Get3D Nvidia

GET3D is a generative model for 3D shape synthesis developed by Nvidia Research and published at NeurIPS 2022. The model generates explicit textured 3D meshes from a collection of 2D images without requiring any 3D ground truth data during training — a key architectural distinction from most prior 3D generative models. GET3D uses a two-branch generator: one branch produces a geometry code that defines the 3D shape using signed distance functions, and a second branch produces a texture field applied to the surface. Both branches are trained adversarially against a 2D discriminator using rendered image pairs, allowing the model to learn 3D structure purely from 2D supervision. Output 3D assets export as textured mesh formats compatible with standard rendering pipelines in tools including Unity and Unreal Engine. GET3D is primarily used by AI and computer vision researchers studying 3D generative model architectures, and by game developers and technical artists exploring how generative models can accelerate early-stage 3D asset creation workflows. The model can generate category-specific 3D assets — vehicles, chairs, buildings, characters, and other object categories — when trained on domain-specific 2D image datasets. This makes it useful for prototyping stylistically consistent asset collections for a game or scene without manually modeling each object. Researchers use it as a baseline model for benchmarking new 3D generation approaches or for studying the properties of GAN-based 3D synthesis compared to diffusion-based alternatives. Find alternatives.