Physically Controllable Relighting of Photographs
Summary
Summarizes a self-supervised, physically-based relighting method that enables explicit control over lighting in photographs by inferring a colored 3D mesh from monocular cues, using path-traced rendering, and a neural renderer for final results. The approach merges traditional rendering's physical accuracy with neural rendering's realism and trains with in-the-wild image collections. This work pushes CG-style lighting control into real-world photo editing.