GNM Head: A Generative aNthropometric Model of the human head
Stylianos Ploumpis, Jan Bednarik, Gaspard Zoss, Ruslan Guseinov, Luca Prasso, Prashanth Chandran, Oliver Boyne, Vasileios Choutas, Timo Bolkart, Daoye Wang, Menglei Chai, Di Qiu, Sebastian Winberg, Gilles Rainer, Lewis Bridgeman, Delio Vicini, Jérémy Riviere, Yannick Boetzel, Alexander Koumis, Jay Busch, Cynthia Herrera, Jacob Still, Scott Ysebert, Peter Lincoln, Sergio Orts Escolano, Christoph Rhemann, Erroll Wood, Thabo Beeler, Stefanos Zafeiriou
cs.CV, cs.GR
2026-07-26
Google open-sources GNM (Generative aNthropometric Model), unifying face, eyeballs, teeth, tongue and neck in one 3D Morphable Model statistical space, fixing FLAME's shell-only gap that omits oral and ocular structure; on 2,000 synthetic-image single-view reconstructions its mean error is 1.683 mm vs FLAME's 2.172 mm.
3D Morphable Models (3DMMs) are an old CV/graphics tool: compress the 3D shape and appearance of the human head into a low-dimensional, controllable latent space. They have become important again because generative large vision models (diffusion) need them as geometric priors for tight spatial control, and neural rendering (3D Gaussian Splatting, NeRF) leans on them to anchor physically plausible surfaces.
But available open head models (FLAME, BFM, LSFM) share a fundamental limitation: they treat the head as a hollow shell, largely omitting intra-oral (teeth, tongue) and fine ocular structure. Without those geometric constraints, generative models lose quality on open-mouth shots and lose fine control over non-verbal cues like lip-tongue coarticulation. This Google report fills that gap.
GNM (a homophone of "genome") is a holistic parametric head framework that unifies facial skin, eyes, teeth, tongue and neck in one statistical space. Three key design choices:
The model is built on high-resolution mesh topology, trained on large-scale high-fidelity 3D scans plus artist-made specialized assets. To turn statistical coefficients into intuitive sliders, the authors add a Semantic Sampler (a dual-CVAE architecture) mapping demographic and expression attributes onto a smooth manifold without unnatural distortion. The accompanying fitting pipeline has collision constraints, tongue convex-hull tests and regularizers, and reconstructs GNM meshes from single- or multi-view images using dense 2D facial landmarks.
Compared with FLAME under an identical pipeline on 2,000 synthetic images with 7,000 ground-truth dense 2D landmarks (synthetic data removes 2D landmark-detection noise):
| Method | Mean error (mm) | Median (mm) |
| FLAME | 2.172 | 2.086 |
| GNM | 1.683 | 1.589 |
GNM's mean error is about 22% lower. The qualitative side matters more: on in-the-wild single-view reconstruction, GNM recovers extreme intra-oral deformations such as wide jaw openings and tongue poses, down to individual teeth, which FLAME cannot do without an internal oral model. Multi-view video and extreme-expression fitting are stable too.
For teams building digital humans, facial mocap, neural rendering, or using 3DMMs to condition diffusion models, GNM is a more complete and higher-quality open prior than FLAME (open for academic and commercial use at github.com/google/GNM). Teeth, tongue and pupils, previously assembled as separate assets, now live in one differentiable, controllable model, which directly helps talking-head synthesis, lip sync, and privacy-safe synthetic face-data generation.
The quantitative error comparison is on synthetic images only. The rationale (remove landmark-detection noise) is sound, but synthetic-image distribution differs from real faces, so the margin on real images may differ. The paper compares GNM against FLAME but does not benchmark newer neural parametric head models (NPHM, imHead) under a unified protocol, and those have their own strengths in topology change and high-frequency detail. Many weights in the Semantic Sampler and fitting pipeline are empirically tuned and would need retuning for new scenarios.