GET3D by NVIDIA

What is GET3D by NVIDIA?

GET3D by NVIDIA is an open research model that generates textured 3D meshes, such as cars, chairs and motorbikes, which open directly in standard 3D software. It was built by NVIDIA’s Toronto AI lab with the University of Toronto and the Vector Institute, and presented at NeurIPS 2022 by a team that includes Jun Gao and Sanja Fidler.

The model never sees 3D files during training. It learns from collections of 2D renders, using two latent codes: one drives a signed distance field for the shape, the other a texture field for colour. DMTet turns the shape into a triangle mesh, a differentiable rasteriser renders it, and separate RGB and silhouette discriminators compare the result with real renders. NVIDIA reported training on about one million synthetic images in roughly two days on A100 GPUs, and inference of around 20 shapes a second on a single GPU.

There is no hosted app or web interface. You download the PyTorch code from GitHub, add pretrained checkpoints from Hugging Face, and run commands from a terminal. The main repository news stops in September 2023 with FlexiCubes support, and the latest change we found was a small link fix in September 2024, so treat it as a finished research release rather than an active product.

Key Features

GET3D’s most useful features are explicit textured mesh output and ready-made checkpoints, because together they let you sample usable 3D objects without running a multi-day training job first.

• Textured mesh generation: Each sample is a triangle mesh with its own texture rather than a voxel grid or point cloud. Adding the textured mesh flag to the inference command writes OBJ files with MTL material data and texture images.
• Pretrained checkpoints: NVIDIA provides checkpoints for the ShapeNet car, chair, motorbike and table categories, hosted on Hugging Face. Inference needs one GPU with at least 16 GB of memory, so you can test results without training.
• Separate shape and texture codes: Geometry and texture come from two independent latent codes, so you can keep one shape and swap its texture, or the reverse. The project page reports good separation between the two.
• Latent interpolation: A single option saves smooth transitions between generated objects, for example morphing one car body into another. Small changes to a latent code also produce close variants of the same object.
• FlexiCubes support: A September 2023 update added FlexiCubes as a drop-in alternative to DMTet for extracting surfaces. You switch it on by appending one flag to the training or inference command.
• Evaluation tools: Built-in options compute FID on rendered images and coverage and MMD scores on generated geometry. Researchers use these to compare new 3D models against GET3D under the same conditions.

Pros

GET3D’s main advantage is free access to the full code of a widely cited 3D generator, with outputs that load in ordinary 3D tools.

• Free and open code: The complete PyTorch training, inference and evaluation code is public on GitHub. Researchers can read every step of the pipeline instead of relying on a closed hosted service.
• Practical output files: Generated objects are standard OBJ meshes with textures, not neural representations. They open in Blender or similar packages, where you can clean, rig or render them like any other asset.
• Fast sampling: NVIDIA reported about 20 shapes a second on a single GPU once a model is trained. That speed suits building large batches of varied props or synthetic training data.
• Fully local processing: Everything runs on your own hardware or cloud instance. No data or generated assets pass through a third-party service, which matters for studios with confidential projects.
• Retrainable on new categories: You can train on your own rendered objects using the supplied ShapeNet rendering scripts as a template. Debug runs work on one GPU with a batch size of four.

Cons

GET3D’s biggest drawback is its licensing, because both the code and the pretrained weights are restricted to non-commercial research or evaluation.

• Non-commercial licences: The code uses the NVIDIA Source Code License and the checkpoints use CC BY-NC-SA 4.0. Studios wanting to ship generated assets must request terms through NVIDIA Research Licensing.
• Demanding hardware: Default training uses eight high-end GPUs, such as V100 or A100 cards, with a batch size of 32. Even inference needs 16 GB of GPU memory, which rules out most laptops.
• No text or photo input: The public release samples objects from random codes within one trained category. Text-guided edits shown in the paper rely on extra fine-tuning that the README does not package as a simple command.
• Ageing software stack: The repository targets Python 3.8, PyTorch 1.9 and CUDA 11.1. Newer drivers and libraries often need patching, so the included Docker setup is the safest route.
• Narrow trained categories: Each checkpoint covers one ShapeNet category trained on synthetic renders. Results outside cars, chairs, motorbikes and tables require your own dataset and days of GPU time.

Who is Using GET3D by NVIDIA?

GET3D is used mainly by 3D generative AI researchers, who cite and benchmark it as a mesh-based baseline, with smaller groups of students and technical artists testing it.

• AI researchers: Academic and industry labs run GET3D as a comparison baseline. Its FID and coverage scripts make side-by-side tests with newer 3D generators straightforward.
• Graphics students: Learners study the code to see differentiable rendering and DMTet in practice. The Colab notebook lets them try inference without owning a large GPU.
• Technical artists: Studio pipeline staff test how generated meshes behave in Blender and game engines. They mostly evaluate it for internal prototypes because of the licence.
• Simulation teams: Engineers building virtual environments explore it for filling scenes with varied vehicles or furniture. NVIDIA pitched it at populating virtual worlds with many distinct objects.

Pricing

GET3D by NVIDIA is free to download, though you pay for the GPU time needed to run or train it (official repository and model card, October 2026).

Some news coverage called it a single-image converter, and one directory lists GLB and USD export, neither of which the official README supports.

Disclaimer: the two licence files, not the $0 price, decide what you may do with generated meshes, so read both before using results in any product.

What Makes GET3D by NVIDIA Unique?

GET3D outputs explicit textured meshes with complex topology, while many earlier 3D generators produced voxels, point clouds or results that needed a neural renderer. Because the mesh and texture come out as ordinary files, they drop into Blender or a game engine without special viewers.

Its training method also matters: it learns 3D structure purely from 2D renders through differentiable rendering and adversarial losses. That combination made it a standard reference point for later work such as FlexiCubes, which NVIDIA then added back into the repository.

Compatibilities and Integrations

GET3D runs on Linux with NVIDIA GPUs and connects to other tools only through exported files, since it has no API or plug-ins.

• Platforms: Linux is recommended, and Docker build files are included for server use.
• Hardware: NVIDIA GPUs with CUDA 11.1 or later; 16 GB of memory for inference.
• Libraries: PyTorch 1.9, nvdiffrast and Kaolin, with Python 3.8.
• Export: OBJ meshes with MTL files and texture images for Blender and other 3D packages.
• Data: Rendering scripts prepare ShapeNet; other datasets need your own renders.

GET3D by NVIDIA Tutorials

The quickest test is a pretrained car checkpoint run in the official Colab notebook, which avoids local setup entirely.

1. Open the Colab notebook linked from the pretrained model page, or clone the repository and build the Docker image.
2. Download the car, chair, motorbike or table checkpoint from Hugging Face.
3. Run the inference command with the visualisation option on a GPU with 16 GB of memory.
4. Add the textured mesh option to export OBJ files with materials.
5. Add the interpolation option to save transitions between samples.
6. Import the OBJ files into Blender to check scale, topology and textures.

How We Rated It

GET3D by NVIDIA scores 3.2 out of 5 overall, because strong local privacy and research value are offset by a hard setup and non-commercial terms.

Summary

GET3D by NVIDIA suits researchers, graphics students and technical artists who want to study or benchmark a well-known mesh-based 3D generator. It produces real textured OBJ files quickly once trained, and it keeps all data on your own hardware. The main limits are a non-commercial licence, an older PyTorch stack and the need for at least a 16 GB GPU. It also cannot turn a prompt or photo into a model without extra work. Start with the Colab notebook and the car checkpoint before committing to any training run.