LETHE: Defending against Unauthorized Image-to-Video Generation via Value Diversity Collapse
LETHE protects photos from unauthorized image-to-video generation by collapsing reference value diversity.
Researchers propose LETHE, a proactive perturbation defense so image-to-video models produce unfaithful or visually corrupted videos from protected photos. A convex geometric analysis identifies the spread of reference value vectors as what governs how later frames read the reference image. LETHE drives those vectors toward uniformity and attracts later-frame attention to the collapsed tokens. On two new human-centric I2V benchmarks, it reports state-of-the-art protection across multiple modern models.
- Reference value-vector spread governs how later I2V frames read an image.
- A mild contraction of that spread can severely corrupt generated video.
- LETHE combines value collapse with attention attraction under typical budgets.
- Two human-centric benchmarks show state-of-the-art protection across modern I2V models.
From the original article
1 min read at arxiv.orgImage-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.
By Boheng Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.