CVPR2024
StyLitGAN: Image-Based Relighting via Latent Control
Anand Bhattad, James Soole, David A. Forsyth
Abstract
Generated Image Relit -1 (w + + d 1 ) Relit -2 (w + + d 2 ) Relit -3 (w + + d 3 ) Relit -4 (w + + d 4 ) Relit -5 (w + + d 5 ) Figure 1 . StyLitGAN identifies directional vectors (di) within StyleGAN's style space (W + ) which, when added to the w + style code, effectively modify the lighting of generated images while preserving their geometry and albedo. This process eliminates the need for per-image search or model fine-tuning. The first column displays images generated from StyleGAN2; subsequent columns illustrate the same scene, each relit using a specific direction. These relighting directions (di) are derived through a forward selection method, ensuring diversity and avoiding cherry-picking. The directional effects are consistent across different scenes: for instance, d1 activates an orange-tinged bedside lamp, d2 a less intense white-tinged lamp, d3 introduces strong directional light from the window, and so on, demonstrating diverse relighting capabilities of StyLitGAN.