SV-GaSRelight: Single-View Gaussian Splatting for 3D Human Relighting

Creating realistic lighting in 3D virtual environments is key to making scenes feel immersive and lifelike. However, many current techniques depend on having multiple camera views or using complex neural rendering, which isn’t ideal for real-time or interactive settings. In this project, we introduce SV-GaSRelight, a new method that brings realistic lighting effects to 3D scenes using only a single image. Our main idea is to use people in the scene as natural "light detectors." Since modern AI models can accurately rebuild a 3D human body from just one photo, we take advantage of that to understand how light behaves in the scene. Once we have the 3D human model, we generate synthetic camera views and simulate how light interacts with surfaces using physics-inspired techniques. To capture lighting more effectively, we also use a neural-based light field representation. We tested our method against other techniques that need multiple camera views and found that our approach delivers similar visual results. We also ran a user study to check how realistic the relit scenes looked, and the feedback confirmed our method works well — all from just a single input image.