Toronto AI Lab

GET3D by NVIDIA is a generative AI model that creates high-quality 3D textured meshes directly from 2D image collections. Unlike previous approaches that often produce limited topology or lack textures, GET3D by NVIDIA generates explicit 3D meshes with complex geometries and high-fidelity textures that are immediately usable in standard 3D rendering engines. The tool works by learning from adversarial losses defined on 2D images, using a differentiable renderer to obtain RGB images and silhouettes during training.

The system utilizes two latent codes—one for geometry and one for texture—allowing for disentangled control over shape and appearance. It employs DMTet to extract 3D surface meshes from signed distance fields and queries texture fields at surface points to generate colors. What makes GET3D by NVIDIA stand out is its ability to produce diverse shapes with arbitrary topology while maintaining high-quality geometry and texture details like wheels on chairs, car lights, and intricate animal features.

This tool primarily benefits 3D content creators, game developers, and virtual world builders who need to rapidly generate large quantities of high-quality 3D assets. Real-world use cases include creating diverse 3D models for gaming environments, generating architectural visualizations, and producing textured assets for VR/AR applications without manual modeling effort.