Method for Creating Animatable Avatars Using Neural Radiance Fields and Two-Dimensional Neural Rendering

This paper presents an approach to creating animatable avatars. The method combines two technologies - neural radiance fields conditioned on a priori information obtained from the parametric FLAME model for the implicit representation of the human head model in three-dimensional space and 2D neural rendering for synthesizing new viewpoints based on the trained representation. An architecture for the neural renderer is proposed, in which both local and global context is effectively processed, demonstrating an improvement in quantitative metrics and the rendering quality compared to previous approaches. Unlike existing approaches that require the presence of 3D scans or a sequence of a large number of images of a person from different viewpoints, the proposed approach uses a synthetic dataset generated based on the parametric FLAME model. The developed approach enables the synthesis of human appearances from new viewpoints and the modification of facial expressions.