Abstract
3D head stylization transforms realistic facial features into artisticrepresentations, enhancing user engagement across gaming and virtual realityapplications. While 3D-aware generators have made significant advancements,many 3D stylization methods primarily provide near-frontal views and struggleto preserve the unique identities of original subjects, often resulting inoutputs that lack diversity and individuality. This paper addresses thesechallenges by leveraging the PanoHead model, synthesizing images from acomprehensive 360-degree perspective. We propose a novel framework that employsnegative log-likelihood distillation (LD) to enhance identity preservation andimprove stylization quality. By integrating multi-view grid score and mirrorgradients within the 3D GAN architecture and introducing a score rank weighingtechnique, our approach achieves substantial qualitative and quantitativeimprovements. Our findings not only advance the state of 3D head stylizationbut also provide valuable insights into effective distillation processesbetween diffusion models and GANs, focusing on the critical issue of identitypreservation. Please visit the https://three-bee.github.io/head_stylization formore visuals.