NVIDIA GET3D

Type: Research Tags: NVIDIA, GET3D, 3D Generation, Generative AI, 3D Shapes, Neural Rendering, NVlabs Related: NVIDIA-NeRF, NVIDIA-Instant-NGP, NVIDIA-GauGAN, NVIDIA-Omniverse, NVIDIA-Research Sources: NVIDIA official documentation Last Updated: 2026-04-10

Summary

GET3D is a generative model from NVIDIA Research Toronto (NVlabs) that learns to synthesize high-quality 3D textured shapes directly from 2D image collections, without requiring 3D supervision. Published at NeurIPS 2022, GET3D uses a GAN-based architecture combined with differentiable rendering to generate diverse 3D meshes with detailed textures, trained on images rendered from 3D datasets. It generates watertight triangle meshes with UV-mapped textures compatible with standard 3D workflows and game engines.

Detail

Purpose

Creating 3D assets is time-consuming and requires specialized skills. GET3D automates 3D content creation by learning a generative model of shapes and textures from 2D supervision only, making it possible to synthesize virtually unlimited novel 3D objects for use in games, VR/AR, e-commerce, and simulation.

Key Features

  • GAN-based 3D generative model trained from 2D image collections
  • Outputs: watertight triangle meshes with UV-mapped albedo textures
  • No 3D supervision required: trained on multi-view rendered images
  • Disentangled geometry and texture: independently sample shape and appearance
  • High-diversity generation: generates varied, non-repetitive 3D assets
  • Fast synthesis: single GPU inference in under 1 second
  • Compatible with standard 3D pipelines: .obj / .glb export
  • Trained on ShapeNet categories: cars, chairs, motorbikes, humans, animals
  • Open-source code on GitHub (NVlabs/GET3D)

Use Cases

  • Game asset generation: synthesize props, vehicles, and furniture
  • VR/AR content creation: procedurally generated 3D environments
  • E-commerce: generate 3D product previews for virtual try-on
  • Simulation: populate synthetic training environments with diverse objects
  • Film: rapid concept art and prop prototyping
  • Research: 3D generative model baseline for papers

Hardware Requirements / Compatibility

  • Training: 4–8x A100 80GB (ShapeNet-scale)
  • Inference: single RTX 3090 / A100
  • CUDA 11.3+
  • PyTorch 1.11+
  • Requires nvdiffrast differentiable renderer

Language Bindings / APIs

  • Python (PyTorch)
  • nvdiffrast for differentiable mesh rendering
  • Export to .obj, .glb for Unity/Unreal/Omniverse
  • Open-source on GitHub (NVlabs/GET3D)

Connections

  • NVIDIA-NeRF — GET3D uses neural rendering like NeRF but generates explicit meshes
  • NVIDIA-Instant-NGP — both are NVlabs neural graphics research projects
  • NVIDIA-GauGAN — GauGAN generates 2D images; GET3D generates 3D geometry
  • NVIDIA-Omniverse — GET3D meshes importable into Omniverse for USD-based workflows
  • NVIDIA-Research — GET3D is from NVIDIA Research Toronto lab (NVlabs)

Resources