StyleMorpheus: Learning a StyleGAN-Based 3D-Aware Morphable Face Model with a Disentangled Style Space

I. Overview

StyleMorpheus learns a neural 3D morphable face model from unconstrained images instead of requiring a large collection of accurately reconstructed 3D scans.

Its style-based latent design separates identity, expression, and appearance controls while retaining photorealistic, 3D-aware rendering.

II. Key Contributions

• Learns a style-based neural 3D morphable model from in-the-wild face images.
• Separates shape- and appearance-related controls across the model to improve disentanglement.
• Supports real-time rendering and downstream editing operations such as style mixing and color manipulation.

III. Methodology

An autoencoder maps face images into a disentangled parametric code space. Shape- and appearance-related style codes control different decoder modules, and style-based adversarial fine-tuning improves photorealistic 3D-aware rendering.

IV. Main Findings

The model is evaluated on face reconstruction and novel-view synthesis and demonstrates controllable identity, expression, and appearance editing at real-time rendering speed.

@Article{Yan_2025_StyleMorpheus,
  author  = {Yan, Peizhi and Ward, Rabab K. and Wang, Dan and Tang, Qiang and Du, Shan},
  title   = {StyleMorpheus: Learning a StyleGAN-Based 3D-Aware Morphable Face Model with a Disentangled Style Space},
  journal = {Neurocomputing},
  year    = {2025},
  volume  = {654},
  pages   = {131329},
  doi     = {10.1016/j.neucom.2025.131329}
}