
Explicit sexual activity2 (0.75%)264 (99.25%)0 (0.00%)
Either category71 (26.69%)182 (68.42%)13 (4.89%)

As shown in Table R4, 68.42% of human judgments identified neither nudity nor explicit sexual activity, despite the positive GPT assessments. However, 26.69% identified at least one unsafe category, indicating that NudeNet can also miss visually recognizable unsafe content. These findings suggest that the disagreement arises from both broader GPT assessments and missed detections by NudeNet, highlighting the complementary nature of the two metrics.

AuthorsOpenReview ↗

Rebuttal by Authors (2/3)

W(b)-1 and Q3. Empirical evidence supporting the theoretical analysis.

A5. We provide three empirical measurements using the S3 evaluation prompt: (1) reprojection leakage, (2) safe–unsafe subspace overlap, and (3) residual unsafe components after text-side intervention.

(i) Reprojection leakage. Using the S3 evaluation prompt, we evaluate the removal–reprojection sequence in Proposition 1 with the full unsafe-prototype span. We measure the remaining unsafe component relative to its original magnitude. As shown in Table R5, global reprojection reintroduces a component equivalent to 49.95% of its original magnitude, directly demonstrating the geometric mechanism analyzed in Proposition 1.

Table R5. Empirical measurement of reprojection leakage.

MethodAfter ErasingAfter Global Reprojection
SAFREE≈0%49.95%

(ii) Safe–unsafe overlap. We measure the overlap between the safe and unsafe subspaces used by TSP, obtaining ∥PSPC∥2=0.6166\|P_SP_C\|_2=0.6166.

(iii) Residual unsafe component of TSP versus SAFREE. We further measure the remaining unsafe components in the final text embeddings after each method's token selection and projection.

Table R6. Remaining unsafe component after each method's complete text intervention.

MethodRemaining unsafe component ↓
SAFREE90.27%
TSP73.90%

TSP retains fewer unsafe-subspace components than SAFREE, reflecting the combined effects of token selection and projection. Notably, TSP also achieves higher token-identification AUC/F1 (0.736/0.413 vs. 0.479/0.251 for SAFREE), demonstrating more accurate unsafe-token selection. Together, these findings help explain TSP's stronger erasure performance in Table 1, while the reprojection analysis confirms the leakage mechanism identified in our theoretical analysis.

W(b)-2 and Q1-1. TSP's improvement in S2.

A6. We clarify that the substantial reduction in S2 is primarily attributable to negative-prompt augmentation. In our implementation, applying TSP also adds the predefined unsafe concept set (Punsafe\mathcal{P}_{unsafe}) to the original negative prompt (c−c^{-}). As shown in Table R7, this augmentation alone reduces the Nude Score from 16.11% to 0%, even without TSP projection. Thus, the large improvement in S2 should not be attributed solely to TSP's safe-subspace projection.

Table R7. Effect of negative-prompt augmentation on Nude Score. Augmented (c^{-}) includes the predefined unsafe concept set Punsafe\mathcal{P}_{unsafe}.

Methodc−c^{-}: Nude ↓Punsafe\mathcal{P}_{unsafe} ∪ c−\cup \space c^{-}: Nude ↓
FMB-only16.11%0%

W(b)-3 and Q1-2. Selected Tokens and simpler conditioning perturbations.

A7. Across 20 S2 prompts, TSP selects 403 of 737. For one S2 prompt, the 21 selected subwords are ▁soft, ly, ▁lit, ▁indoor, ▁scene, ▁with, ▁woman, ▁standing, ▁plain, ▁wall, ▁lighting, ▁diffuse, ▁calm, ▁producing, ▁natural, ▁cinema, tic, ▁atmosphere, ▁a, ▁fixed, ▁camera. The table shows representative changes from before to after projection. After-projection percentages indicate the magnitude remaining. Table R8 reports representative embedding changes, normalized to each token's original magnitude.

Table R8. Change the before and after projection of selected tokens.

Selected tokenEmbedding norm (before → after)Unsafe-component magnitude (before → after)
woman100% → 25.80%100% → 65.01%
standing100% → 24.29%100% → 73.38%
lighting100% → 24.95%100% → 76.97%
camera100% → 20.04%100% → 68.38%

We further compare TSP with simpler conditioning perturbations under matched intervention strengths.

Table R9. Comparisons to simpler intervention.

InterventionMatching criterionNudeNet detection ratio ↓Imaging Quality ↑
Global embedding attenuationTotal displacement2.72%0.5688
Random-direction projectionPer-token output norm0.00%0.7062–0.7094
Token-wise attenuationPer-token output norm0.00%0.7064
Token-wise attenuationPer-token displacement0.00%0.7087
TSP safe projectionReference0.00%0.7068

As shown in Table R9, token-wise interventions using TSP-selected tokens achieve comparable erasure performance and imaging quality regardless of the projection direction, whereas global attenuation substantially degrades quality. These results highlight the importance of selecting which tokens to modify, rather than relying solely on the projection direction. This is further supported by TSP's higher token-identification AUC/F1 (0.736/0.413 vs. 0.479/0.251 for SAFREE).

AuthorsOpenReview ↗

Rebuttal by Authors (3/3)

W(b)-4 and Q5. Robustness evaluation.

A8. Following the reviewer's request, we evaluate UNITE under semantic distribution shifts, multiple target objects, and incomplete localization masks.

(i) Semantic shift. We evaluate 60 annotated prompts (12 per category) using a fixed concept dictionary that excludes the tested aliases and paraphrases.

Table R10. Token-selection robustness under semantic shifts.

TypeAUC ↑ (TSP / SAFREE / VideoEraser)TSP target-token miss rate ↓TSP non-target modification rate ↓
Clean0.919 / 0.178 / 0.7810.00%65.83%
Aliases/paraphrases0.597 / 0.389 / 0.58635.29%60.83%
Context/order changes0.733 / 0.433 / 0.40820.59%50.00%
Multiple target mentions0.753 / 0.317 / 0.44418.75%34.90%

TSP achieves the highest AUC across all four settings, although performance decreases for unseen aliases and paraphrases. The last two columns distinguish missed target tokens from modifications outside annotated target spans.

(ii) Multiple targets objects. We additionally evaluate images containing two cats to assess whether UNITE can erase multiple instances of the same concept.

Table R11. Erasing multiple target objects.

MethodCat detection ratio ↓≥2 cats detection ratio ↓
Base100%100%
UNITE0%0%

Qwen identifies two cats in every sampled Base frame but no cats in the corresponding UNITE frames. These results demonstrate successful erasure of multiple target instances in the evaluated setting.

(iii) Incomplete localization masks. We further perturb the detector masks in S2 to assess sensitivity to missed target regions.

Table R12. Robustness test against incomplete mask.

Mask interventionMask coverageNudeNet detection ratio ↓
Dilated mask100.00%0% (0/1,620)
Right-shifted mask71.06%11.91% (193/1,620)
Partial mask49.92%89.14% (1,444/1,620)

As shown in Table R12, erasure performance degrades as mask coverage decreases, highlighting FMB's dependence on accurate localization. Together, these experiments demonstrate UNITE's ability to erase multiple target instances while identifying its limitations under unseen textual expressions and incomplete spatial detection.

W(b)-5 and Q4. S4 scenario.

A9. UNITE targets unsafe concepts explicitly present in image or text conditioning, whereas S4 concerns unsafe generation from otherwise safe inputs. We could not identify reproducible S4 cases during our searching, preventing a controlled assessment. We acknowledge this as an applicability boundary of UNITE and leave such cases for future investigation.

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

This paper proposes UNITE, a training-free method for concept erasing in image-to-video (I2V) generation. It addresses unsafe concepts introduced by the input image, text prompt, or both. UNITE consists of two components. FMB (First-frame Masking and Blending) detects unsafe regions in the input image and regenerates them during sampling. TSP (Token-wise Safe Projection) projects concept-related text tokens onto a predefined safe subspace. Experiments on nudity and object erasing show that UNITE outperforms existing training-free methods while maintaining overall video quality.

Critical strengths
  1. The paper addresses a meaningful problem in I2V concept erasing. Existing text-based methods cannot effectively remove unsafe concepts introduced by the input image.
  2. The proposed method requires no additional training and achieves strong erasing results, especially in S2 and S3. The experiments also include component ablations, different detectors, and video quality evaluations.
Critical weaknesses
  1. FMB mainly applies region detection, masking, and blending to the first-frame latent, while TSP follows the common idea of concept subspace projection. Moreover, ConceptGuard (EMNLP 2026) has already studied unsafe concepts from both image and text conditions, including the same three scenarios. Therefore, the novelty of the proposed method and its differences from existing work remain unclear.
  2. Since FMB relies on an external detector to locate unsafe regions, a natural alternative is to edit the input image before video generation. ConceptGuard uses FLUX.1 Kontext for image editing and also suppresses unsafe text embeddings. However, the paper only compares against SAFREE and VideoEraser, which leave the input image unchanged. Without comparisons with such image editing methods, it is unclear whether FMB provides advantages over existing alternatives.
  3. In Table 5, adding TSP to FMB reduces the Nude score from 18.97 to 0.04 in S2, even though the text prompt is already safe. It is unclear why text projection provides such a large improvement in this setting. Moreover, the theoretical analysis does not provide empirical evidence of the proposed concept leakage or the actual overlap between safe and unsafe subspaces.
  4. In Table A4, the Dynamic Degree of UNITE drops to 5.00 in S1, compared with 90~100 for the baselines. Although the authors attribute this to suppressing unsafe actions, it remains unclear whether the method can preserve normal motion while removing unsafe concepts.
Questions for authors

Please see the weaknesses above. I am willing to reconsider my rating if these concerns are properly addressed.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors (1/2)

We appreciate Reviewer BdLP for your constructive feedback. We address your concerns with additional comparisons and analyses below.

W1. Difference against ConceptGuard.

A1. We thank the reviewer for highlighting this relevant prior work and will incorporate it into the final version. ConceptGuard primarily focuses on multimodal risk identification, using an external image editor (FLUX.1 Kontext) for visual erasure and projection-based text editing. In contrast, FMB releases detected unsafe regions from first-frame conditioning and regenerates them during video denoising, without requiring an additional generative image editor. For text-side erasure, ConceptGuard scores tokens using their orthogonal-complement components, whereas TSP directly measures each token's unsafe-subspace component and projects selected tokens onto a predefined safe subspace.

We further validate these differences experimentally. On 60 independently annotated prompts, TSP achieves higher unsafe token-identification AUC/F1 (0.736/0.413) than ConceptGuard (0.363/0.141) (Table R1). Moreover, Table R2 shows that ConceptGuard's external editing pipeline fails in object erasing, while UNITE achieves more effective erasure through the video-generation process.

Table R1. Unsafe token identification.

MethodAUC ↑F1 ↑
SAFREE0.4790.251
VideoEraser0.4540.202
ConceptGuard†0.3630.141
TSP0.7360.413

†In the official repo, they only released the weights without executable inference/forward code, so we reproduce ConceptGuard based on the paper.

W2. Compare with FLUX.1 Kontext like ConceptGuard.

A2. As requested, we compare FMB with FLUX.1 Kontext, the external image editor adopted by ConceptGuard. We replace FMB with FLUX.1 Kontext for reference-image editing, using 28 inference steps, guidance scale 2.5, seed 0, and the instruction "Remove [target] from the image completely. Fill in the background naturally. Keep everything else unchanged." For a controlled comparison, both methods use the same TSP module for text-side erasure, with all other video-generation settings held identical.

Table R2. Comparisons FLUX.1 Kontext and FMB.

MethodMickey Mouse: S2Mickey Mouse: S3Hello Kitty: S2Hello Kitty: S3
CLIP ↓LPIPS ↑CLIP ↓LPIPS ↑CLIP ↓LPIPS ↑CLIP ↓LPIPS ↑
FLUX + TSP0.98790.15650.98590.19800.92530.16130.93480.2390
UNITE0.64460.64840.64520.63020.57500.56940.61870.5177

Table R3. Computational cost.

MethodExternal image-editing time ↓
FLUX + TSP10.68 s
UNITE0 s — no separate editor

As shown in Table R2, UNITE consistently outperforms the FLUX.1 Kontext baseline across both concepts and scenarios. In particular, FLUX.1 Kontext frequently retains recognizable target identities, whereas FMB effectively regenerates these regions within the video model's denoising process. Furthermore, FLUX.1 Kontext requires an additional 10.68-second image-editing pass (Table R3), while FMB operates directly within the existing I2V generation process. These results demonstrate the effectiveness of in-model regeneration over external image preprocessing in our evaluated setting.

W3-1. Why Nude score reduces in S2 by adding TSP?

A3. These reduction is mainly caused by the negative prompt (c−c^{-}). When apply the TSP, we also add the concept set (Punsafe\mathcal{P}_{unsafe}) to the original negative prompt. As shown in Table R4, by adding the concept set to the negative prompt, Nude Score also decease to 0% without TSP.

The observed reduction in Nude Score is largely attributable to the negative prompt (c−c^{-}). In our implementation, applying TSP also augments the original negative prompt with the predefined unsafe concept set (Punsafe\mathcal{P}_{unsafe}). As shown in Table R4, this augmentation alone reduces the Nude Score from 16.11% to 0%, even without TSP.

Table R4. Effect of negative-prompt augmentation on Nude Score. Augmented (c−c^{-}) includes the predefined unsafe concept set Punsafe\mathcal{P}_{unsafe}.

Methodc−c^{-}: Nude ↓Punsafe\mathcal{P}_{unsafe} ∪ c−\cup \space c^{-}: Nude ↓
FMB-only16.11%0%
AuthorsOpenReview ↗

Rebuttal by Authors (2/2)

W3-2. Empirical evidence supporting the theoretical analysis.

A4. The basic empirical evidence is presented in Table 1 that existing methods cannot effectively erase in S3. As requested, we provide direct measurements of concept-component reintroduction and safe–unsafe subspace overlap.

Table R5. Empirical results concept leakage.

MethodAfter ErasingAfter Global Reprojection
SAFREE≈0%49.95%

First, we test the removal–reprojection sequence analyzed in Proposition 1 using the S3 prompt. We measure the Frobenius norm of the selected tokens’ unsafe components, normalized by its original unsafe component. The concept leakage is nearly eliminated by the initial removal, but global reprojection reintroduces a component equivalent to 49.95% of its original magnitude. This directly demonstrates the reintroduction mechanism analyzed in Proposition 1.

Separately, for the rank-one safe and unsafe subspaces used in our evaluated UNITE configuration, we obtain ∣∣PSPC∣∣2=0.6166||\mathcal{P}_\mathcal{S}\mathcal{P}_\mathcal{C}||_2=0.6166. This measurement quantifies the non-orthogonality accounted for by our residual-component bound.

W4. Can be UNITE preserve normal motion?

A5. Yes, general I2V motion remains comparable to the other erasing baselines. We already show this in Table 4 and A5 use the same 355 image-prompt pairs with five seeds.

Table R6. General VBench-I2V comparison. Full results presented in Table A5 in appendix.

MetricBaseSAFREEVideoEraserUNITE
Total Score89.2889.2789.0089.12
Motion Smoothness98.7398.8399.1798.93
Dynamic Degree32.5629.0117.1326.93

UNITE's Dynamic Degree is just 2.08 points below SAFREE and 9.80 above VideoEraser, with near-original Total Score. These results indicate that the substantial Dynamic Degree reduction in S1 is specific to the unsafe-action prompt, rather than a general degradation of motion generation.

Thank you again for your time and thoughtful feedback. As the discussion deadline approaches, we would appreciate hearing about any remaining questions or concerns while there is still time for us to address them. Please feel free to let us know if any part of our response needs further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.