Concepts de base
Proposing a method for highly stylized 3D face models with desired topology using surface deformation network and paired exemplar.
Résumé
The article introduces a novel method, LeGO, for generating highly stylized 3D face models with desired topologies. It focuses on training a surface deformation network with 3DMM and translating its domain to the target style using a paired exemplar. The method achieves stylization of 3D face mesh by mimicking the style of the target through differentiable renderer and directional CLIP losses. The Mesh Agnostic Encoder (MAGE) is utilized during inference to enable mesh-agnostic stylization for input with diverse topologies. The results demonstrate successful production of highly stylized face meshes according to given styles and desired topologies, suitable for applications like image-based stylized avatar generation and facial animation.
Introduction:
Recent advances in 3D face stylization have made significant strides in few to zero-shot settings.
Existing methods based on statistical 3D Morphable Models (3DMM) lack sufficient variation for practical applications.
Key Elements for Stylized Avatar Creation:
Avatar creation in a desired topology compatible with conventional CG pipelines.
Extending stylization capabilities beyond 3DMM.
Generating animatable stylized avatars using blend shapes.
Proposed Method:
Leverages surface deformation network trained with FLAME decoder to achieve highly stylized 3D face models.
Utilizes directional CLIP-based domain adaptation method for identity preservation while reflecting desired style.
Hierarchical rendering scheme captures local and global facial features effectively during training and inference stages.
Related Work:
Various methods like GANs, DMs, text-based geometry deformation, and surface deformation have been explored for generating high-quality and diverse stylized 3D faces.
Stats
"During fine-tuning, we employ a directional CLIP-based domain adaptation method."
"We also demonstrate example applications of our method including image-based stylized avatar generation."
"MAGE is composed of pre-trained encoders from Neural Face Rigging (NFR) and latent mapping networks."
"The resulting stylized face model can be animated by commonly used 3DMM blend shapes."