Reproduce “ReDirector: Creating Any-Length Video Retakes with Rotary Camera Encoding” 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
Byeongjun Park, Byung-Hoon Kim, Hyungjin Chung, Jong Chul Ye
ReDirector provides a camera-controlled video retake generation system capable of processing dynamically captured, arbitrary-length input videos using a rectified Rotary Position Embedding (RoPE) and a novel Rotary Camera Encoding (RoCE). The method establishes improved geometric consistency and camera accuracy, demonstrating robust generalization to diverse camera trajectories and video lengths over previous approaches.
Key insights
- A rectified application of shared 3D RoPE across input and target videos is essential for length-agnostic positional encoding and spatiotemporal alignment.
- Embedding camera pose information as a camera-conditioned phase shift within RoPE (RoCE) provides a powerful mechanism for encoding fine-grained, physically grounded multi-view relationships.
- Integrating geometry-aware attention with RoCE helps disentangle dynamic objects from static backgrounds, maintaining geometric consistency even over long temporal spans.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.