SGFNeRF: Shape Guided 3D Face Generation in Neural Radiance Fields

NeRF (Neural Radiance Fields), as an implicit 3D representation, has demonstrated the capability to generate highly realistic and dynamically consistent images. However, its hierarchical sampling approach introduces a significant amount of redundant computation, leading to erroneous geometric information, particularly in high-frequency facial details. In this paper, we propose SGFNeRF, a novel 3D face generation model by integrating a 2D CNN-based generator and face depth priors optimization method in the same framework. We employ a Gaussian distribution for sampling to extract facial surface information. Additionally, we design a feature decoder to incorporates depth uncertainty into our method, enabling the method to explore regions further away from face surfaces while preserving its ability to capture fine-grained details. We conduct experiments on the FFHQ dataset to evaluate the performance of our proposed method. The results demonstrate a significant improvement compared to previous approaches in terms of various evaluation metrics.