ControlFace: Feature Disentangling for Controllable Face Swapping.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Current face swapping methods often fail to preserve identity similarity due to ineffective feature extraction.
• Existing approaches struggle with accurately disentangling identity information for controllable manipulation.

Purpose of the Study:

• To propose a novel face swapping framework, ControlFace, that enhances identity preservation and enables controllable attribute transfer.
• To disentangle source face structure and texture for more precise identity feature manipulation.

Main Methods:

• ControlFace disentangles source face structure and texture into separate feature embeddings.
• These embeddings are injected into corresponding feature mappers and fused within the StyleGAN latent space.
• This approach allows for controlled transfer of specific identity features.

Main Results:

• The proposed ControlFace framework demonstrates superior performance in identity feature transfer compared to state-of-the-art methods.
• Experiments show high-quality face image generation with improved identity similarity.
• The method enables controllable face swapping, offering users greater flexibility.

Conclusions:

• ControlFace effectively disentangles identity information, leading to enhanced face swapping capabilities.
• The framework achieves superior identity transfer and high-fidelity image generation.
• This work advances controllable face swapping through precise identity feature manipulation.