Arxiv paper – TrajectoryCrafter: Redirecting Camera Trajectory for Monocular Videos via Diffusion Models

In this episode, we discuss TrajectoryCrafter: Redirecting Camera Trajectory for Monocular Videos via Diffusion Models by Mark YU, Wenbo Hu, Jinbo Xing, Ying Shan. TrajectoryCrafter is a new method that precisely redirects camera paths in monocular videos by separating view changes from content generation. It uses a dual-stream conditional video diffusion model that combines point cloud renders with source videos to ensure accurate views and coherent 4D content. By training on a hybrid dataset of monocular and multi-view videos with a double-reprojection strategy, TrajectoryCrafter achieves robust performance across diverse scenes.