LFSD-VQVAE: 3D Face Generation Based on Local Feature Swapping

Currently, the study of latent disentanglement in 3D shape generation still fails to achieve accurate control of local features in the generated results. To tackle this challenge, we propose a 3D face generation method based on local feature swapping, referred to as LFSD-VQVAE. We segment a 3D face using a pre-defined face template and map the 3D face onto a canonical sphere, serialize the points in each region of the canonical sphere, and then perform group learning to obtain a set of shape combinations with semantic information. We perform shape combination swapping among different 3D faces and use shared codebook strategy to quantize the feature from the encoder, followed by learning a transformer that implements the 3D face generation task. Our local feature swapping aims to smooth the transition between adjacency attributes of the face and safeguard the continuity and naturalness of the face surface, effectively overcoming the limitations of existing face generation models in controlling local feature disentanglement. We conduct experiments on UHM datasets containing neutral expression. Compared with several state-of-the-art methods, the proposed method achieves the highest VP value, demonstrating its superior effectiveness in controlling local feature disentangling.