Relit-LiVE

Relit-LiVE is a video relighting framework from a research team (authors include Weiqing Xiao, Hong Li, Xiuyu Yang, Hao Zhao and Beibei Wang), presented as a SIGGRAPH 2026 conference paper. It uses a diffusion process to produce relit video frames together with viewpoint-aligned, per-frame warped environment maps, without requiring prior camera poses.\n\nThe method combines an RGB-intrinsic fusion renderer, which feeds raw reference images into the rendering pathway to retain global illumination cues and material appearance, with per-frame environment map prediction. Training is multi-stage: latent interpolation creates multi-illumination supervision, and cycle-consistent self-supervised illumination learning improves temporal coherence on real videos. The project page lists image relighting at 1024 x 1472 and video relighting at 480 x 832 over 57-frame sequences. Applications shown include scene editing and video delighting.\n\nAccording to a YouTube creator, the tool lets users change light warmth, hardness and angle and project environment maps, and the code has been released, requiring a CUDA GPU with about 24 GB VRAM. The project page links to code, but no repository, weights or license details were available in the provided material, so these remain unconfirmed.