Reproduce “AnchorSync: Global Consistency Optimization for Long Video Editing” 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
Zichi Liu, Yinggui Wang, Tao Wei, Chao Ma
AnchorSync, a two-stage diffusion-based framework, enables high-quality long video editing by ensuring both global consistency and temporal coherence. It achieves superior performance in long-term consistency with I-I CLIP Sim* of 97.84 and I-I CLIP Sim† of 94.11, and better structural preservation with the lowest Warp Error (3.78) and Canny Error (10.36) compared to existing methods.
Key insights
- Decoupling long video editing into sparse anchor frame processing and dense intermediate frame interpolation effectively manages global consistency and local coherence separately.
- Bidirectional Attention and progressive pairwise denoising across anchor frames are crucial for propagating consistent edits and structural alignment throughout the entire video sequence.
- Multimodal conditioning (optical flow and edge maps) during interpolation, combined with bidirectional temporal frame fusion, effectively preserves original video motion and structure, preventing flickering and artifacts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.