Lume-Palette relights indoor scenes with spatially controlled multi-view AI
A new two-stage framework achieves photorealistic, multi-view consistent indoor scene relighting...
Deep Dive
Chenjian Gao, Linning Xu, and Tianfan Xue propose Lume-Palette, a progressive framework for spatially controllable multi-view indoor relighting. It decouples the task into illumination distillation (extracting canonical palettes from a pretrained diffusion model) and illumination casting (mapping target lighting from coarse 3D geometry). An asymmetric multi-view conditioning strategy handles dense multi-modal inputs. Experiments show photorealistic results across synthetic and real scenes.
Key Points
- Two-stage decoupling: illumination distillation then illumination casting for realistic relighting
- Asymmetric multi-view conditioning selectively injects spatial context from dense inputs
- Tested on synthetic and real-world scenes with photorealistic, multi-view consistent outputs
Why It Matters
Enables precise, photorealistic lighting control in 3D scenes for VR, film, and architectural visualization.