GNM Head: A Generative aNthropometric Model of the human head

GNM (Generative aNthropometric Model) is a parametric 3D head model designed to overcome the limitations of traditional models like FLAME or the Basel Face Model. While existing models typically treat the head as a hollow shell, GNM provides a holistic representation that includes the face, neck, eyeballs, teeth, and tongue. By unifying these structures into a single statistical manifold, GNM allows for precise, semantically disentangled control over facial expressions and internal mouth movements, which is essential for high-fidelity animation and photorealistic rendering.

The model is built on a massive dataset of over 5,000 individuals, captured using a multi-view system that provides high-resolution 3D geometry. To ensure anatomical accuracy, the authors combined these real-world scans with high-quality, artist-sculpted assets for internal structures. The GNM framework uses a linear identity and expression basis, combined with a skeletal rig for articulation. A key innovation is the use of localized part-based smoothness constraints, which allow the model to deform specific regions (like the tongue or lips) without causing unwanted distortions in adjacent areas like the skull or ears.

As generative AI and neural rendering techniques (such as 3D Gaussian Splatting) become more prevalent, there is an increasing need for robust geometric priors that can ground synthetic imagery in physical reality. GNM addresses the "Uncanny Valley" by providing the necessary geometric constraints for realistic speech, co-articulation, and eye movement. By making the framework publicly available, the authors aim to provide a standard, high-fidelity foundation for researchers and developers working in telepresence, AR, and digital human animation.