Diffusion Model-based Video Perspective Reconstruction Technique
ReCamMaster, jointly developed by Zhejiang University and Racer Technology team, adopts an advanced text-to-video diffusion modeling architecture. The core technological innovation of the tool lies in combining the Wan2.1 model with the camera trajectory control algorithm, realizing the breakthrough of generating multi-view videos from a single video source. The team especially optimizes the spatio-temporal consistency processing between video frames, so that the generated new-view video can maintain action coherence. In terms of technical implementation, the system first performs feature extraction and 3D scene understanding on the input video, then re-renders the viewpoints according to the user-specified camera trajectory parameters, and ultimately outputs a multi-angle video with professional visual effects.
This diffusion model-based approach is better able to handle occlusion in complex scenes compared to traditional computer vision techniques, and generates sensible content beyond the original picture through deep learning. The team's open source version of the code already includes full Wan2.1 model support, but the highest precision internal model is not fully open yet due to commercial considerations.
This answer comes from the articleReCamMaster: Rendering Tool for Generating Multi-View Videos from a Single VideoThe
