Skip to content
Get packs

TRELLIS

Generates 3D assets from text or images, exportable as meshes, Gaussians, or radiance fields.

GitHub Download source
Stars
14.2k
Forks
1.4k
Licence
MIT
Language
Python
Last update
10 days ago
Type
AI model / agent

What it is

TRELLIS is a large 3D generation model from Microsoft Research. You give it a text prompt or an image, and it produces a 3D object in formats such as meshes, 3D Gaussians, or radiance fields. It is built on a shared 'structured latent' representation that can be decoded into different output formats, and it also supports local editing and generating variants of an object.

  • Generate a 3D asset from a single input image
  • Generate a 3D asset from a text prompt using the TRELLIS-text models
  • Create variants of an existing generated object
  • Make local edits to a generated 3D asset
  • Export generated results as Gaussians or meshes

Great for

  • Researchers exploring text- or image-to-3D generation
  • Developers with an NVIDIA GPU who want to prototype 3D asset pipelines in Python
  • Creators who need quick 3D starting points from a single image
  • People who want to export one generated asset in several formats

Not ideal for

  • Users without an NVIDIA GPU with at least 16GB of memory, since the README says this is necessary
  • Windows users, because the code is tested only on Linux and Windows setup is not fully tested
  • Anyone wanting a polished no-code tool, since the project is a research codebase run from the command line and Python
  • Text-only workflows expecting top quality, since the README recommends generating images first for better results
Step-by-step guide: how to use TRELLISInstall, first run, examples, tips and FAQs
TRELLIS vs TRELLIS.2Shap-E vs TRELLIS

How to get it

  1. 1

    Clone and build from source

    You'll need: Linux · NVIDIA GPU with at least 16GB memory · CUDA Toolkit (tested with 11.8 and 12.2) · Conda (recommended) · Python 3.8 or higher

    Clone the repository with submodules

    git clone --recurse-submodules https://github.com/microsoft/TRELLIS.git

    Enter the project folder

    cd TRELLIS

    Create environment and install dependencies

    . ./setup.sh --new-env --basic --xformers --flash-attn --diffoctreerast --spconv --mipgaussian --kaolin --nvdiffrast
  2. 2

    Run it

    • Install the prerequisites: Linux, an NVIDIA GPU with at least 16GB memory, CUDA Toolkit, and Conda
    • Clone the repository with submodules and enter the folder
    • Run the setup script to create the trellis conda environment
    • Load a pretrained model such as microsoft/TRELLIS-image-large from Hugging Face in Python
    • Run the minimal example at example.py to generate a first 3D asset
  3. 3

    Get the most out of it

    • Use the image-conditioned model (TRELLIS-image-large), which the README recommends for better performance
    • Run the Gradio demo app.py or app_text.py to try generation interactively, if dependencies are installed with --demo
    • Export Gaussians or meshes from the same generation, as shown in the updates for app.py and example.py
    • Try multi-image conditioning for TRELLIS-image, noting it is tuning-free and may not suit every input
    • Use the variant generation example (example_variant.py) to create alternatives of an object

Downloads

Source code (.zip)main branch · straight from GitHubOfficial website & docstrellis3d.github.io

Strengths

  • Produces 3D output in several formats from one latent representation, per the README
  • Pretrained models are published on Hugging Face and load by repository name
  • Released under the MIT license with training code and a dataset toolkit

Trade-offs

  • Requires a high-end NVIDIA GPU and tested only on Linux, per the README
  • Text-conditioned models are described by the authors as less creative and detailed than image-conditioned ones
  • Installation involves many dependencies and CUDA version considerations that the README warns may take a while

How it compares

ProjectStarsTypeLevelUpdatedBest for
TRELLIS (this)14.2kAI model / agentadvanced10 days agoResearchers exploring text- or image-to-3D generation
shap-e12.3kAI model / agentintermediate2.4 years agoResearchers exploring text-conditioned 3D generation
three.js117kLibraryintermediate4 days agoWeb developers adding interactive 3D graphics to a website
react-three-fiber33.4kLibraryintermediate15 days agoReact developers who want to add 3D scenes to a web app
editor25.3kAppintermediate6 days agoPeople who want to sketch or model a building in 3D on their own machine

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.