acceptodds
Under review as a conference paper at ICLR 2027

LETHE: Defending against Unauthorized Image-to-Video Generation via Value Diversity Collapse

24 views

Abstract

Image-to-video (I2V) diffusion models can animate a single reference image into high-quality, subject-consistent videos, posing threats of misuse as anyone can turn personal photos into fabricated videos without consent. A promising solution is to proactively add an adversarial perturbation to the image before release, such that videos generated from it would be unfaithful or visually corrupted. However, due to an insufficient understanding of what governs how subsequent frames read the reference image in modern I2V models, existing protections struggle to offer strong protection under common perturbation budgets. In this paper, we approach the problem by exploring this underlying reading mechanism. Through a convex geometric analysis, we theoretically identify the spread of the reference value vectors as a key factor governing this reading process, and empirically find that even a mild contraction of this spread can severely degrade the generated video, making it an efficient lever for protection. Guided by this insight, we propose LETHE, a novel protection method that directly disrupts this reading process. LETHE pairs two complementary objectives, where a value collapse objective drives the reference value vectors toward uniformity, while an attention attraction objective draws subsequent frames toward these collapsed tokens to amplify corruption. We further construct two human-centric I2V protection benchmarks and conduct extensive experiments with multiple modern I2V models, which demonstrate the state-of-the-art protection performance and versatility of LETHE.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
video generation, image-to-video models, adversarial examples

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_e68TOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LETHE, a proactive image protection method designed to prevent unauthorized image-to-video (I2V) generation. The central observation is that subsequent video frames acquire information from the reference image through attention operations, where the diversity of the reference value vectors constrains the diversity of the information that can be retrieved.

Critical strengths
  • Novel and well-motivated method: The paper provides an interesting perspective on I2V protection by targeting the reference-reading mechanism through value collapse and attention attraction. The theoretical analysis and mechanism experiments support the proposed design.
  • Strong experimental performance: LETHE consistently outperforms existing methods across different I2V models and datasets. The additional robustness and transferability experiments further strengthen the empirical evaluation.
Critical weaknesses
  • Gap between theoretical analysis and practical protection: The theoretical analysis focuses on value diversity within individual attention operations, but does not directly characterize the behavior of the complete I2V generation process. Therefore, the theoretical results provide motivation rather than a guarantee of protection effectiveness.
  • High computational overhead: Protecting a single image requires approximately 7–8 minutes on H100 hardware. This limits practical scalability, and the paper lacks a direct comparison of protection-generation costs with existing methods.
  • Limited generalization to challenging real-world images: The evaluation primarily focuses on carefully selected images with clear, frontal, and unobstructed faces. More experiments on occluded faces, non-frontal views, multiple people, and low-quality images are needed to establish practical robustness.
  • Incomplete component ablation: The paper does not evaluate attention attraction independently, making it difficult to fully isolate the necessity and contribution of each proposed objective.
Questions for authors

See weakness and solve my concerns.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/2)

Dear Reviewer e68T, we sincerely thank you for your precious time and valuable comments. We are deeply encouraged by your recognition of our novel and well-motivated method, and our strong performance. We hope the following responses address your concerns.


W1: Gap between theoretical analysis and practical protection.

A1: Thank you for this thoughtful comment! We fully agree that Prop. 1 characterizes a single attention operation rather than the complete generation process, and thus serves as a motivation rather than a guarantee. In fact, this is also why we did not rely on the bound alone in our paper (see "Empirical Validation" in Sec. 3 and the limitations in App. D.3). Besides, we also bridge this gap empirically from two aspects:

  • Intervention on the full model (Sec. 3). Directly contracting the reference values by only 20%–30% during the complete generation process already severely degrades the video (Fig. 2), whereas random perturbations of the same magnitude do not, i.e., the degradation is specific to the reduced spread.
  • Mechanism verification on protected images (App. D.1). In the full model, the images protected by LETHE indeed exhibit a notably reduced value spread and readout diversity (Fig. 8), which is accompanied by the degraded videos in Sec. 5. Moreover, the same collapse is also consistently observed on four other I2V models with different architectures (please kindly refer to A4 of Reviewer cwqY for details). These results are consistent with our analysis, i.e., the collapse predicted for a single attention operation is indeed observed in the complete generation process.

As such, we believe our theory provides a useful and empirically supported insight for designing the protection, although a formal end-to-end guarantee for the multi-layer and multi-step generation process remains challenging, which we leave as our future work. We will make this clearer in our revision.


W2: Computational overhead.

A2: Thank you for the practical comment! Following your suggestion, we measure the per-step time and peak GPU memory of all methods on both Wan2.1 and CogVideoX. For fairness, all methods are run on a single H100 GPU under identical settings (bf16 precision, block-wise gradient checkpointing, and 17/49 frames for Wan2.1/CogVideoX during optimization):

ModelMetricPhotoGuardSDS+AdvDMI2VGuardAnti-PromptOurs
Backprop through DiT✗✗✓✓✓✓
Wan2.1Time / step (s) ↓0.122.054.567.564.234.38
Peak Mem. (GB) ↓4.8545.8547.8850.9347.4046.18
ISM-A ↓0.7340.6760.6080.5820.5240.428
CogVideoXTime / step (s) ↓0.341.575.469.626.074.45
Peak Mem. (GB) ↓1.7412.2318.8823.1221.4618.39
ISM-A ↓0.4440.4090.4360.4920.3420.297
  • From the table, we have two observations. (1) The peak memory is largely determined by the surrogate model. Merely running a forward pass of the DiT (with backpropagation through the VAE only) already takes 44.7 / 12.2 GB on Wan2.1 / CogVideoX. LETHE adds only a small increment on top of it, and consumes less memory than AdvDM/I2VGuard/Anti-Prompt. (2) The time cost of LETHE is on par with AdvDM/Anti-Prompt and notably lower than I2VGuard, as our objectives are directly computed on the intermediate features (K/V) of a single forward-backward pass, without additional computation (e.g., I2VGuard's dual forward passes). PhotoGuard and SDS+ are cheaper in both aspects as they do not backpropagate through the DiT, yet their protection is much weaker.
  • We believe this overhead is acceptable, as it is only a one-time offline cost per image and can be offered as a service by parties with sufficient resources (e.g., social media platforms or dedicated protection providers). We will further discuss this scenario in our Appendix.

We will add these results and more discussion in our revision.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/2)

W3: Generalization to challenging real-world images (occluded faces, non-frontal views, multiple people, and low-quality images).

A3: Thank you for this constructive suggestion! We focused on clear frontal faces mainly because ArcFace (used for ISM) is most reliable on them, whereas by design, LETHE does not rely on any face-specific priors/component. Following your suggestion, we further build a challenging set of 100 images (25 for each of the four categories you mentioned). As ArcFace is less reliable here, we instead report two DINO-based metrics of VBench, i.e., I2V-Subject (similarity to the reference image) and Subject-Consistency (consistency across frames), together with QAF and QAV (on Wan2.1).

MetricCleanI2VGuardAnti-PromptOurs
I2V-Subject↓0.9660.9220.9240.843
Subject-Cons.↓0.9450.9340.8890.834
QAF↓0.7860.4920.6980.364
QAV↓0.8360.5590.7610.448

From the table, LETHE remains the most effective on all four metrics (e.g., QAF: 0.364 vs. 0.492 of the second-best), and the same trend also holds within each of the four categories and all metrics. These results indicate that the effectiveness of LETHE extends beyond clear frontal faces. We will add these results in our revision and release these images under the same terms as a part of our benchmark's challenging subset. Thank you again for making our evaluation stronger!


W4: Additional ablation study on attention attraction objective.

A4: Thank you for this helpful suggestion! Following your suggestion, we further evaluate the variant with Laa\mathcal{L}_{aa} only (mean of three runs, Wan2.1), which completes Tab. 2 as follows:

DatasetVariantISM-A↓ISM-F↓QAF↓QAV↓
CelebV-TextClean0.81730.87250.91230.9462
Laa\mathcal{L}_{aa} only0.65140.72820.55900.6403
Lvc\mathcal{L}_{vc} only0.58990.68270.52240.6009
Lvc+Laa\mathcal{L}_{vc}+\mathcal{L}_{aa}0.42800.53080.39700.4976
Kinetics-700Clean0.75840.82420.79740.8381
Laa\mathcal{L}_{aa} only0.62830.71640.54730.6293
Lvc\mathcal{L}_{vc} only0.50840.61270.46000.5324
Lvc+Laa\mathcal{L}_{vc}+\mathcal{L}_{aa}0.41250.49570.38520.4745

From the table, we have two observations (note that all variants use the same budget of ϵ=16/255\epsilon=16/255). (1) Each objective alone is already effective. Lvc\mathcal{L}_{vc} takes effect by directly collapsing the diversity of the reference values, while Laa\mathcal{L}_{aa} alone disturbs the attention distribution over the reference tokens and thus also degrades the generation, although generally less effectively than Lvc\mathcal{L}_{vc}. (2) Their combination is notably stronger than devoting the entire budget to either one (e.g., ISM-A on CelebV-Text: 0.651/0.590 → 0.428), as Laa\mathcal{L}_{aa} draws subsequent frames to read more from the collapsed values and thus notably amplifies the effect of Lvc\mathcal{L}_{vc}. This indicates that the two objectives are complementary, and both are necessary.

We will add these results to Tab. 2 in our revision. Thank you again for your insightful and constructive comments!

Thank you for your thorough responses and additional experiments. The revisions have addressed my concerns. I appreciate the authors' efforts in improving the manuscript. I have no further questions or concerns and will keep my positive score.

AuthorsOpenReview ↗

Thank You for Your Positive Feedback!

Dear Reviewer e68T, thank you so much for your positive feedback and for maintaining your positive score! It encourages us a lot! We also sincerely appreciate your prompt response and your constructive comments, which have helped us further improve our paper. We will ensure that the revised version includes all the new experiments (e.g., the complete ablation study and the evaluation on challenging images) and the corresponding discussions.

Reviewer_z3aMOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the proactive protection of personal images against unauthorized image-to-video (I2V) generation by modern DiT-based models. The authors first examine how subsequent-frame tokens read the reference (first-frame) tokens through attention. From a convex-hull perspective, they theoretically show that the variance of the reference-frame value vectors is a key factor in generation, and an intervention experiment further demonstrates that even slightly contracting the reference values can severely degrade the generated videos. Motivated by these observations, the paper proposes LETHE, which combines two objectives: a value collapse loss that minimizes the normalized variance of the reference value vectors, and an attention attraction loss that increases the average attention logit from subsequent-frame queries to the reference keys. The authors also introduce two human-centric I2V protection benchmarks. Experiments show that the proposed method outperforms five baselines, transfers to four other I2V models, and remains effective under common transformations and purification.

Critical strengths
  1. This paper studies a timely and important problem. DiT-based I2V models can now animate a single personal photo into a realistic video, whereas existing protections, even at the common budget of 16/255, remain largely ineffective against these models and leave the subject recognizable. Protecting personal images against this threat is therefore a practical and pressing problem.

  2. The protection objective is novel and is derived directly from the mechanism by which subsequent frames read the reference frame. The intervention experiment in Section 3 provides intuitive and convincing motivation, and the empirical gains are substantial. The analysis, the intervention, and the validation form a coherent logical chain. The attention-based perspective may also inform protection for other reference-conditioned generation tasks, which I find insightful.

  3. The paper is clearly written and easy to follow. The authors also construct two human-centric I2V protection benchmarks and plan to release them, which should facilitate the evaluation of future I2V protection methods.

  4. The evaluation is comprehensive to a large extent.

Critical weaknesses

Overall, I do not find major weaknesses in this paper, as both the method and the evaluation are clear and convincing. However, I have two moderate concerns and a few questions.

  1. This paper lacks a perceptual evaluation of the protected images. All methods are compared under the same ℓ∞\ell_\infty budget of 16/255, but the paper does not report how visible the perturbation is on the protected images. This matters for the intended use case: a user who shares a selfie on social media is unlikely to adopt a protection that leaves noticeable patterns on the face, and visible patterns may also prompt an attacker to apply purification before generation. It therefore remains unclear to what extent the stronger protection of LETHE comes at the cost of imperceptibility. Note that the visibility of the perturbation depends not only on the protection objective but also on the optimizer and on whether EOT is used, both of which may differ across methods. For a fair comparison, I suggest a controlled ablation in which all methods share the same optimizer and the same number of steps and run without EOT, with the protected images evaluated using metrics such as PSNR and LPIPS, as well as recent MLLM-based single-image quality metrics such as DeQA-Score (You et al., 2025) and Q-Insight (Li et al., 2025). This would clarify whether the objectives of LETHE improve effectiveness at the expense of imperceptibility.

  2. This paper does not include an evaluation against adaptive attacks. The mechanism of LETHE is quite specific: it mainly reduces the spread of the value vectors of the reference tokens (Eq. (4)) and increases the attention logits to these tokens (Eq. (5)). Since an attacker who uses an open-source I2V model has full control over its inference process, such a mechanism may be countered at inference time. For example, the attacker could first record the typical spread σclean\sigma_{\text{clean}} of the reference values in each layer and head on a few clean images. At inference, whenever the measured σV\sigma_V falls clearly below this level, the attacker rescales the values as vi←vˉ+s (vi−vˉ)v_i \leftarrow \bar{v} + s\,(v_i - \bar{v}) with s=σclean/σVs = \sigma_{\text{clean}} / \sigma_V, which reverses the contraction described in Section 3, and similarly restores the attention mass on the reference tokens to its clean level. This strategy is training-free, cheap, and applicable to most attention-based I2V models. It is unclear whether such a simple adaptive attack can undo the effect of LETHE.

  3. The author should also discuss the efficiency of the proposed method.

References

  1. You Z, Cai X, Gu J, et al. Teaching large language models to regress accurate image quality scores using score distribution[C]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025: 14483-14494.
  2. Li W, Zhang X, Zhao S, et al. Q-insight: Understanding image quality via visual reinforcement learning[J]. Advances in Neural Information Processing Systems, 2025, 38: 36802-36827.
Questions for authors
  1. (May affect my assessment) How do LETHE and the baselines compare in visual imperceptibility, measured by PSNR, LPIPS, or recent MLLM-based single-image quality metrics such as DeQA-Score and Q-Insight, when all methods share the same optimizer and run without EOT?

  2. (May affect my assessment) How does LETHE perform against adaptive attacks, such as the value re-spreading and attention re-balancing strategy described in W2?

  3. (Clarification) Wan2.1-I2V-14B additionally injects CLIP image features of the reference image through a separate cross-attention pathway, whereas CogVideoX-5B-I2V conditions on the reference image mainly through the latent of the first frame. Since the CLIP pathway is not considered by LETHE, how does this architectural difference affect the protection?

  4. (Suggestion) The mechanism of LETHE appears to be related to attention sinks (Xiao et al., 2024) and rank collapse in self-attention (Dong et al., 2021). A brief discussion of these connections would help position the analysis. What is the authors' view on this?

  5. (Suggestion) What are the efficiency and computational cost of the proposed method?

References

  1. Xiao G, Tian Y, Chen B, et al. Efficient streaming language models with attention sinks[C]//International Conference on Learning Representations. 2024, 2024: 21875-21895.

  2. Dong Y, Cordonnier J B, Loukas A. Attention is not all you need: Pure attention loses rank doubly exponentially with depth[C]//International conference on machine learning. PMLR, 2021: 2793-2803.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/2)

Dear Reviewer z3aM, we sincerely thank you for your precious time and thoughtful comments. We are deeply encouraged by your recognition of the timely problem, our principled, mechanism-derived objective, and our benchmarks contributions, and we hope the following responses address your concerns.


W1 & Q1: Imperceptibility under a controlled setting.

A1: Thank you very much for this constructive suggestion! Following your protocol, we run all methods with the same optimizer (sign-PGD with a step size of ϵ/20\epsilon/20 and zero initialization, i.e., the protocol of AdvDM), the same number of steps (100), and without EOT, such that the only difference among methods is the optimization objective. We then report PSNR, LPIPS (VGG), DeQA-Score, and Q-Insight of the protected images (CelebV-Text, Wan2.1, ϵ=16/255\epsilon=16/255).

MethodPSNR ↑LPIPS ↓DeQA ↑Q-Insight ↑
Clean––4.0133.898
AdvDM28.960.5823.0523.271
SDS+28.280.5612.8563.008
PhotoGuard31.320.4273.4743.486
I2VGuard29.820.4472.9763.246
Anti-Prompt30.010.4993.2983.303
Ours29.930.4383.1003.378

From the table, LETHE ranks top-3 on all four metrics (top-2 on LPIPS/Q-Insight), i.e., its imperceptibility is comparable to that of existing methods (e.g., within 0.1 dB of Anti-Prompt in PSNR). PhotoGuard, which only attacks the VAE encoder, obtains the best scores, yet its protection is the weakest. Meanwhile, LETHE still achieves the best protection performance under this implementation. These results suggest that the objectives of LETHE do not improve effectiveness at the expense of imperceptibility. We will add these results in our revision.


W2 & Q2: Adaptive attacks.

A2: Thank you very much for proposing this insightful and well-designed adaptive attack! Following your description, we implement it on Wan2.1. Specifically, we first record the clean spread σclean\sigma_{\text{clean}} and attention mass of each layer and head on 30 held-out clean images. At inference, whenever σV<τ⋅σclean\sigma_V<\tau\cdot\sigma_{\text{clean}}, we re-spread the reference values via vi←vˉ+s(vi−vˉ)v_i\leftarrow\bar v+s(v_i-\bar v) with s=σclean/σVs=\sigma_{\text{clean}}/\sigma_V, and similarly re-balance the attention mass to its clean level. We set τ\tau to the most aggressive value that still matches benign generation (τ=0.8\tau=0.8; see "Clean + Adaptive Attack"). The results on 50 CelebV-Text images are as follows:

SettingISM-A↓ISM-F↓QAF↓QAV↓
Clean0.8310.8730.9120.948
Clean + Attack0.8230.8610.9040.933
LETHE0.4410.5240.3710.471
LETHE + Attack0.4720.5790.3470.435

From the table, this adaptive attack only marginally weakens LETHE: (1) it recovers only 8% / 16% of the identity gap on ISM-A / ISM-F, although re-spreading is triggered in 34.8% of the heads (average s=1.88s=1.88), and (2) the video quality even drops further, and the recovery fails. This is possibly because rescaling only restores the magnitude of the spread, but not the token-specific information originally carried by the reference values. We will add these results in our revision.


Q3: The CLIP pathway in Wan2.1.

A3: Thank you for this insightful question! We fed clean and protected images into the two pathways separately. Protecting only the latent input gives an ISM-A of 0.437, close to the full protection (0.428), while protecting only the CLIP input gives 0.803, close to the clean level (0.817). This agrees with Wan2.1 (Wan et al., 2025), where the CLIP branch primarily provides global semantics instead of detailed information (e.g., face details). Hence, the identity is mainly reproduced from the first-frame tokens. We will add this to App. D.


Q4: Connections to attention sinks (Xiao et al., 2024) and rank collapse (Dong et al., 2021).

A4: Thank you for pointing out these insightful connections! We fully agree that both are closely related to our analysis. (1) Rank collapse. Dong et al. show that pure self-attention naturally drives all tokens toward a rank-one state as depth grows, which practical models counteract with skip connections and MLPs. In contrast, LETHE deliberately induces a similar collapse from the input side, as Lvc\mathcal{L}_{vc} is minimized when all reference values coincide. (2) Attention sinks. Sink tokens absorb a large share of attention while contributing little content. From this view, LETHE actively turns the reference tokens into an attention sink, where Laa\mathcal{L}_{aa} attracts the attention and Lvc\mathcal{L}_{vc} removes the distinct content. The key difference is that natural sinks are benign since they barely contribute to the output, whereas ours do not carry less content but homogenized content, i.e., subsequent frames are drawn to read a token-agnostic average instead of the reference details, which in turn corrupts the generated video. We will add this discussion in our revision.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/2)


W3 & Q5: Efficiency.

A5: Thank you for the practical comment! Following your suggestion, we measure the per-step time and peak GPU memory of all methods on both Wan2.1 and CogVideoX. For fairness, all methods are run on a single H100 GPU under identical settings (bf16 precision, block-wise gradient checkpointing, and 17/49 frames for Wan2.1/CogVideoX during optimization):

ModelMetricPhotoGuardSDS+AdvDMI2VGuardAnti-PromptOurs
Backprop through DiT✗✗✓✓✓✓
Wan2.1Time / step (s) ↓0.122.054.567.564.234.38
Peak Mem. (GB) ↓4.8545.8547.8850.9347.4046.18
ISM-A ↓0.7340.6760.6080.5820.5240.428
CogVideoXTime / step (s) ↓0.341.575.469.626.074.45
Peak Mem. (GB) ↓1.7412.2318.8823.1221.4618.39
ISM-A ↓0.4440.4090.4360.4920.3420.297
  • From the table, we have two observations. (1) The peak memory is largely determined by the surrogate model. Merely running a forward pass of the DiT (with backpropagation through the VAE only) already takes 44.7 / 12.2 GB on Wan2.1 / CogVideoX. LETHE adds only a small increment on top of it, and consumes less memory than AdvDM/I2VGuard/Anti-Prompt. (2) The time cost of LETHE is on par with AdvDM/Anti-Prompt and notably lower than I2VGuard, as our objectives are directly computed on the intermediate features (K/V) of a single forward-backward pass, without additional computation (e.g., I2VGuard's dual forward passes). PhotoGuard and SDS+ are cheaper in both aspects as they do not backpropagate through the DiT, yet their protection is much weaker.
  • We believe this overhead is acceptable, as it is only a one-time offline cost per image and can be offered as a service by parties with sufficient resources (e.g., social media platforms or dedicated protection providers). We will further discuss this scenario in our Appendix.

We will add these results and more discussion in our revision. Thank you again for your comprehensive and helpful comments!

Thank you for the detailed rebuttal. The new experiments are well designed and address most of my concerns.

Imperceptibility (W1/Q1). The controlled comparison shows that LETHE is comparable to the baselines on PSNR, LPIPS, DeQA-Score, and Q-Insight. This resolves my concern that the stronger protection comes at the cost of visibility.

Adaptive attack (W2/Q2). The value re-spreading and attention re-balancing attack recovers only a small part of the identity gap. The explanation that rescaling restores the magnitude but not the token-specific information is reasonable.

CLIP pathway, related work, and efficiency (Q3 to Q5). The pathway ablation, the discussion of attention sinks and rank collapse, and the cost table are all clear and satisfactory.

Please include these results and discussions in the revised version. I have no further questions at this stage. I will adjust my score accordingly.

AuthorsOpenReview ↗

Thank You for Your Positive Feedback & Raising the Score!

Dear Reviewer z3aM, thank you so much for your positive feedback and for raising the score! It encourages us a lot! We also sincerely appreciate your prompt response and your insightful suggestions (e.g., the controlled imperceptibility evaluation and the adaptive attack), which have greatly strengthened our paper. We will ensure that the revised version includes all the new experiments and the corresponding discussions.

Reviewer_cwqYOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper tackles the problem that existing I2V image protection methods fail to sufficiently disrupt the identity and appearance of an original image. The paper proposes LETHE which reduce the diversity of reference values via Value Collapse Loss, and focuses more on collapsed information for latter frames via Attention Attraction Loss. LETHE significantly outperforms SoTA I2V image protection methods such as I2VGuard and Anti-Prompt, and also shows strong black-box transferability and robustness after purification.

Critical strengths
  • Clear motivation: The paper shows a theoretical analysis of the source-conditioning mechanism in I2V models and provides a reasonable motivation for the proposed method.
  • Effective method: The proposed methods including Value Collapse Loss and Attention Attraction Loss effectively disrupt reference propagation.
  • Strong results: The proposed method outperforms existing SoTA I2V image protection methods (I2VGuard and Anti-Prompt) over multiple I2V models (Wan & CogVidX) and strong black-box transferability and robustness against various purification methods including transformation and generation approaches.
Critical weaknesses
  • No quantitative imperceptibility evaluation: The paper does not provide quantitative measures of the perceptual quality of protected images, which is one of the core evaluation in this task. An evaluation using metrics such as PSNR, SSIM, and LPIPS, along with comparisons against existing methods, would help clarify the trade-off between protection effectiveness and imperceptibility.
  • High computational cost: Generating a perturbation requires approximately 7–8 minutes per image on a single H100 NVL (94GB) GPU. which may limit its scalability and practical deployment. It would be helpful to provide trade-off between protection effectiveness and computational cost across diverse models and competing methods.
  • Limited evaluation: The paper mainly uses the human-centric benchmarks such as CelebV-Text and Kinetics-700. It is well-motivated but it is also limited to images containing clearly visible human faces Additional experiments on general I2V benchmark such as VBench (which is also used in I2VGuard and Anti-Prompt) would be beneficial to show the method’s effectiveness on general visual content beyond human subjects.
  • Limited validation across different conditioning methods: LETHE shows strong transferability across different I2V architecture. However it is unclear that the same value diversity collapse mechanism is responsible for the observed protection effects in these models. Since these models employ different source-image conditioning mechanisms, it remains unclear whether the proposed value collapse mechanism operates consistently across architectures.
Questions for authors

See weaknesses

7 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1/2)

Dear Reviewer cwqY, we sincerely thank you for your precious time and valuable comments. We are deeply encouraged by your recognition of our clear motivation, effective method, and strong empirical results, and we hope the following responses address your concerns.


W1: Quantitative imperceptibility evaluation.

A1: Thank you for the insightful suggestion! Following your suggestion, we report PSNR, SSIM, and LPIPS (VGG) between the protected and clean images, together with the protection strength (ISM-A from Tab. 1) for reference. All methods are evaluated with Wan2.1 under the same budget of ϵ=16/255\epsilon=16/255.

MethodPSNR↑SSIM↑LPIPS↓ISM-A↓
AdvDM28.960.6310.5820.608
SDS+30.100.7410.5090.676
PhotoGuard29.010.6400.3750.734
I2VGuard26.460.5320.5580.582
Anti-Prompt28.550.5930.4940.524
Ours29.880.7130.4300.428

From the table, LETHE ranks top-2 on all three metrics, i.e., its perturbation is comparable to or less perceptible than those of existing methods. These results suggest that LETHE achieves notably stronger protection at a comparable level of imperceptibility. We will add these results in our revision.


W2: Computational cost.

A2: Thank you for the practical comment! Following your suggestion, we measure the per-step time and peak GPU memory of all methods on both Wan2.1 and CogVideoX. For fairness, all methods are run on a single H100 GPU under identical settings (bf16 precision, block-wise gradient checkpointing, and 17/49 frames for Wan2.1/CogVideoX during optimization):

ModelMetricPhotoGuardSDS+AdvDMI2VGuardAnti-PromptOurs
Backprop through DiT✗✗✓✓✓✓
Wan2.1Time / step (s) ↓0.122.054.567.564.234.38
Peak Mem. (GB) ↓4.8545.8547.8850.9347.4046.18
ISM-A ↓0.7340.6760.6080.5820.5240.428
CogVideoXTime / step (s) ↓0.341.575.469.626.074.45
Peak Mem. (GB) ↓1.7412.2318.8823.1221.4618.39
ISM-A ↓0.4440.4090.4360.4920.3420.297
  • From the table, we have two observations. (1) The peak memory is largely determined by the surrogate model. Merely running a forward pass of the DiT (with backpropagation through the VAE only) already takes 44.7 / 12.2 GB on Wan2.1 / CogVideoX. LETHE adds only a small increment on top of it, and consumes less memory than AdvDM/I2VGuard/Anti-Prompt. (2) The time cost of LETHE is on par with AdvDM/Anti-Prompt and notably lower than I2VGuard, as our objectives are directly computed on the intermediate features (K/V) of a single forward-backward pass, without additional computation (e.g., I2VGuard's dual forward passes). PhotoGuard and SDS+ are cheaper in both aspects as they do not backpropagate through the DiT, yet their protection is much weaker.
  • We believe this overhead is acceptable, as it is only a one-time offline cost per image and can be offered as a service by parties with sufficient resources (e.g., social media platforms or dedicated protection providers). We will further discuss this scenario in our Appendix.

We will add these results and more discussion in our revision.


W3: Evaluation on VBench.

A3: Thank you for this constructive suggestion! By design, LETHE is content-agnostic, as its objectives only operate on the values and attention of the reference tokens without relying on any face-specific component. Following your suggestion, we further evaluate LETHE on VBench-I2V with Wan2.1. Specifically, we remove the human categories to focus on general visual content, and sample 50 images across the 9 remaining categories. As ISM is no longer applicable, we instead report two DINO-based metrics of VBench, i.e., I2V-Subject (similarity to the reference image) and Subject-Consistency (consistency across frames), together with QAF and QAV.

MetricCleanI2VGuardAnti-PromptOurs
I2V-Subject ↓0.9870.9470.9350.868
Subject-Cons. ↓0.9820.9570.9160.838
QAF ↓0.8820.6280.7600.388
QAV ↓0.9280.7310.8340.562

From the table, LETHE remains the most effective on all four metrics (e.g., QAF: 0.388 vs. 0.628 of the second-best), indicating that its effectiveness generalizes to general visual content beyond human subjects. We will add these results in our revision.

AuthorsOpenReview ↗

Rebuttal by Authors (Part 2/2)


W4: Mechanism validation across conditioning models.

A4: Thank you for this insightful question! As discussed in Sec. 3, the core mechanism of LETHE is to collapse the readout diversity of the reference tokens, i.e., D=1M∑j∥oj−oˉ∥2D=\frac{1}{M}\sum_j \lVert o_j-\bar{o} \rVert^2 with oj=∑iajivio_j=\sum_i a_{ji} v_i. Following your comment, we verify whether this also holds on the transferred models. Specifically, we transfer the perturbations crafted on Wan2.1 to the four models without modification, and measure the change of normalized readout diversity (relative to clean images) where the reference is read, i.e., the self-attention over the first-frame tokens for DiT-based models and the spatial self-attention for SVD (which concatenates the reference latent to every frame). Results are averaged over all layers, heads, and denoising steps on 50 images, and we also report the variation across random seeds for reference.

SettingLightX2VAccVideoLTX-VideoSVD
Clean (different seed)+2%+2%+1%−1%
Ours−70%−66%−30%−24%

From the table, the same collapse of readout diversity is consistently observed on all four transferred models, far beyond the variation across seeds. While this evidence is correlational, together with our intervention in Sec. 3 (where directly contracting the reference values already degrades the generated video), it is consistent with our hypothesis in App. D.3 that the transferred protection may also share the same mechanism, despite the different conditioning designs of these models.

We will add these results and more discussion in our revision. Thank you again for your valuable and constructive comments!

I appreciate the authors' efforts to thoroughly address my concerns. Given the impressive results, I have a few additional questions regarding the experiments.

  1. VBench evaluation on other models: The VBench evaluation was conducted only on Wan2.1. I wonder whether similar improvements can be observed on other models, such as LTX-Video. Since these models have different architectures, additional results would help demonstrate how consistently LETHE performs across different I2V models.

  2. Computational cost: Thank you for providing the additional computational cost analysis. However, I think reporting the total wall-clock time required to protect a single image would be more meaningful than reporting only the per-step time. Different methods may require different numbers of optimization steps, so the per-step time does not necessarily reflect the actual computational cost (e.g., LETHE uses 100 iters while Anti-Promtp originally uses 50 iters). In particular, I2VGuard requires an additional reference video generation process, which I would expect to introduce considerable computational overhead. It is unclear whether this cost is included in the reported measurements. Could the authors provide the total protection time per image for each method, including any additional video generation or preprocessing costs? This would provide a clearer and more practical comparison of computational efficiency.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.