Towards Disentangling Latent Space for Unsupervised Semantic Face Editing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Facial attribute control in StyleGAN is challenging due to entangled latent spaces.
• Supervised methods require difficult-to-obtain labeled data and limit attribute editing.
• Unsupervised, disentangled latent space editing is crucial for versatile semantic face manipulation.

Purpose of the Study:

• To develop a novel technique for disentangling latent spaces in generative adversarial networks (GANs).
• To enable unsupervised, independent control over facial attributes in generated images.
• To improve the quality and versatility of semantic face editing.

Main Methods:

• Introduced Structure-Texture Independent Architecture with Weight Decomposition and Orthogonal Regularization (STIA-WO).
• Developed STGAN-WO by applying STIA-WO to GANs, using style vectors for weight matrix control and orthogonal regularization.
• Implemented a structure-texture independent architecture with two latent vectors for disentangled control of image components.

Main Results:

• STGAN-WO successfully disentangles facial attributes in the latent space.
• Unsupervised semantic editing is achieved by manipulating latent codes in coarse and fine layers for texture and structure, respectively.
• Experimental results demonstrate superior attribute editing performance compared to state-of-the-art methods.

Conclusions:

• The proposed STIA-WO technique effectively disentangles latent spaces for unsupervised semantic face editing.
• STGAN-WO offers a powerful tool for precise and independent control of facial attributes in GAN-generated images.
• This approach overcomes limitations of supervised methods and enhances the capabilities of generative models for image editing.