Google’s GNM: How 636 Parameters Enable a Controllable Moving 3D Face

Google GNM Head – a parameterized 3D head model

Google released the open‑source GNM (Generative aNthropometric Model) Head, which follows the classic 3D Morphable Model approach. The model treats a human head as a standard geometry plus two deformation dictionaries: one for identity (static shape) and one for expression (dynamic facial motions). Pose and global translation are added on top, and the final mesh is obtained by linear blend‑skin weighting.

Model composition and parameter breakdown

The v3.0 model file contains 17,821 vertices, 35,324 triangles and four joints. Its controllable parameters are:

Identity (253 dimensions) : head (170), eyes (3), teeth (80)

Expression (383 dimensions) : left eye (100), right eye (100), lower face (150), tongue (32), pupil (1)

Pose (4 joints) : neck, head, left eye, right eye

Global control (3 dimensions) : X / Y / Z translation

The official README lists 383 expression dimensions, but the PDF definition mistakenly records the tongue as 31 dimensions; the actual model uses 32, so the article follows the v3.0 file.

Why the fine‑grained split matters

Each parameter maps to a deterministic mesh, enabling exact reproducibility, interpolation and real‑time tracking. By modeling eyes, mouth, lower face, tongue and pupil independently, GNM avoids unwanted coupling (e.g., mouth movement triggering a blink) and applies smooth blending at region boundaries.

Internal structures are also detailed: teeth identities derive from 5,000 procedural tooth sets, the tongue combines scanned geometry with artist‑crafted bindings, and the eyeballs use a dual‑sphere approximation for cornea and sclera, leaving a separate pupil‑scale parameter.

Real‑time demonstration

Developer Chris Offner connected MediaPipe facial landmarks to GNM, showing a live mesh that follows blinking, mouth opening and other expressions. Swapping the identity vector instantly changes the face to a different person, proving that identity and expression can be processed separately.

The community quickly built extensions: a Three.js browser demo, Blender integration, a Rust rewrite of the core deformation pipeline, and a desktop tool with 52 blendshapes, animation export and GLB output. All of these rely on a few matrix operations and run on NumPy, JAX, PyTorch or TensorFlow under an Apache 2.0 license, without requiring large GPUs.

Hugging Face Space – an interactive front‑end

A Hugging Face Space wraps the upstream notebook into a Gradio + Three.js UI. Instead of exposing all 636 sliders, it offers 38 controls (10 identity, 16 expression, 9 pose, 3 translation). Users can select semantic expressions such as “happy”, “surprise” or “wink”, which a CVAE sampler first generates as a parameter set, then fine‑tune manually.

The evaluation of a single forward pass takes roughly 40 ms on a free CPU instance, and the Three.js scene updates only the vertex buffer, avoiding full scene reconstruction and keeping the camera stable.

Limitations and open issues

Only the head is released; textures, hair, clothing, lighting and rendering are left to other systems.

The training data uses binary gender labels and four broad population groups, limiting demographic coverage.

Documentation is still incomplete; the formal definition PDF contains a one‑dimensional mismatch.

GNM is not an out‑of‑the‑box facial capture product; developers must still handle landmark detection, parameter fitting, smoothing and export.

Why GNM matters

GNM provides a high‑quality, controllable geometry layer that can be combined with perception models (e.g., MediaPipe) on one side and rendering pipelines (Blender, Three.js, game engines) on the other. It complements generative AI—diffusion models supply rich appearance, while GNM supplies precise, programmable motion.

GNM Head is best viewed as an open‑source building block for avatars, real‑time avatars, gaze‑estimation data generation, or research tools that need a unified topology, rather than a complete digital‑human solution.