Continuous Shape-to-Texture Face Aging With Flow-Based Prior Latent Age Modulation and Attentional Alignment StyleGAN

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Synthesis

Background:

• Generative models, particularly GANs, struggle with accurately synthesizing continuous facial aging, failing to capture age-related shape-to-texture changes.
• Existing methods lack precise control over the aging process, resulting in unnatural transitions.

Purpose of the Study:

• To develop an innovative facial age transformation framework for generating continuous shape-to-texture aging facial images.
• To enable precise and reversible mapping between age attributes and latent space for smooth aging transitions.
• To accurately depict facial characteristics from shape to texture corresponding to specific ages.

Main Methods:

• Prior Latent Age Modulation (PLAM) utilizes normalizing flows for precise, reversible mapping between age attributes and the prior latent space.
• Attentional Feature Fusion (AFF) dynamically weights and fuses age attribute features with StyleGAN content features.
• The framework generates continuous shape-to-texture aging facial images by manipulating latent space.

Main Results:

• The proposed framework successfully generates realistic facial aging images with continuous progression.
• Quantitative and qualitative analyses demonstrate the method's effectiveness in capturing age-related shape and texture changes.
• The approach achieves superior performance in facial aging tasks compared to existing methods.

Conclusions:

• The developed facial age transformation framework effectively addresses limitations in current generative models.
• The integration of PLAM and AFF enables accurate and smooth synthesis of aging facial images.
• This method offers a significant advancement in realistic facial image synthesis for age progression.