NeRF-inspired 3D face portraits generation

• Main Author: Min Min
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/182442
• Institution: Nanyang Technological University

3D face reconstruction has always been a hot topic in the field of computer vision. Many researchers have used 3D deformable models for 3D face reconstruction, but most of these methods focus on using a single image. Due to the lack of multi-view information in a single image, the reconstruction quality is often suboptimal. This project explores and studies 3D face reconstruction in the following ways: First, to better experiment and validate the methods, a new facial dataset was independently constructed, and 3D reconstruction was performed on multiple faces. Second, methods such as NeRF(Neural Radiance Fields), which are primarily applied to scene reconstruction, have not been widely explored in the context of face reconstruction. This project innovatively investigates the performance of TensoRF(Tensorial Radiance Fields) in the field of face reconstruction. Based on the self-built dataset, the project tests face reconstruction under various scenarios and explores the impact of lighting conditions and local highlights on TensoRF’s performance. At the same time, 3D Gaussian Splatting and HRN(Hierarchical Representation Network) methods were reproduced, and their results were compared with TensoRF’s performance. A detailed comparison was made, discussing and summarizing the advantages of TensoRF in reconstructing high-frequency facial details. Finally, based on TensoRF’s performance, this project creatively employed a simple UNet architecture for further image enhancement of high-frequency skin details, achieving promising results.