Google Releases Parametric 3D Head Model That Runs on CPUs

Google has released gmn, a differentiable parametric 3D head model that runs on CPUs, according to an announcement from the HuggingApps X account on July 24, 2026. The model is designed to generate facial movements and expressions in real time and is available as an interactive demo on Hugging Face Spaces. gmn, which stands for Google Neural Model, is a differentiable parametric model, meaning it can be optimized directly through gradients, similar to neural networks. This allows highly precise adjustments to facial shape and expression without requiring a GPU, a significant breakthrough for applications on resource constrained devices. According to Google's official GitHub repository, gmn is part of an open ecosystem of parametric human models and perception stacks, starting with GNM Head.

The model is based on real world scan data to produce realistic 3D heads and accurate facial expressions. Google has also released add ons for Blender and 3ds Max, enabling direct integration into 3D artists' workflows. The interactive demo hosted on Hugging Face Spaces allows users to try the gmn model directly in a browser. Users can manipulate facial parameters in real time, observing how small changes to eyebrow shape, mouth, or eyes produce different expressions. This demonstrates the model's ability to handle subtle and complex facial variations. The ability to run the model on a CPU is a key feature that distinguishes gmn from other 3D face models that typically require GPUs.

This opens the door for applications on mobile devices, laptops, and even edge devices where GPU access is limited. Developers can integrate the model into real time applications such as virtual avatars, games, and video conferencing tools without specialized hardware. gmn is open source, with source code available on GitHub under Google's organization. This allows the research and developer community to modify, extend, and integrate the model into their own projects. Google also provides add ons for Blender and 3ds Max, industry standard tools for 3D artists, facilitating adoption among professionals. The parametric approach means the model uses a set of adjustable parameters to control facial shape and expression. This differs from static mesh models that require manual re creation for each variation.

With gmn, users can quickly generate a wide range of facial variations by simply changing parameters, saving time and resources in 3D content production. The model is based on real world scan data, making generated shapes and expressions more realistic than manually created models. Google collected data from multiple subjects to ensure the model can handle the diversity of human facial shapes. This is important for applications such as digital avatars that need to look natural and convincing. The availability of add ons for Blender and 3ds Max indicates that Google is targeting not only researchers but also professional 3D artists. These add ons allow users to import the gmn model directly into their scenes and start animating without writing code.

This accelerates production workflows for film, games, and virtual reality. gmn is also differentiable, meaning it can be used in machine learning pipelines. Developers can train the model for tasks such as face tracking, expression synthesis, or even face generation from audio. This capability makes gmn a flexible tool for various AI applications involving human faces. With the release of gmn, Google strengthens its position in the open source parametric human model ecosystem. This aligns with industry trends toward more open and accessible models, enabling faster innovation in computer graphics and AI. The model is expected to be used in a variety of applications, from entertainment to education and healthcare. The interactive demo on Hugging Face Spaces has attracted attention from the AI and graphics communities.

Users can access it directly through a browser without installation, demonstrating the model's ease of use. It also serves as a learning tool for those wanting to understand how parametric 3D models work. Google has not announced specific plans for further development of the gmn model, but the GitHub repository indicates it is an active project. The community is expected to contribute by adding new features, improving performance, or integrating the model with other tools. The open source nature ensures the model will continue to evolve over time. Overall, the release of gmn marks a step forward in democratizing 3D face technology.

With CPU compatibility, support for industry standard tools, and an open source license, the model has the potential to become a new standard for applications requiring real time facial simulation.