Enhancing the Latent Space of 3D-GANs: A Model-Agnostic Paradigm for Hierarchical Disentanglement

Recent advancements in 3D Generative Adversarial Networks (3D-GANs) have made significant strides in generating high-fidelity, multi-view-consistent facial images from single-view 2D image databases. These cutting-edge networks integrate the capabilities of StyleGAN and NeRF, offering differentiated control over pose and attributes through distinct subspaces of the latent space. Despite these advancements, a significant challenge remains: the overlapping of these subspaces, which complicates distinct pose and attribute manipulation. Furthermore, previous studies have largely overlooked the crucial task of segregating identity-centric and identity-free information. To address the above issues, our study introduces a model-agnostic, hierarchical disentanglement framework (in terms of underlying 3D architecture), designed to refine the latent space of 3D-GANs. We achieve this by employing an equivariant network to minimize the pose-attribute overlap and a reversible network to delineate identity from identity-free features. Notably, our research marks the first implementation of pose-invariant feature extractors to enable pose equivariance, alongside the innovative use of reversible networks in latent space disentanglement. The proposed latent space, EDSpace, has demonstrated marked improvements over traditional latent spaces in 3D-GANs. It consistently enhances quality metrics across various datasets, methods, and models, underscoring the framework's robustness and effectiveness in advancing the capabilities of 3D-GANs.