Disentangled 3D Gaussian Splatting

Recent advances in novel view synthesis have been greatly accelerated by the efficient, photorealistic rendering capabilities of 3D Gaussian Splatting (3DGS), bringing volumetric video technology closer to practical adoption. However, to fully realize the capabilities of 3DGS-based volumetric video, enhancing its geometric fidelity and relighting performance remains a critical challenge. In this work, we demonstrate that a controlled disentanglement of geometry and appearance modeling in 3DGS mutually enhances both aspects, leading to optimized convergence. Motivated by this observation, we propose a novel disentangled 3DGS representation that introduces a disentangling factor to regulate the interaction between geometry and appearance. Combined with well-designed pruning and densification strategies, our method improves performance in both novel view synthesis and surface reconstruction. We incorporate normal and visibility as additional Gaussian attributes, optimized in a staged manner, enabling high-quality relighting through a deferred shading pipeline. Furthermore, we extend our framework to the temporal domain, incorporating customized optimization strategies to produce temporally stable volumetric videos with relighting capabilities. Extensive experiments on both custom and public datasets demonstrate that our approach achieves competitive results in rendering quality and geometric reconstruction compared to existing methods.