3D Portrait Stylization with Adaptive Semantic Editing based on GAN Latent Codes

Yantao Song, Xiangchong Jia, Jieru Jia, Yudong Liang, Xinyan Liang

Abstract: Stylized 3D portrait generation faces two key challenges: preserving facial identity during artistic transformations and enabling precise semantic control over edited outputs. While existing methods based on 3D GANs can adapt to new styles through fine-tuning, they often suffer from geometric distortions and excessive texture oversmoothing, particularly in cartoon styles requiring exaggerated features. To address these limitations, we propose a novel 3D portrait stylization framework that decomposes the task into three independent subproblems: controlled generation, geometric stylization, and texture stylization. By leveraging GAN latent codes as conditional inputs, our approach simultaneously achieves multi-style generation and attribute manipulation without compromising geometric or textural fidelity. Extensive experiments on paired multi-style face datasets, constructed from benchmark methods, demonstrate the consistent superiority of our method consistently outperforms state-of-the-art methods in terms of visual quality and editing accuracy.

External IDs:dblp:journals/pr/SongJJLL26