How much memory does it need?

TRELLIS.2 generates 3D shape and materials from one image. Use a photo where the object is clear and separated from the background. The repository requires Linux and at least 24 GB of NVIDIA GPU memory, recommends CUDA Toolkit 12.4, and says it was tested on A100/H100. This setup therefore targets a Linux GPU meeting those conditions. [Official TRELLIS.2 repository](https://github.com/microsoft/TRELLIS.2)

Model setup recipes

Create a 3D asset from one photo with TRELLIS.2

Start by checking Linux and an NVIDIA GPU with at least 24 GB.

TRELLIS.2 is a 4B-parameter model for generating textured 3D assets from one input image. Its official example reads a photo and saves sample.glb plus a sample.mp4 preview. The repository lists a 24 GB minimum NVIDIA GPU¹ memory requirement and uses a Linux environment.

Prepare the photo and check the hardware

TRELLIS.2 generates 3D shape and materials from one image. Use a photo where the object is clear and separated from the background. The repository requires Linux and at least 24 GB of NVIDIA GPU memory, recommends CUDA² Toolkit 12.4, and says it was tested on A100/H100. This setup therefore targets a Linux GPU meeting those conditions. Official TRELLIS.2 repository

An image workspace examining a product photo’s shape and material — Use the shape and material cues in one photo as input for 3D generation.

Set up the repository and environment

Clone the repository with submodules and run setup.sh to install dependencies. The README recommends Python 3.8+ and CUDA Toolkit 12.4; building CUDA extensions can take time. Run the script from a Linux shell; downloading the checkpoint³ also requires Hugging Face access.

Run the official setup script

git clone -b main https://github.com/microsoft/TRELLIS.2.git --recursive
cd TRELLIS.2
. ./setup.sh --new-env --basic --flash-attn --nvdiffrast --nvdiffrec --cumesh --o-voxel --flexgemm

Convert the example image to GLB

After installation, run the official example.py. In the file, replace the default path in Image.open('assets/example_image/T.png') with your photo’s path and save it. The microsoft/TRELLIS.2-4B weights generate the shape and save sample.glb and the turntable preview sample.mp4 in the run directory.

Generate a sample 3D asset

# Edit example.py: replace Image.open('assets/example_image/T.png') with your photo path
python example.py
# Outputs: sample.glb and sample.mp4

A local workbench prepared with product images and drafts — A clear product photo makes the generated shape easier to inspect.

Inspect the GLB geometry and surfaces

When the command finishes, open sample.glb in a 3D viewer and rotate it. Inspect the back, thin parts, and areas around holes; finish by checking the GLB you will deliver, not just the sample.mp4 preview.

For product drawings or measurement models that need exact dimensions, do not use a single-photo GLB as-is. Blocking out a shape quickly and producing a model for manufacturing need different preparation.

A board comparing image outputs at different sizes — After generation, rotate the GLB in a viewer to inspect it.

Terminology notes

GPU — A processor designed to handle many calculations in parallel. It performs model computations during AI inference.
Back to the text
CUDA — A software platform for general-purpose computing on NVIDIA GPUs.
Back to the text
Checkpoint — A file containing saved model weights and related state. Versions or tasks in one model family may use different checkpoints.
Back to the text

Frequently asked questions

What should you check first in this article?: TRELLIS.2 is a 4B-parameter model for generating textured 3D assets from one input image. Its official example reads a photo and saves `sample.glb` plus a `sample.mp4` preview. The repository lists a 24 GB minimum NVIDIA GPU memory requirement and uses a Linux environment.
How much memory does it need?: TRELLIS.2 generates 3D shape and materials from one image. Use a photo where the object is clear and separated from the background. The repository requires Linux and at least 24 GB of NVIDIA GPU memory, recommends CUDA Toolkit 12.4, and says it was tested on A100/H100. This setup therefore targets a Linux GPU meeting those conditions. [Official TRELLIS.2 repository](https://github.com/microsoft/TRELLIS.2)