Image Relighting

Image relighting studies how to change the direction, color, intensity, or spatial placement of light in images and scenes without altering their underlying content. Approaches range from diffusion and transformer models conditioned on lighting controls to intrinsic-image decomposition and physics-guided methods that estimate geometry, materials, shadows, and illumination. For 3D and multi-view settings, relightable neural representations such as Gaussian splats aim to keep lighting edits consistent across viewpoints, while video methods also seek temporal stability. A central challenge is closing the gap between synthetic training data and complex real-world lighting while balancing physical accuracy, fine-grained control, and convincing image generation.