What AI says about Face Synthetics

The project shows that face analysis models can be trained entirely on synthetic data and still generalize to real-world images, addressing the domain gap between synthetic and real data. It creates realistic, diverse synthetic faces by procedurally generating 3D identities, expressions, textures, hair, and clothing, then rendering with a physically-based engine, and provides a dataset of 100,000 synthetic faces with 2D landmarks and per-pixel segmentation labels. The dataset is available for non-commercial research, enabling tasks like dense landmark detection and face parsing with synthetic data and achieving competitive performance on real data.

Hosted on GitHub Pages