Improved latent diffusion-based IC-DGAN framework for high-resolution multi-feature and expression manipulation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Generative Models

Background:

• Facial expression and multi-feature manipulation are crucial for media and biometrics.
• Existing methods struggle with semantic consistency, pose/illumination sensitivity, and computational cost.

Purpose of the Study:

• To propose an improved latent diffusion-based deep generative adversarial network (IC-DGAN) for precise semantic multi-feature and facial expression manipulation.
• To address limitations of current facial editing techniques, including semantic inconsistency and high computational demands.

Main Methods:

• Developed the IC-DGAN framework integrating multiple generators/discriminators, K-means clustering, and constructive pre-training.
• Utilized Scale-Invariant Feature Transform (SIFT) and latent diffusion models for attribute disentanglement and manipulation.
• Enabled robust attribute editing (age, gender, expression) by mapping portraits to the latent space, minimizing visual distortions.

Main Results:

• IC-DGAN achieved reduced unintended portrait variations by 12.3% and enhanced manipulation accuracy by 8.7%.
• The framework obtained a Fréchet Inception Distance (FID) of 25.94, outperforming state-of-the-art methods.
• Demonstrated high-resolution, realistic portrait generation with synchronized multi-level decomposition.

Conclusions:

• IC-DGAN offers a robust solution for high-fidelity facial editing, overcoming significant challenges in the field.
• The proposed framework advances precise semantic multi-feature and facial expression manipulation capabilities.
• Results highlight the potential for improved applications in media entertainment and biometric forensics.