PStyle-3D: Example-Based 3-D-Aware Portrait Style Domain Adaptation.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Generating high-quality artistic portraits is a key challenge in computer vision.
• Existing 3-D generative models struggle with preserving facial details in artistic portraits due to 3-D structure complexities.
• 2-D generative models are often preferred for artistic portraits despite limitations in 3-D consistency.

Purpose of the Study:

• To develop a novel approach for generating high-quality, multiview-consistent 3-D artistic portraits.
• To address the challenges of integrating artistic style with 3-D facial geometry.
• To improve upon existing 3-D generative models for portrait synthesis.

Main Methods:

• A novel domain feature extraction module to capture features from natural faces and artistic portraits.
• Integration of extracted features into a 3-D representation for multiview consistency.
• A new domain adaptation module, inspired by StyleGAN, to regulate target domain features in S space within each StyleBlock.

Main Results:

• The proposed approach successfully generates high-quality 3-D artistic portraits.
• The method demonstrates superior performance in preserving 3-D geometric information compared to existing techniques.
• Enhanced multiview consistency was achieved in the generated 3-D artistic portraits.

Conclusions:

• The novel domain adaptation approach effectively fuses source and target domain features for 3-D artistic portrait generation.
• The method overcomes limitations of previous 3-D generative models in artistic portrait synthesis.
• This work advances the state-of-the-art in creating realistic and consistent 3-D artistic facial representations.