Controllable and Identity-Aware Facial Attribute Transformation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial attribute modification without paired datasets is challenging.
• Previous methods require ground-truth images or numerous models, limiting scalability and attribute range.
• Existing approaches often fail to preserve facial identity during attribute transformations.

Purpose of the Study:

• To develop a single model for controllable and identity-aware facial attribute transformations.
• To overcome the scalability and identity preservation issues of prior methods.
• To enable efficient, multi-attribute facial image manipulation.

Main Methods:

• Utilized a generative adversarial network (GAN) with a multitask conditional discriminator.
• The discriminator was trained to recognize identity, distinguish real/fake images, and identify facial attributes.
• Learned meaningful image representations in a lower-dimensional latent space, associating vector parts with identity and attributes.

Main Results:

• The proposed method achieves controllable and identity-aware transformations across multiple facial attributes using a single model.
• The model successfully preserves facial identity and attributes during transformations.
• Learned latent space representations allow for generating new faces and diverse transformations (e.g., thinner/chubbier).
• Achieved faster transformations due to single image encoding and efficient vector manipulation.

Conclusions:

• The proposed GAN framework offers a scalable and extensible solution for facial attribute editing.
• It significantly outperforms existing methods like CycleGAN in terms of efficiency and flexibility.
• The approach enables diverse facial image manipulations while preserving identity, with potential for generating novel faces.