Reproduce “SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chao Wang, Zijin Yang, Yaofei Wang, Yuang Qi, Weiming Zhang, Nenghai Yu, Kejiang Chen
SWIFT, or Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution, offers a method to identify the originating generative model of a video by leveraging the intrinsic temporal mapping of its 3D Variational Autoencoder. The system achieved an average attribution accuracy of 94.0% across five state-of-the-art video generation models, demonstrating efficiency and requiring minimal or no training samples.
Key insights
- The intrinsic temporal up/down-sampling within a 3D VAE creates a "Pixel Frames (many) <-> Latent Frame (one)" mapping, which, if corrupted, leads to a measurable increase in reconstruction loss.
- This differential in reconstruction loss between normal and temporally corrupted video segments serves as a reliable, model-specific fingerprint for attribution.
- Adaptive thresholding using non-parametric methods like KDE allows for robust attribution with minimal (few-shot) or even zero-shot samples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.