- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes Immune2V, which adds a bounded adversarial perturbation to an image before release so that videos generated from it by image-to-video (I2V) models are distorted, as a defense against deepfakes made from personal photos. Treating the reference image as entering generation through a structural and a semantic stream, it combines a structural loss that pushes the perturbed image's structural condition away from the clean one (with a temporal coverage term on Wan 2.1) and a semantic loss that pulls early DiT features under aligned prompts toward those obtained with collapse-inducing prompts. Experiments on Wan 2.1 and HunyuanVideo-I2V compare against four baselines and include a loss ablation, input transformations, and cross-model transfer.
Critical strengths
S1. Optimizing and evaluating on 13B to 14B open-source DiT models (Wan 2.1 and HunyuanVideo-I2V) keeps the setting close to a realistic threat, and perturbations optimized on Wan 2.1 still lower all five metrics on CogVideoX and DynamiCrafter without re-optimization.
S2. The structural attack is built on a concrete observation about Wan 2.1. Because the reference image is zero-padded into a pseudo-video before the causal VAE, an image-level perturbation mostly affects the first few temporal latent slices, and the temporal coverage term is designed to spread the disruption across all of them. This is the part of the paper I found most informative.
Critical weaknesses
W1. The evaluation does not adequately support the claimed protection Sec. 3.1 defines success as "severe and persistent degradation across the entire generated video", but Table 1 reports only five averaged metrics. On Wan 2.1, SC drops from 95.85 to 90.34, which does not show how many videos actually become unusable. There is no per-sample failure criterion, human evaluation, or variance. The paper is motivated by deepfakes made from personal photos, yet the 50-image face subset has no identity-related evaluation, although Anti-I2V, one of the baselines, measures identity preservation with ArcFace. The robustness claim in Sec. 5.2 is also stronger than Table 3 shows. After JPEG compression, about two-thirds of the SC drop is lost, and results under diffusion-based purification are not reported.
W2. The distinction from Anti-I2V drawn in Sec. 2 does not hold, and the ablation isolates neither the specific design choices nor the two streams Sec. 2 states that existing I2V immunization methods "mostly attack the old UNet denoising backbone", but Anti-I2V already reports results on DiT models such as CogVideoX-5B and OpenSora and also targets VAE and denoiser representations, which is where the two losses here operate. What remains specific to this paper is the collapse-prompt feature target, the expectation over prompts, and the temporal coverage term. Table 2 supports combining the two objectives but does not isolate the contribution of these choices. Nor does it necessarily isolate the two streams. Eq. (8) uses computed from the whole , so unless the gradient through the structural condition is stopped, also acts on the structural stream. The motivating analysis is also underspecified. Fig. 3 does not state which image-level attack is used, and Fig. 4 does not specify how the semantic distance is computed.
Questions for authors
- Could you state the threat model, in particular which generation settings the attacker is assumed to control? On Wan 2.1, which CFG scale is used at test time, and are the random seeds different from those used to build the collapse trajectories? If possible, how much protection remains with different step counts, CFG scales, or prompt styles?
- What parameters are used for the three transformations in Table 3, and what are the five metrics when the same transformations are applied to clean images?
- How is the semantic distance in Fig. 4 computed, and which attack produces the "Struct. Attacked Image"? Fig. 5 draws the backward path of only through the semantic encoder. Is the gradient through the structural condition detached in the implementation?
- For each baseline adapted to the two models, which modules are attacked, and what are the iteration count and per-image runtime? On Wan 2.1, PhotoGuard's PIC (84.88) is close to that of the clean input (85.26), while alone, which also attacks the encoder latent, reaches 82.71. Which encoder and target latent does PhotoGuard use? What PSNR, SSIM, or LPIPS do the immunized images of each method reach?
- The paper states that the code will be released upon acceptance. Could you share an anonymized implementation during the discussion, in particular for the four reimplemented baselines?