GNM Head: Google Open-Sources Complete 3D Parametric Human Head Model, Apache 2.0

Google Research has released GNM Head, a 3D parametric human head model with 383 expression blend shapes, 170 shape components for head identity, and 80 shape components for teeth, all in a single unified parameter space. The model covers the full scope of human head anatomy, including controllable eyeballs, teeth, and tongue, areas that prior open parametric models handled separately or excluded entirely. It is released under Apache 2.0, replacing research only restrictions that blocked commercial use of earlier foundational head models.

The License Gap in Prior Head Models

The standard open 3D head model in production pipelines has been FLAME, released in 2017 by the Max Planck Institute for Intelligent Systems. FLAME's influence on facial AI research is substantial: most published methods for 3D face reconstruction, expression transfer, and neural face rendering use FLAME as the geometric foundation. Its reach into research is well documented.

The commercial use problem is equally well documented. FLAME is distributed under a non commercial research license. Studios, developers, and individual artists who want to use FLAME in a production pipeline or a commercial product need to contact the institute for a separate agreement. That process effectively excludes smaller teams and independent productions for whom a formal research licensing negotiation is not practical.

SMPL-X, a related model that covers the full human body including face and hands, carries the same restriction. Both FLAME and SMPL-X are research landmarks that established the standard approach to parametric human head modeling. Neither one was ever freely available for commercial use. GNM Head is the first model at this level of completeness to ship under Apache 2.0.

The release arrives as parametric 3D representations are seeing increased use as conditioning signals for neural rendering and generative models. When a generative model produces a head, it needs a compact representation to receive expression and identity control. A parametric model that covers both the surface geometry and the internal anatomy in a single parameter vector is more useful as a conditioning interface than a model that covers only the external face. GNM Head is designed to function in that role.

The timing also reflects the accumulation of production use cases that require legal clarity on commercial rights. Studios integrating 3D head models into AI assisted workflows, where outputs enter commercial productions, could not use FLAME or SMPL-X without a separate license negotiation. GNM Head removes that barrier while expanding the model's technical capabilities beyond what FLAME provides.

383 Expression Blend Shapes Versus Prior Models

FLAME includes around 50 expression blend shapes. SMPL-X includes a similar count at the face region. GNM Head includes 383, more than seven times as many, covering subtle intermediate facial states that the smaller expression space cannot represent.

The practical consequence appears in performance retargeting workflows. When a captured performance is mapped to a parametric head model, the model's blend shape count sets the ceiling on how much of the original performance survives the mapping. A model with 50 blend shapes rounds each expression to the nearest available shape. A model with 383 shapes can represent the nuance between a half smile and a quarter smile, between a brow raise from surprise and a brow raise from skepticism, between a mouth shape for the phoneme "th" and the phoneme "f" under real speech conditions.

For audio driven animation, the expression space determines lip sync quality. Phoneme to viseme mappings require distinct mouth shapes for sounds that are visually similar but acoustically different. A 383 shape space gives lip sync systems enough target resolution to hold up under close inspection, including the kind of close scrutiny that digital doubles in major productions receive in post production review.

The 383 shapes also include expressive states that FLAME does not cover at all. Micro expressions, partial expressions, and asymmetric facial movements, where one side of the face moves differently from the other, require blend shapes that address each side independently. GNM Head's larger expression space accommodates this range.

The gap becomes more concrete when compared to real time capture standards. Apple's ARKit face tracking, which drives Face ID and AR features on iPhone, uses 52 blend shapes as its standard expression interface. GNM Head's 383 shapes represent a 7x expansion over that consumer face tracking baseline, meaning a parametric model built on GNM Head can receive and retain far more of a live performance than one built on ARKit scale expression spaces.

Performances captured with mobile face capture rigs, a common approach for previz and secondary character work where cost and speed matter, can now be retargeted to GNM Head parameters with significantly less expression loss than was possible when FLAME was the destination model. The larger target space preserves the subtleties that differentiate good screen performance from flat movement.

Complete Internal Anatomy in One Parameter Space

Prior open head models represent the external surface of the face. They do not model the geometry of the eyeballs, the teeth, or the tongue as separate controllable components. Productions that need those elements for close up work have historically handled them with separate rigs or manual geometry, created outside the parametric model and then joined in a separate integration step.

GNM Head places all four components in a single unified parameter space. The head surface uses 170 shape components. Teeth shape variation uses 80 shape components. Expression blend shapes use 383 entries. Eyeball rotation adds directional control over gaze. A single parameter vector drives identity, expression, and anatomy together.

This matters in practice for scenes that include open mouth expressions, wide angles, or dental shots. A unified space means the teeth and tongue geometry are always consistent with the head surface under any parameter combination. When a rigged head turns, opens its mouth, and raises its brow simultaneously, the internal anatomy moves with it without manual correction. In a separately assembled rig, maintaining that consistency across all three movements requires manual intervention or constraint setup that does not generalize across characters.

The unified parameter space also simplifies fitting. Algorithms that fit a parametric head model to image or video input can optimize a single vector rather than running separate fitting pipelines for the head surface, the teeth, and the gaze direction. GNM Head's fitting procedure is a single optimization problem by design.

Four Framework Backends

GNM Head supports NumPy, JAX, PyTorch, and TensorFlow through the same interface. A research team building a custom training pipeline in JAX and a production team running inference in PyTorch can both use the same parametric model without conversion. The model itself is framework agnostic.

The support for all four frameworks removes a common integration obstacle for teams that use a different framework than the one a model was originally released in. FLAME and SMPL-X have community conversions to PyTorch but those conversions are unofficial and do not always reflect the most current model version. GNM Head provides official support across all four frameworks from the initial release.

Framework agnosticism also matters for teams building hybrid pipelines where different stages use different frameworks. A pipeline that uses TensorFlow for 3D Gaussian Splatting reconstruction, PyTorch for the generative model, and NumPy for post processing utilities can use GNM Head parameters natively at each stage without conversion overhead.

Official framework support also matters for maintenance over time. Community conversions of FLAME to PyTorch have existed for years, but they require manual updates whenever the underlying model changes. When Google updates GNM Head, the four backend interface updates in the same commit. Production pipelines that pin to a specific GNM Head commit get a consistent model across all frameworks without tracking separate community conversion repositories.

The single code path across four frameworks also simplifies testing. A team that verifies GNM Head behavior in NumPy for quick prototyping can run the same parameter set through PyTorch for production inference without worrying about numerical differences introduced by conversion. Parametric consistency across frameworks is a precondition for using the model as a reliable intermediate representation in a multi stage pipeline.

Performance Capture, Rigging, and NeRF Conditioning

The paper documents GNM Head applications in four areas: performance capture, facial rigging, audio driven animation, and NeRF/3DGS conditioning.

In performance capture, GNM Head serves as the output template that a captured performance is fitted to. The 383 expression blend shapes provide enough resolution for the fitted parameters to represent the full range of a performance, including expressions that simpler models would round to the nearest available blend shape. Productions using facial capture for digital double work can use GNM Head as the standard parametric representation that all downstream steps operate on.

For facial rigging, GNM Head provides a foundation that character artists can build on rather than starting from scratch. A rigger who uses GNM Head as the base parameterization inherits the 383 expression shapes and the teeth and eye anatomy without designing those components. Custom blend shapes for a specific character can be layered on top of the GNM Head foundation.

NeRF and 3D Gaussian Splatting methods benefit from a parametric head model that provides geometric conditioning for novel view synthesis. Conditioning a NeRF or 3DGS reconstruction on GNM Head parameters makes it possible to control the expression and identity of a reconstructed head by adjusting the parameter vector rather than by modifying the underlying scene representation directly. This is particularly useful for relighting and view synthesis tasks where the head geometry needs to respond to expression changes while the scene rendering remains stable.

Audio driven animation with GNM Head targets the lip sync use case directly. Fitting viseme parameters to the 383 shape space at each frame of a speech audio track gives the lip sync system enough degrees of freedom to represent coarticulation, the overlapping mouth shapes that occur between consecutive phonemes in natural speech. Lip sync systems built on FLAME's 50 shapes produce transitions that look mechanical under the close shot distances that dialogue scenes typically use.

The four documented application areas are not a closed list. The Apache 2.0 license and four backend framework support mean that researchers and developers can apply GNM Head to additional tasks, including real time avatar driving, synthetic training data generation, or expression transfer for stylized characters, without requesting permission or negotiating a research agreement.

Teeth Shape Components

The 80 teeth shape components in GNM Head's unified parameter space address a geometry layer that FLAME and SMPL-X left unparameterized. Prior workflows that needed controllable teeth geometry required a separate teeth model, typically a static mesh library or a custom rig built per character, that was assembled with the head model in a separate integration step.

A parameterized teeth space matters for productions where characters have distinctive dental features, for dental and medical visualization where accurate teeth geometry is the primary output, and for any close range render where the mouth is open or partially open. GNM Head's 80 teeth shape components represent individual variation in tooth size, spacing, and alignment, not just a single open or closed state.

The teeth parameterization is also relevant for lip sync workflows where the visible teeth position affects the plausibility of a phoneme. Dental geometry that is inconsistent with an open mouth expression produces a visible discontinuity that post processing cannot cleanly fix. A unified parameter space that links teeth shape to expression blend shapes prevents that discontinuity at the model level.

Apache 2.0 and the Blender Plugin

The Apache 2.0 license permits commercial use, modification, and redistribution with attribution. Prior models required contacting the research group for a commercial license. That process excluded smaller teams and independent production companies for whom a formal research agreement was not practical.

A community Blender plugin for GNM Head appeared on GitHub within 12 hours of the model's release. The plugin connects GNM Head's parameter space to Blender's shape key and bone rigging system, making the model accessible to artists who work in Blender without writing custom integration code. The speed of the community response indicates that GNM Head addresses a gap that Blender users in VFX and character animation had been waiting for.

The plugin is separate from the official Google Research release and is maintained by the community. Its rapid availability reflects a pattern common to Apache 2.0 releases of foundational 3D tools. The permissive license enables community integration work to begin immediately, without waiting for official integration support.

For indie VFX artists, the combination of Apache 2.0 licensing and a day one Blender plugin means that GNM Head's 383 shape expression space is accessible through a free, widely used 3D application with no licensing cost. The same head model that a major studio's performance capture pipeline uses for digital double work is now available to independent animators with an internet connection.

Access: 29 Contributors, GitHub and arXiv

GNM Head was developed by 29 contributors at Google Research. The paper was posted to arXiv on July 26, 2026. The model, code, and assets are available on GitHub at google/GNM. The repository includes the model weights, the four backend inference library, and documentation for fitting GNM Head parameters to image or video input.

At 29 contributors, GNM Head represents an unusually large authorship for a single parametric model release. Projects at this scale typically require specialists across geometry modeling, expression capture, rigging, and framework engineering to contribute distinct components. The breadth of the author list is consistent with a system that covers as many distinct anatomy and application layers as GNM Head does.

For teams starting from scratch, the recommended entry point is the repository's quick start documentation, which walks through installing the Python package, loading a GNM Head instance from the provided weights, and generating a mesh from a parameter vector. The multi framework interface means the quick start works regardless of whether a team's existing codebase uses PyTorch or TensorFlow.

The arXiv paper's appendix covers the mathematical formulation of the unified parameter space in detail, including the PCA basis construction for head shape components and the blend shape optimization used to derive the 383 expression shapes from a large scale capture dataset. Teams building custom fitting algorithms or extending the model to new anatomy regions will find the appendix the most relevant technical reference.

For character animation pipelines that combine head geometry with full body motion, GNM Head pairs with motion tools in the same open source ecosystem. ARDY from NVIDIA Research generates full body motion from text at 30 frames per second with kinematic constraints. GNM Head provides the head geometry layer that body motion systems need for complete character representation.

For motion forecasting in dynamic scenes, MolMoMotion from AI2 addresses body prediction from visual input, a complementary capability to GNM Head's head parameterization. For the performance capture context that GNM Head fits into directly, Andy Serkis has argued for 25 years that digital performance deserves recognition as authorial work. GNM Head's open sourcing provides the parametric foundation layer that makes that argument's technical premise increasingly accessible.

AI assisted character work for video production can be combined with the AI FILMS Studio video workspace, which provides access to the latest text-to-video and image-to-video generation models.

Fitting to Live Capture Input

The paper documents a fitting algorithm that estimates GNM Head parameters from image or video input. The fit runs as a single optimization over the unified parameter vector, estimating identity shape, expression, teeth geometry, and gaze direction together from a single image rather than running separate pipelines per anatomy component.

Single image fitting is the entry point for workflows that need a parametric head estimate from photography rather than from dedicated 3D scanning hardware. A portrait photograph, a video frame, or a sequence of stills can produce a GNM Head parameter estimate that includes teeth and eye geometry without requiring an oral cavity scan or an eye tracker.

For audio driven animation, the fitting pipeline converts reference footage into GNM Head parameters that serve as training or calibration data. The 383 expression space provides enough resolution for the fitted expression trajectories to capture lip sync nuances that smaller expression spaces would lose during the parameter estimation step.

The unified fitting optimization supports batch processing workflows. A scan session that yields multiple expressions of the same subject can be processed as a batch, producing a parameter archive that covers the subject's expression range. That archive serves as the foundation for downstream synthesis, including generating new expressions of the same subject, transferring expressions to different identity parameters, or conditioning a generative model on specific identity plus expression combinations.

The fitting algorithm provides a quality metric for how well the estimated parameters reconstruct the input image. That metric can flag frames where the fit quality falls below a threshold, sending them for manual review rather than passing bad parameter estimates downstream into an automated pipeline.

For episodic productions, where dozens of capture sessions need to maintain consistent quality across an episode run, that automated quality signal reduces review time substantially.

The fitting pipeline's input is not limited to controlled studio footage. Papers in the GNM Head lineage demonstrate parameter estimation from unconstrained video footage, portrait photographs, and images captured under uncontrolled lighting. A production that has archival footage of a subject can fit GNM Head parameters to that footage and use the resulting parameter archive to guide synthesis, rigging, or rendering work that the original footage was not shot for. The Apache 2.0 license means those fitted parameters can be used in commercial deliverables without restriction.

---

Sources

arXiv: GNM Head paper GitHub: google/GNM Blender plugin: derQwertzus/GNM-Head-Add-on-for-Blender