ReCamMaster
ReCamMaster is an ICCV'25 Best Paper Finalist research project that enables camera-controlled generative rendering from a single video. Developed by researchers from Zhejiang University, Kuaishou Technology, CUHK, and HUST, the system allows users to re-capture real-world video footage with novel camera trajectories using a simple and effective video conditioning scheme. The software leverages advanced generative models to synthesize new views of existing videos, effectively simulating virtual camera movements through complex scenes. Key features include support for diverse camera paths, the release of a training and inference codebase, and the provision of a pre-trained model checkpoint compatible with Wan2.1. The project is accompanied by the MultiCamVideo Dataset, a multi-camera synchronized video dataset rendered in Unreal Engine 5 to support training and evaluation. While the open-source repository provides a functional reference implementation, the authors note that performance may vary compared to thei