google/gnm-v3

GNM (Generative aNthropometric Model) - v3.0

Official release weights for Google's Generative aNthropometric Model (GNM).

---

Model Overview

GNM (pronounced *genome* /ˈdʒiː.noʊm/, in reference to the human genome) is a state-of-the-art parametric 3D statistical model of the human head developed by Google. Learned from a large dataset of high-resolution 3D scans, GNM provides fine-grained, disentangled control over facial identity, expressions, and head pose, complete with controllable internal anatomy (eyeballs, teeth, and tongue).

3D Morphable Models (3DMMs) are widely used across computer vision, computer graphics, and generative AI for representing human geometry and appearance. GNM introduces a state-of-the-art parametric representation of the human head accompanied by multi-framework backend support and semantic parameter sampling.

The model is released under the Apache 2.0 permissive license, suitable for both academic research and commercial applications.

---

Key Features & Anatomy

• Detailed 3D Face Geometry: Generates a dense 3D mesh consisting of the

cranial/facial skin, eyes, teeth & gums, and tongue.

• Controllable Anatomy:
• Skin: Full cranial, facial, and neck coverage.
• Eyes: Controllable internal eyeball structures (sclera, pupil, iris)

and external cornea.

• Teeth & Gums: Articulated upper and lower dentition.
• Tongue: Articulated tongue surface supporting inner-mouth deformation

and speech articulations.

• Disentangled Parameter Spaces: Full orthogonal control over:
• Identity: Controls subject-specific facial features and anatomical

proportions (head, eyeball, teeth).

• Expression: Animates the face with a rich set of blendshapes across the

eyes, lower face, tongue, and iris.

• Head Pose & Joint Rotations: Controls kinematic rotation of the neck

and bilateral eyeball gaze (axis-angle format).

• Translation: Controls global 3D Cartesian positioning.
• Multi-Topology UV Mapping: Structured UV layouts provided for both quad

(quad_uvs, shape [Q, 4, 2]) and triangulated (triangle_uvs, shape [T, 3, 2]) topologies across five logical regions (skin, upper teeth/gums, lower teeth/gums, tongue, eye interior, and eye exterior).

• Multi-Framework Backend Support: Native support for NumPy, JAX,

PyTorch, and TensorFlow via a unified GNM interface.

• Semantic Parameter Sampling: Compatible with pre-trained IdentitySampler

and ExpressionSampler models for generating identity and expression parameters from high-level labels.

---

Model Organization Structure

Both Hugging Face Hub and Kaggle Models follow an identical organization architecture:

• 1 Model / Repository per MAJOR Version:
• Hugging Face Hub: google/gnm-v3
• Kaggle Models: google/gnm-v3
• Variants for MINOR Versions:
• Hugging Face Hub: Each minor version is organized into version

subdirectories (e.g. v3_0/gnm_head.npz), with the root README.md documenting the major release and version-specific README.md files documenting each version directory.

• Kaggle Models: Each minor version is published as a distinct model

variation under the other framework (e.g. google/gnm-v3/other/gnm_head_v3_0).

---

Model Files in This Release

The following model files and SHA-256 checksums are staged in this release:

| Model File | Variant | Version | SHA-256 Checksum | | :--- | :--- | :---: | :--- | | gnm_head.npz | Head | 3.0 | 61d78bbfb4ad8e0b38495804a4caef3214d3df00f8c3f68761e63b41ce3747eb |

---

Quickstart & Usage

Install the gnm package from the GNM GitHub repository:

git clone https://github.com/google/gnm.git
cd gnm/gnm/shape
pip install -e .

# Or with framework-specific backend support:
pip install -e ".[jax]" # JAX
pip install -e ".[pytorch]" # PyTorch
pip install -e ".[all]" # All supported frameworks

Loading the Model

from gnm.shape import gnm_numpy # Or gnm_jax, gnm_pytorch, gnm_tensorflow

# 1. Load from Hugging Face Hub (via huggingface_hub SDK or HTTPS CDN fallback)
gnm = gnm_numpy.GNM.from_huggingface(
version=gnm_numpy.GNMMajorVersion.V3,
variant=gnm_numpy.GNMVariant.HEAD,
)

# 2. Load from Kaggle Models (via kagglehub SDK)
gnm = gnm_numpy.GNM.from_kaggle(
version=gnm_numpy.GNMMajorVersion.V3,
variant=gnm_numpy.GNMVariant.HEAD,
)

# 3. Load via HTTPS CDN (no external SDK required, automatic local caching)
gnm = gnm_numpy.GNM.from_remote(
version=gnm_numpy.GNMMajorVersion.V3,
variant=gnm_numpy.GNMVariant.HEAD,
)

Basic Parameter Manipulation

import numpy as np

# Zero parameters produce the neutral template face mesh
identity = np.zeros(gnm.identity_dim)
expression = np.zeros(gnm.expression_dim)
rotations = np.zeros((gnm.num_joints, 3)) # Axis-angle rotations (neck, eyes)
translation = np.zeros((3,))

# Evaluate 3D vertices [V, 3]
vertices = gnm(identity, expression, rotations, translation)
triangles = gnm.triangles # [F, 3]

---

Model Limitations & Ethical Considerations

This model was trained on datasets using demographic categories standard in 3DMM literature. These categories do not fully represent the spectrum of human gender identities or the full diversity of the global population. Please refer to the Technical Report for a detailed discussion of dataset statistics, fairness evaluations, and limitations. Users should consider the potential implications for fairness and representation in their specific applications.

---

Citation

If you use GNM or this model release in your research, please cite:

@article{ploumpis2026gnmhead,
title={GNM Head: A Generative aNthropometric Model of the human head},
author={Ploumpis, S. and Bednarik, J. and Zoss, G. and Guseinov, R. and Prasso, L. and Chandran, P. and Boyne, O. and Choutas, V. and Bolkart, T. and Wang, D. and Chai, M. and Qiu, D. and Winberg, S. and Rainer, G. and Bridgeman, L. and Vicini, D. and Riviere, J. and Boetzel, Y. and Koumis, A. and Busch, J. and Herrera, C. and Still, J. and Ysebert, S. and Lincoln, P. and Escolano, S. O. and Rhemann, C. and Wood, E. and Beeler, T. and Zafeiriou, S.},
year={2026},
eprint={2607.23687},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2607.23687},
}

---

Links & Resources

• GitHub Repository:...