acceptodds
Desk rejected at ICLR 2027

RVDGuard: Robust Protection against Unauthorized Image-to-Video Generation via First-Step Velocity Deflection

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Image-to-video (I2V) generation models can animate a single photo into a realistic video, raising serious concerns about privacy violations and deceptive content creation. Existing proactive protection methods add imperceptible perturbations to images before release so that videos generated from them are degraded. However, we find that existing I2V protections suffer from two robustness limitations. On the input side, their perturbations are largely neutralized by common resizing, cropping, and compression. On the output side, the generated videos often remain visually clean, so an attacker can restore them or regenerate new videos from unaffected frames. To address these limitations, we propose RVDGuard, a robust protection framework that requires only a single denoiser evaluation per transformation at each optimization step. For input robustness, RVDGuard applies Expectation over Transformation, passing original and protected images through shared transformation instances and averaging the objective over them. For output robustness, RVDGuard deflects the direction of the first-step velocity predicted by the I2V model, maximizing a per-temporal-slice cosine dissimilarity between the predictions for protected and original images under a shared initial noise. Because the first update sets the course of the entire sampling trajectory, this deflection propagates through all subsequent steps and corrupts every generated frame. Experiments on CogVideoX-5B-I2V, Wan2.2-TI2V-5B, and Wan2.1-I2V-14B show that the VBench-I2V scores of videos generated from images protected by RVDGuard drop by on average relative to those generated from original images, compared with for the strongest baseline, and that RVDGuard remains effective under common image transformations and transfers to an unseen I2V model.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Image protection, Image-to-Video

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_J3znOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper addresses the robustness shortcomings of existing image-to-video (I2V) generation protection methods by proposing the RVDGuard defense framework. It achieves its objectives by applying input-side transformation and optimizing only the first step. Experiments validated the effectiveness of the framework on three mainstream I2V models: CogVideoX and the Wan series.

Critical strengths
  1. Precise problem identification: Poor robustness of input perturbations and the fact that the output remains usable are two common issues in real-world deployments, making this research valuable.

  2. The method is designed to be simple and efficient.

Critical weaknesses
  1. This approach is more of an engineering combination than a methodological innovation: The EoT method proposed in 2018 has been widely adopted, while the “first-step velocity perturbation” method simply simplifies existing methods to the first step (similar to “Perturbing the early stages” mentioned in DiffusionGuard ICLR’25).

  2. The paper only verifies the robustness of its own method and does not compare it with baseline methods, failing to demonstrate the motivation behind the proposed robustness.

Questions for authors
  1. Theoretically, why is First-Step Velocity Deflection more effective than full optimization? Are there any model-specific limitations?

  2. How much has robustness improved compared to baseline methods?

  3. Since the goal is to improve robustness, how does the method perform under diffusion-based purification attacks?

  4. Why do some of the results shown in the table 4 (transferability) outperform the non-transfer results?

  5. Only showing one visual result isn't enough to illustrate the findings. More visualizations and videos are needed.

4 comments
AuthorsOpenReview ↗

Response to Reviewer J3zn (1/4)

Dear Reviewer J3zn,

We appreciate your recognition of the precise problem identification and the simplicity of the method. We respond to your concerns on novelty positioning, baseline robustness, and purification below; W2 and Q2 are merged as they address the same issue.

W1: Limited methodological innovation and combination of existing techniques.

A1: We respectfully disagree that RVDGuard merely simplifies existing methods to the first step. We agree that EoT is widely used and that DiffusionGuard already optimizes at the first denoising step, so neither is our claimed novelty. Our contribution is what is optimized at that step and why it suffices for I2V.

DiffusionGuard maximizes the norm of the first-step predicted noise of the protected image alone, motivated by inpainting models copying unmasked details at the first step. We instead maximize the directional deviation between the first-step velocities of the protected and the original image under a shared noise. This distinction matters in I2V: injecting the induced velocity difference at step 0 alone degrades the video almost as much as injecting it at every step, whereas a random change of the same magnitude has almost no effect (Tables 5 and 6). Perturbations optimized for different steps are nearly orthogonal (Table 7), so multi-step optimization multiplies the cost for mixed gains (Table 9).

The two components are also not independent: optimizing step 0 alone needs one denoiser pass per transformation, which is what makes averaging over transformations at every iteration tractable.

AuthorsOpenReview ↗

Response to Reviewer J3zn (2/4)

W2&Q2: Missing robustness comparison against baseline protection methods.

A2: We add a robustness comparison with all five baselines and report the results in Table R1. The baselines fail to withstand common transformations. PhotoGuard, Anti-I2V, and I2VGuard, their protection is largely erased by simple post-processing. Immune2V is the most robust baseline and still lowers quality-related metrics such as Imaging Quality. RVDGuard, in contrast, keeps Subject Consistency between 0.694 and 0.802 under all transformations.

Table R1. Robustness results on Wan2.2-TI2V-5B.

SettingMethodSubject Consistency ↓I2V Subject ↓Aesthetic Quality ↓Imaging Quality ↓
JPEG compression, q=50q = 50PhotoGuard0.9400.9520.5900.666
Immune2V0.9040.8780.5060.610
Anti-I2V0.9410.9670.6180.665
Anti-Prompt0.9460.9670.6050.671
I2VGuard0.9370.9540.5950.657
RVDGuard0.7020.8680.4670.567
Gaussian noise, σ=0.05\sigma = 0.05PhotoGuard0.9390.9490.5910.630
Immune2V0.8920.8670.4860.648
Anti-I2V0.9410.9660.6030.640
Anti-Prompt0.9380.9630.5990.638
I2VGuard0.9350.9530.5980.614
RVDGuard0.7170.8720.4920.585
Resizing, ×0.5\times 0.5PhotoGuard0.9540.9560.6100.623
Immune2V0.9370.8900.5180.523
Anti-I2V0.9520.9710.6220.636
Anti-Prompt0.9560.9700.6220.632
I2VGuard0.9450.9590.6100.610
RVDGuard0.7650.8890.5040.601
Gaussian blur, σ=1.0\sigma = 1.0PhotoGuard0.9530.9580.6120.600
Immune2V0.9360.8900.5180.493
Anti-I2V0.9530.9730.6290.613
Anti-Prompt0.9520.9700.6230.610
I2VGuard0.9410.9590.6110.587
RVDGuard0.6940.8500.4890.558
Center cropping, ratio 0.7PhotoGuard0.9390.9200.5780.625
Immune2V0.9150.8510.4820.577
Anti-I2V0.9430.9330.5990.662
Anti-Prompt0.9400.9320.5850.639
I2VGuard0.9370.9130.5860.627
RVDGuard0.8020.9380.5340.588
AuthorsOpenReview ↗

Response to Reviewer J3zn (3/4)

Q1: Theoretically, why is First-Step Velocity Deflection more effective than full optimization? Are there any model-specific limitations?

A3: Theoretical support. We have added a formal proposition to Appendix C. For a flow-matching model integrated with a deterministic linear solver, such as Euler or the UniPC sampler of Wan, the final latent satisfies zN=z0+∑k=0N−1hkvkz_N = z_0 + \sum_{k=0}^{N-1} h_k v_k with ∑k=0N−1hk=−σ0\sum_{k=0}^{N-1} h_k = -\sigma_0, where σ0>0\sigma_0 > 0 is the initial noise scale. Since clean and protected generations share z0z_0, writing Δvk=vkδ−vk\Delta v_k = v_k^\delta - v_k gives zNδ−zN=−σ0Δv0+Rz_N^\delta - z_N = -\sigma_0 \Delta v_0 + R and hence ∥zNδ−zN∥≥(1−ρ)σ0∥Δv0∥\lVert z_N^\delta - z_N \rVert \ge (1 - \rho) \sigma_0 \lVert \Delta v_0 \rVert, where R=∑k=1N−1hk(Δvk−Δv0)R = \sum_{k=1}^{N-1} h_k (\Delta v_k - \Delta v_0) and ρ=∥R∥/(σ0∥Δv0∥)\rho = \lVert R \rVert / (\sigma_0 \lVert \Delta v_0 \rVert). The leading term −σ0Δv0-\sigma_0 \Delta v_0 is exactly the shift in the model's initial clean-video estimate, and when the two velocities have comparable norms, its magnitude grows monotonically with the cosine dissimilarity we maximize. Measured with the official samplers across 100 images per model, ρ\rho stays strictly below 1 on every image, with a mean of 0.68 on Wan2.1-I2V-14B and 0.74 on Wan2.2-TI2V-5B, retaining 51% and 55% of the first-step deviation in the final latent zNz_N. By contrast, a random first-step change of the same norm yields ρ=1.00\rho = 1.00 and retains only 4%–5%.

Comparison with full optimization. While the ideal goal is to maximize the final latent deviation ∥zNδ−zN∥\lVert z_N^\delta - z_N \rVert, back-propagating through the unrolled multi-step sampler largely fails in practice (0.900 vs. 0.659 in Table 9), whereas joint optimization over ten steps splits the fixed ℓ∞\ell_\infty budget across nearly orthogonal step-wise directions (Tables 7–8), performing stronger on some metrics and weaker on others at ten times the computational cost (Table 9). Our first-step objective directly maximizes the leading term −σ0Δv0-\sigma_0 \Delta v_0 with a single clean denoiser evaluation.

Model-specific limitations. The exact linear decomposition above holds for flow-matching models with deterministic linear solvers, namely Wan2.1 and Wan2.2. For CogVideoX, a vv-prediction model sampled with DDIM, an analogous weighted decomposition holds where ρ\rho averages 0.90 and 21% of the deviation is retained, still far above 3.5% for a random change, consistent with Table 6 where later-step injections also remain active. Finally, the method requires the model's first-step prediction to be conditioned on the input image.

AuthorsOpenReview ↗

Response to Reviewer J3zn (4/4)

Q3: Missing evaluation under diffusion-based purification attacks.

A4: We evaluate two representative diffusion-based purification attacks: DiffPure [1], and IMPRESS [2]. For DiffPure, we encode the image using Stable Diffusion 1.5 and add noise at step 60 of 1000 before reverse denoising. For IMPRESS, we follow its default setup with 1000 Adam optimization steps with a learning rate of 5×10−35 \times 10^{-3}, and an LPIPS loss weight of 0.01 with threshold 0.1. We evaluate the first 100 images of VBench-I2V with Wan2.2-TI2V-5B, and we will extend it to the full evaluation set in the revised version. As shown in Table R2, existing protection baselines are largely neutralized by both purification attacks, four of the five baselines return to at least 0.956, and Immune2V stays at 0.908 and 0.949. In contrast, RVDGuard maintains significant degradation under both purifications, keeping Subject Consistency down to 0.712 and 0.836, and Imaging Quality to 0.496 and 0.605. This demonstrates that deflecting the first-step velocity direction provides substantially stronger resilience against diffusion-based purification than existing baselines.

Table R2. Quantitative results against diffusion-based purification on Wan2.2-TI2V-5B.

SettingMethodSubject Consistency ↓I2V Subject ↓Aesthetic Quality ↓Imaging Quality ↓
DiffPure [1]Clean0.9730.9840.7220.713
PhotoGuard0.9600.9480.5980.714
Anti-Prompt0.9630.9540.6080.719
Immune2V0.9080.8700.5850.705
Anti-I2V0.9650.9670.6140.722
I2VGuard0.9590.9460.6040.701
RVDGuard0.7120.8610.5350.496
IMPRESS [2]Clean0.9720.9880.7070.704
PhotoGuard0.9580.9680.6070.688
Anti-Prompt0.9560.9710.6070.692
Immune2V0.9490.9170.5920.701
Anti-I2V0.9680.9780.6080.707
I2VGuard0.9590.9650.6040.669
RVDGuard0.8360.9160.5870.605

Q4: Unclear reason why transfer results in Table 4 outperform non-transfer results.

A5: The transfer results do not outperform the non-transfer results overall, and only a subset of metrics are lower because different target models have different sensitivities to image-conditioning perturbations, making raw scores not directly comparable across architectures. Specifically, DynamiCrafter is more sensitive to conditioning perturbations than the three DiT models, which is evident when comparing Table 4 with Table 1: although all baselines are optimized on Wan2.2-TI2V-5B alone, their Motion Smoothness on DynamiCrafter is lower than on Wan2.2-TI2V-5B across every single baseline, and four of the five baselines also yield lower I2V Subject and I2V Background on DynamiCrafter. RVDGuard follows the same pattern: four of the seven metrics remain lower in the non-transfer settings on Wan2.1-I2V-14B and CogVideoX-5B-I2V, and Subject Consistency, Background Consistency, and Aesthetic Quality remain lower on Wan2.2-TI2V-5B than on DynamiCrafter.

Q5: Insufficient visual results and video demonstrations.

A6: We thank the reviewer for this suggestion and have added six additional multi-model qualitative comparisons in Appendix J of the revised manuscript, along with the corresponding generated video files in the Supplementary Material.

References

[1] Nie, Weili, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Animashree Anandkumar. "Diffusion models for adversarial purification." International Conference on Machine Learning (ICML), PMLR, 2022.

[2] Cao, Bochuan, Changjiang Li, Ting Wang, Jinyuan Jia, Bo Li, and Jinghui Chen. "IMPRESS: Evaluating the resilience of imperceptible perturbations against unauthorized data usage in diffusion-based generative AI." Advances in Neural Information Processing Systems 36 (2023): 10657-10677.

Sincerely,

Authors

Reviewer_7TBwOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RVDGuard protects images against unauthorized image-to-video generation by optimizing bounded input perturbations. Its objective maximizes the cosine dissimilarity between clean and protected first-step velocity predictions, averaged across future temporal slices. Shared transformation instances provide robustness to input processing, while first-step optimization avoids differentiating through the full sampling trajectory.

Critical strengths
  1. Simple objective with meaningful supporting analysis. Comparing clean and protected predictions under shared noise and transformations isolates the perturbation’s effect, while first-step optimization avoids unrolling the sampler. Loss ablations and single-step injection experiments provide useful support for these design choices.

  2. Substantial gains with relevant robustness checks. The reported average relative VBench degradation is 25.7%, compared with 14.1% for the strongest baseline across three models. Evaluation also includes unseen transformation types, an unseen generation seed, and restored-frame regeneration, strengthening the empirical case beyond direct generation alone.

Critical weaknesses
  1. Confounded transfer comparison. Section 5.5 optimizes RVDGuard against three surrogate models, whereas all baselines transfer perturbations optimized on Wan2.2 alone. Table 4 therefore demonstrates transferability but does not isolate the advantage of the proposed objective from the benefits of ensemble optimization.

  2. Protection claims exceed the evaluation criterion. Lower average VBench scores do not establish that no usable, identity-preserving frame remains. The restoration-and-regeneration experiment is helpful, but the paper lacks an explicit usability criterion and per-image attacker success rates to substantiate its categorical claim that every generated frame is unusable.

Questions for authors
  1. How is “normalized ℓ2” defined in Figure 4b? If it denotes squared distance between individually unit-normalized velocities, it equals twice the cosine dissimilarity in Eq. 2; what distinguishes the compared objectives?

  2. What text condition is used during perturbation optimization, and does the defender need access to the attacker’s eventual generation prompt?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 7TBw (1/2)

Dear Reviewer 7TBw,

We thank you for the careful reading and for recognizing the simplicity and supporting analysis of our objective. Your two critical points, the confounded transfer comparison and the over-strong “every frame” claim, are well taken, and we address them directly below.

W1: Confounded transferability comparison with multi-model ensemble optimization.

A1: We have added an experiment that optimizes RVDGuard on Wan2.2-TI2V-5B alone and transfers it to DynamiCrafter, matching the exact surrogate setting of all baselines. Following Appendices B, F, and G, this evaluation is conducted on the first 100 images of VBench-I2V, and we will extend it to the full set in the revised version. As shown in Table R1, single-surrogate RVDGuard still achieves the lowest values across all seven metrics among all single-surrogate methods. These results show that RVDGuard retains a transfer advantage when all methods use the same single surrogate.

Table R1. Transferability to DynamiCrafter.

MethodSurrogate modelsI2V Subject ↓I2V Background ↓Subject Consistency ↓Background Consistency ↓Aesthetic Quality ↓Imaging Quality ↓Motion Smoothness ↓
PhotoGuardWan2.2-TI2V-5B0.8830.9040.8870.9200.5910.6640.914
Anti-PromptWan2.2-TI2V-5B0.9130.9290.9100.9290.6160.6690.932
Immune2VWan2.2-TI2V-5B0.8720.8940.8790.9160.5920.6580.910
Anti-I2VWan2.2-TI2V-5B0.9220.9340.9240.9370.6300.6730.938
I2VGuardWan2.2-TI2V-5B0.8820.9040.8830.9170.5890.6610.917
RVDGuardWan2.2-TI2V-5B0.7470.8090.7550.8380.5120.5670.904

W2: Overstated protection claims without explicit frame usability criteria.

A2: We agree that video-level average VBench scores alone do not establish that no usable frame remains, and we have added a framewise DINOv2-similarity criterion with per-image protection success rates. Following Appendix F, we inspect every individual generated frame after excluding the first frame: a frame is counted as usable if its DINOv2 similarity to the original image reaches the 5th percentile of frames generated from clean images, regardless of visual artifacts, and protection succeeds on an image only if no single generated frame exceeds this threshold. As shown in Table R2 on the 100 images of VBench-I2V, RVDGuard achieves the highest per-image protection success rate on Wan2.1-I2V-14B and Wan2.2-TI2V-5B, and is comparable to Immune2V on CogVideoX-5B-I2V despite Immune2V using four times the perturbation budget. Since protection does not succeed on every single image, we have softened the categorical "every frame" statements in the revised manuscript, and will include the full-set per-image rates and user study in the revised version.

Table R2. Per-image protection success rate: the percentage of images for which no generated frame is usable.

MethodWan2.1-I2V-14BWan2.2-TI2V-5BCogVideoX-5B-I2V
Clean2%0%0%
PhotoGuard39%4%19%
Anti-Prompt82%23%38%
Immune2V74%51%84%
Anti-I2V15%1%1%
I2VGuard68%4%26%
RVDGuard93%70%83%
AuthorsOpenReview ↗

Response to Reviewer 7TBw (2/2)

Q1: Unclear definition of normalized L2 distance in Figure 4b.

A3: The normalized ℓ2\ell_2 distance in Figure 4b is defined on each slice as ∥v0,bδ−v0,b∥2/∥v0,b∥2\lVert v_{0,b}^\delta - v_{0,b} \rVert_2 / \lVert v_{0,b} \rVert_2, which normalizes the difference only by the clean velocity norm without unit-normalizing v0,bδv_{0,b}^\delta. Writing r=∥v0,bδ∥2/∥v0,b∥2r = \lVert v_{0,b}^\delta \rVert_2 / \lVert v_{0,b} \rVert_2 and θ\theta for the angle between the two velocities, this objective equals r2+1−2rcos⁡θ\sqrt{r^2 + 1 - 2r\cos\theta}, which reduces to a monotone function of the cosine dissimilarity 1−cos⁡θ1 - \cos\theta in Equation 2 only when r=1r = 1. When rr is unconstrained, normalized ℓ2\ell_2 spends part of its gradient scaling the magnitude ratio rr rather than rotating θ\theta; scaling the step length along the same direction is largely absorbed by later denoising steps, whereas rotating the direction steers the trajectory off the clean manifold. Consistent with Table 5, on the 100 images of VBench-I2V with Wan2.2-TI2V-5B, an unaligned random first-step change of the same magnitude retains only 4% of its relative deviation in the final latent zNz_N, compared with 55% for RVDGuard, and we have made these definitions explicit in Section 5.3.

Q2: Unclear text prompt condition during perturbation optimization.

A4: During optimization, RVDGuard conditions on the official VBench-I2V text prompt paired with each image, which is also the prompt used for generation in Table 1. The defender does not need the attacker's eventual prompt for the protection to remain effective. To verify this, we keep the protected images unchanged and generate each video with the official prompt of a different image in the dataset, which is unrelated to the image content. As shown in Table R3 on VBench-I2V with Wan2.2-TI2V-5B, although the degradation is milder than under the matched prompt, RVDGuard still preserves clear protection under unrelated prompts, keeping Subject Consistency down at 0.729, Aesthetic Quality at 0.523, and Imaging Quality at 0.591. The protection is thus weakened but remains effective under a prompt that differs completely from the one used in optimization. Moreover, an attacker who wants to animate the image in practice would typically describe its content rather than using an unrelated prompt, staying close to the matched setting. We have clarified the prompt setting in Section 5.1.

Table R3. Prompt robustness on Wan2.2-TI2V-5B.

MethodI2V Subject ↓I2V Background ↓Subject Consistency ↓Background Consistency ↓Aesthetic Quality ↓Imaging Quality ↓Motion Smoothness ↓
Matched prompt
RVDGuard0.8280.8500.6610.8170.4520.5490.976
Unrelated prompt
Clean0.9830.9860.9240.9440.6420.7140.986
RVDGuard0.9020.9210.7290.8310.5230.5910.963

Sincerely,

Authors

Reviewer_2tHkOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method to immunize images against malicious I2V generation. The authors note that existing methods are not robust to common image transformations like crop, resize, gaussian noising and JPEG purification. Furthermore, the immunized result still has semantic content which can be restored and used for malicious I2V generation. The proposed method RVDGuard uses expectation over transformations to optimize a perturbation that is robust to common transformations. The adversarial objective involves maximising the cosine dissimilarity between velocities predicted with the clean and perturbed images as conditionings.

Critical strengths
  1. Identifies immunization robustness as a limitation of existing I2V models
  2. Simple idea of deflecting the first-step velocity peredcition and promising results
Critical weaknesses
  1. The vulnerability of image immunization to simple transformations has already been already studied in the I2I setting. EditShield [1], for example, incorporates EoT into perturbation optimization to improve robustness against such transformations. The input robustness component is thus an application of the same concept to I2V models rather than a core novelty.
  2. Using the timesteps for for immunization: This is a somewhat known result in the comunity. Works like DeContext[2] and VidFreeze[3] explicitly show similar results in their ablations, showing that using initial sampling steps during optimizations are much more effective.
  3. Insufficient qualtitative results. The authors show only 1 qualititive sample of theire result. Please consider adding more results, at least in the supplementary. 4.It is unclear which prompt was used for optimization. The authors should clarify whether they assume knowledge of the prompt beforehand. If not, then prompts used using optimization and inference should be explicitly metioned. A study of prompt robustness should also be done.
  4. The method is tested against 3 models (2 of which are Wan). I suggest the method also be tried on other model families like Hunyuan and LTX, which are also opensource and popularly used in I2V protection literature.
  5. Potential inconsistincy between qualitative and quantitative results. The qualitative result in Fig3 shows near complete loss of semantic information in generated video from RVDGuard-immunized-image. Quantitively this should corespond to a huge drop in all VBench metrics (except motion smoothness). But table 1 does not show drop which are as significant as expected. I highly recomend sharing more result samples with more frames to be more transparent about the results.

Minor Points: 7. The authors can Include other I2V protection baselines that induce temporal freezing in videos for completeness like Medusa [4], VidFreeze [3]. 8. Missing Human Evaluation for noise impreptibilty and protection effectiveness.

[1] Chen, Ruoxi, et al. "Editshield: Protecting unauthorized image editing by instruction-guided diffusion models." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

[2] Shen, Linghui, Mingyue Cui, and Xingyi Yang. "DeContext as Defense: Safe Image Editing in Diffusion Transformers." arXiv preprint arXiv:2512.16625 (2025).

[3] Chowdhury, Rohit, et al. "Vid-freeze: Protecting images from malicious image-to-video generation via temporal freezing." arXiv preprint arXiv:2509.23279 (2025).

[4] Yu, Hongwei, et al. "MEDUSA: Motion Elimination in Diffusion Using Spectral Attack." Forty-third International Conference on Machine Learning

Questions for authors
  1. What was the prompt used in optimization?
  2. Are the colored frame artifacts shown in the video uniformly observed in the generated video? (Sharing the code or generated videos from their method would've helped to confirm this)
  3. What was the number of frames used during optimization?
  4. Can you do a study of how the I2V metrics change as you keep increasing the number of generated frames for already immunized images? Coresponding qualititive result would be appreciated. It would be insightful to see the limit of protection provided by RVDGuard
6 comments
AuthorsOpenReview ↗

Response to Reviewer 2tHk(1/4)

Dear Reviewer 2tHk,

We sincerely thank you for your detailed and constructive review. Your critiques on novelty positioning, model coverage, and qualitative evidence have directly shaped the revised version. Below, we respond to every Weakness (W), Question (Q), and Minor point (M); merged items are marked accordingly.

W1&W2: Limited novelty in transformation robustness and early timestep optimization.

A1: We respectfully disagree that RVDGuard is an application of EoT combined with a known early-timestep result. Our contribution is the first-step velocity deflection objective and the mechanistic findings behind it. EoT is a standard tool, and we do not claim it as a novelty.

EditShield applies EoT to a latent-shift objective for image editing. Our objective compares the first-step velocities of the protected and the original image under a shared noise, so the same transformation instance must be applied to both images.

DeContext and Vid-Freeze show in ablations that attention objectives work better at early timesteps. Our finding is different: for the velocity predicted by the denoiser, step 0 alone is sufficient, and what matters is the direction of the velocity difference rather than its magnitude. Injecting this difference at step 0 alone degrades the video almost as much as injecting it at every step, whereas a random direction of the same magnitude has almost no effect (Tables 5 and 6). Optimizing only step 0 also needs a single denoiser pass per transformation, which is what makes transformation averaging affordable.

W3&W5&Q2: Insufficient qualitative results and potential inconsistency with quantitative metrics.

A2: We thank the reviewer for this careful observation and have added six further qualitative comparisons in Appendix J, covering two more images on all three models with all baselines, along with the corresponding videos in the supplementary material. To explain the gap between Figure 3 and Table 1, we classify every protected video in Table 1 by two video-level criteria: whether it still resembles the input image in DINOv2 similarity, and whether its visual quality degrades in Imaging Quality, using the 5th percentile of clean videos as the threshold for both. Three types of outcomes occur. First, videos that degrade into noise, textures, or colored blocks, as in Figure 3, account for 60%, 42%, and 85% of the images on Wan2.1-I2V-14B, Wan2.2-TI2V-5B, and CogVideoX-5B-I2V; directly answering Q2, when such colored-frame artifacts occur, they appear uniformly across all generated frames after the initial conditioning frame. Second, videos that remain sharp and smooth without colored artifacts, but whose content becomes unrelated to the input image, account for 35%, 45%, and 5%. This second type keeps Imaging Quality, Aesthetic Quality, and Motion Smoothness close to clean videos, explaining why these averages drop less in Table 1 than Figure 3 suggests. Third, protection fails on the remaining 5%, 13%, and 10%. We will include this breakdown in the revised version.

AuthorsOpenReview ↗

Response to Reviewer 2tHk (2/4)

W4a&Q1: Unclear optimization prompt settings and missing prompt robustness study.

A3: We have clarified the prompt settings in Section 5.1 and conducted a prompt robustness study in Table R1. During optimization, RVDGuard conditions on the official VBench-I2V text prompt paired with each image (line 256-257), so Table 1 corresponds to the matched-prompt setting. Since the defender does not know the attacker's prompt in practice, we additionally evaluate the opposite extreme without re-optimizing the protected images: each video is generated using the official prompt of a different image in the dataset that is unrelated to the input image content. As shown in Table R1 on VBench-I2V with Wan2.2-TI2V-5B, although the degradation is milder than under the matched prompt, RVDGuard still maintains clear protection under unrelated prompts, keeping Subject Consistency down at 0.729, Aesthetic Quality at 0.523, and Imaging Quality at 0.591. This shows that protection is weakened but remains effective.

Table R1. Prompt robustness on Wan2.2-TI2V-5B.

MethodI2V Subject ↓I2V Background ↓Subject Consistency ↓Background Consistency ↓Aesthetic Quality ↓Imaging Quality ↓Motion Smoothness ↓
Matched prompt
RVDGuard0.8280.8500.6610.8170.4520.5490.976
Unrelated prompt
Clean0.9830.9860.9240.9440.6420.7140.986
RVDGuard0.9020.9210.7290.8310.5230.5910.963

W4b: Limited model evaluation and missing Hunyuan and LTX families.

A4: We add LTX-Video 0.9.5 and HunyuanVideo-1.5 on 100 images of VBench-I2V and we will extend it to the full set in the revised version. For LTX-Video, we use its official setting with 768×512 resolution, 121 frames, 40 steps, and guidance scale 3, and apply RVDGuard with the same budget as in the paper. As shown in Table R2, RVDGuard is as effective as on the models of Table 1, reducing Subject Consistency from 0.965 to 0.622 and Imaging Quality from 0.709 to 0.406. For HunyuanVideo-1.5, we optimize and evaluate with 49 frames at its official 480p setting. The protection is weaker on this model: the generated videos keep a high visual quality but drift away from the input image; for example, the camera leaves the scene or the background is replaced, which lowers Subject Consistency from 0.959 to 0.813 and I2V Subject from 0.986 to 0.855.

Table R2. Quantitative results of RVDGuard on additional I2V model families.

ModelMethodI2V Subject ↓I2V Background ↓Subject Consistency ↓Background Consistency ↓Aesthetic Quality ↓Imaging Quality ↓
LTX-Video 0.9.5Clean0.9830.9880.9650.9670.5990.709
LTX-Video 0.9.5RVDGuard0.7250.7720.6220.8210.3830.406
HunyuanVideo-1.5Clean0.9860.9850.9590.9620.7200.703
HunyuanVideo-1.5RVDGuard0.8550.8950.8130.8690.6430.620

Q3: Unclear number of video frames used during optimization.

A5: Optimization uses the same number of frames as the generation: 49 frames for CogVideoX-5B-I2V and Wan2.2-TI2V-5B, and 17 frames for Wan2.1-I2V-14B, as listed in Section 5.1 (lines 247–249).

AuthorsOpenReview ↗

Response to Reviewer 2tHk (3/4)

Q4: Missing study on protection performance under increasing generated frames.

A6: We add a study on how the protection of RVDGuard changes when the attacker generates longer videos than those used during optimization. Due to the limited rebuttal period, we conduct it on Wan2.1-I2V-14B, whose protected images are optimized with 17-frame videos, and generate 17, 49, 81, and 161 frames from the same protected images. We evaluate 50 images for the four lengths, and will include the complete study on all models in the revised version. We provide four generated videos in the supplementary material. As shown in Table R3, RVDGuard keeps good protection up to the 81-frame length. At 49 and 81 frames, about three and five times the optimization length, Subject Consistency stays at 0.726 and 0.638, compared with 0.966 and 0.927 for videos generated from original images. At 161 frames, about ten times the optimization length, the protection becomes weaker, with Subject Consistency of 0.801, although all metrics remain below those of original images. This reveals a limitation of RVDGuard: its effect is diluted when the video grows far beyond the length used during optimization.

Table R3. Protection of RVDGuard on Wan2.1-I2V-14B when generating longer videos than the 17 frames used for optimization.

FramesMethodI2V Subject ↓I2V Background ↓Subject Consistency ↓Background Consistency ↓Aesthetic Quality ↓Imaging Quality ↓Motion Smoothness ↓
17Clean0.9850.9860.9720.9630.6430.7130.984
RVDGuard0.6890.7290.5470.7690.3530.4310.964
49Clean0.9870.9870.9660.9630.6330.7060.978
RVDGuard0.7980.8320.7260.8470.4720.5650.954
81Clean0.9650.9770.9270.9410.6040.7020.964
RVDGuard0.7270.7780.6380.8110.4190.4650.927
161Clean0.9720.9760.9570.9510.6310.7140.979
RVDGuard0.8380.8720.8010.8720.5790.6240.978

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

Related papers

Discussion (0)

Sign in to comment.