3DFaceController: Region-Controllable Face Synthesis via Decomposed and Recomposed Neural Radiance Fields

Kangneng Zhou, Yaxing Wang, Shuang Song, Jie Zhang, Ping Li

Abstract: Advancements in neural radiance fields (NeRFs) have enhanced 3D face synthesis quality. While some methods use semantic maps to guide synthesis, region-controllable synthesis remains challenging. We propose 3DFaceController, a framework for decompositional and recompositional generative radiance fields enabling region-controllable face synthesis. 3DFaceController decomposes the global face field into local fields via signed distance functions (SDF), allowing independent rendering of local components and explicit generation of physically valid 3D structures. A style-based generator with a Spatial-Semantic-Recomposition (SSR) module then synthesizes high-resolution images by combining global and local features without additional optimization. Experiments show 3DFaceController achieves state-of-the-art performance in photorealism and disentanglement.