Fast 2-step regularization on style optimization for real face morphing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• StyleGAN excels at high-quality face synthesis.
• Inverting real faces into StyleGAN's latent space is an active research area.
• Morphing real faces using latent representations is currently limited by high training costs and large labeled datasets.

Purpose of the Study:

• To propose a novel method for morphing real faces based on StyleGAN.
• To address the limitations of existing methods regarding training costs and data requirements.
• To enable faster, more diverse, and realistic face morphing.

Main Methods:

• Proposing a novel method by adding regularization to style optimization.
• Labeling latent vectors via synthesized faces to create a labeled dataset.
• Utilizing logistic regression to discover interpretable directions in latent space.
• Optimizing latent vectors and directions using appropriate regularization.
• Employing learned directions across different layer representations for real face morphing.

Main Results:

• The proposed method achieves faster and more diverse realistic output compared to existing techniques.
• Reduced training costs and data requirements for StyleGAN-based face morphing.
• Successful demonstration of real face morphing using learned latent directions.

Conclusions:

• The novel regularization-based approach significantly advances StyleGAN-based real face morphing.
• This method offers a more efficient and effective solution for generating diverse and realistic morphed faces.
• The released code and dataset facilitate further research in this domain.