Reproduce “FaceVerse: a Fine-grained and Detail-controllable 3D Face Morphable Model from a Hybrid Dataset” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, Yebin Liu
FaceVerse introduces a fine-grained and detail-controllable 3D neural face model, specifically optimized for East Asian faces, constructed using a novel hybrid dataset and a coarse-to-fine conditional StyleGAN architecture. The model achieves high generalization and fidelity, enabling parametric control over intricate facial geometry and texture details.
Key insights
- Combining large-scale low-fidelity data with small-scale high-fidelity data through a coarse-to-fine neural architecture effectively addresses the trade-off between generalization and detail in 3D face modeling.
- A conditional StyleGAN architecture, particularly with a normal discriminator, can learn to generate and control parameter-changeable, fine-grained geometric and textural details for 3D faces.
- Hierarchical fitting, moving from base parameters to detail latent codes and noise, provides a robust and effective strategy for reconstructing high-fidelity, controllable 3D faces from single 2D images.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.