TensoRF-GAN: Improving 3D-Aware Portrait Image Synthesis with Tensor Decomposition

Existing methods for synthesizing portraits with 3D-aware can produce remarkable high-quality portrait images while maintaining strong 3D consistency. However, the localized receptive field of the implicit function in NeRF poses a challenge for the generator to effectively encode the overall global structure of the feature field. Simultaneously, NeRF relies on volume rendering, which can incur high computational costs for generating high-resolution results, thereby intensifying the challenges in the training process. To address these challenges, we introduce TensoRF-GAN, a novel framework for efficient and high-quality 3D-aware portrait synthesis. Specifically, we integrate the high-quality and 3D view-consistent image synthesis method VolumeGAN with the Vector-Matrix (VM) decomposition technique and effectively optimize the 3D feature volume into factorized low-rank tensor components. From the experiments, our method not only demonstrates enhanced efficiency in terms of both training time and reduced GPU memory usage but also improves the image quality for the 3D-aware image synthesis.