The framework jointly predicts relit video frames and viewpoint-aligned per-frame warped environment maps. This joint formulation helps enforce geometry-light consistency without relying on prior camera poses. Relit-LiVE combines RGB-intrinsic fusion rendering, environment-map prediction, and robust multi-stage training with synthetic and real-world datasets to reduce common relighting failures such as spatial inconsistency, original-light leakage, and temporal flicker.
Relit-LiVE is useful for virtual production, video editing, VFX, scene relighting research, and data generation where lighting must be changed after capture. Its ability to infer aligned environment maps alongside the relit video makes it more useful than a simple style transfer or prompt-based relighting approach. With public code linked from the project, it is listed as a free open-source research tool.
