Introducing Relit-LiVE: A Novel Framework for Physically Consistent Video Relighting

Researchers have developed Relit-LiVE, an innovative video relighting framework that enhances the stability and physical consistency of relit videos without prior camera pose knowledge. Unlike traditional methods that rely on accurate intrinsic decomposition, which often leads to visual distortions and artifacts, Relit-LiVE incorporates raw reference images into the rendering process. This approach allows the model to recover essential scene cues lost in traditional methods. The framework also features a unique environment video prediction formulation that generates relit videos alongside per-frame environment maps in a unified diffusion process, ensuring strong geometric-illumination alignment. Experimental results show that Relit-LiVE outperforms existing video relighting and neural rendering techniques across various benchmarks, supporting applications such as scene-level rendering, material editing, and streaming video relighting. The project can be accessed at https://github.com/zhuxing0/Relit-LiVE.