Core Concepts
HeadEvolver enables the generation of stylized head avatars from text guidance through locally learnable mesh deformation.
Abstract
HeadEvolver introduces a novel framework for creating high-quality digital head avatars using text guidance. The method focuses on preserving facial features and semantic consistency while deforming a template mesh. By incorporating per-triangle weighted Jacobians, the approach allows for fine-grained control over local shape changes guided by text prompts. The framework supports various applications such as motion retargeting, texture transfer, geometry editing, and texture editing. Extensive experiments demonstrate the effectiveness of HeadEvolver in generating diverse and stylized 3D head avatars suitable for interactive editing and animation.
Stats
Generated results include "Vincent Van Gogh" and "Young Elf Girl"
Texture editing involves changing hair colors to red
Self-intersection ratio is 2.08%
Quotes
"HeadEvolver can deform a template mesh to stylized head avatars under text guidance."
"Our method is capable of generating diverse and stylized 3D head avatars as high-quality digital assets."