Multi-Source Face Relighting: Pre-fusion and Post-fusion Approaches for Improved Lighting Consistency and Detail Preservation

Multi-source face relighting is a key challenge in computer vision due to its applications in virtual reality, face recognition, and image editing, yet existing methods struggle with lighting consistency, shadow modeling, and detail preservation under complex multi-light-source conditions. This paper introduces two approaches, pre-fusion and post-fusion, to address these challenges by extending single-light-source relighting methods to multi-source scenarios. The pre-fusion approach unifies multiple light source parameters before relighting, improving lighting consistency and shadow modeling, while the post-fusion approach fuses individual relit images at the pixel level, preserving fine details with flexible control. Evaluations on a custom dataset and the FFHQ dataset show that pre-fusion outperforms post-fusion in terms of MSE, PSNR, and DSSIM. Experimental results confirm that pre-fusion effectively enhances lighting consistency and preserves details, offering a practical solution for real-world multi-source face relighting.