Skip to the content
Know your tech before you buy it. Makers: get listed from $49/mo

TRELLIS

trellis3d.github.io

  • Windows
  • Linux
  • Web
  • Self-hosted

What TRELLIS is

TRELLIS is a free, open-source research model that creates 3D assets from text or image prompts. It can output radiance fields, 3D Gaussians, and meshes through a shared Structured Latent representation. Prompts can generate asset variants or target local regions for edits. Pretrained image-to-3D and text-to-3D models range from 342 million to 2 billion parameters and were trained on 500,000 3D objects. Example exports include textured meshes in GLB and 3D Gaussian files in PLY. A Gradio demo can be installed and run locally. Running the project requires an NVIDIA GPU with at least 16 GB of memory, CUDA for some submodule builds, and Python 3.8 or newer; Conda is recommended. Linux is tested, while Windows setup is not fully tested. The project is strongest for artistic-style assets and has limited photorealistic real-world object generation. Materials are intended for academic and research use, not commercial use.

Who it suits

TRELLIS suits researchers, developers, and artists creating prompt-driven 3D assets on compatible NVIDIA hardware. It is less suitable for commercial projects or users needing photorealistic real-world objects.

Good to know, and what to watch

  • Text-to-3D and image-to-3D models
  • Generates meshes, radiance fields, and 3D Gaussians
  • Local Gradio demo available
  • Exports GLB and PLY files
  • Supports regional prompt editing
  • Requires NVIDIA GPU with 16 GB memory
  • Windows setup is not fully tested
  • Commercial use is not intended
  • Photorealistic object generation is limited

The verdict

TRELLIS offers several 3D representations, editable regions, and local demo options for research-oriented workflows. Hardware, operating-system, and usage restrictions should be checked before setup.

What the maker publishes

Demo
The repository provides a Gradio web demo that users can run locally after installing its demo dependencies. source
Dependencies
Installation requires CUDA for compiling some submodules and Python 3.8 or higher, with Conda recommended for dependency management. source
Editing
It can generate variants of an asset and manipulate targeted local regions using text or image prompts. source
Exports
The usage example exports textured meshes as GLB and 3D Gaussians as PLY files. source
Formats
The example exports textured meshes as GLB files and 3D Gaussian representations as PLY files. source
Hardware
The code requires an NVIDIA GPU with at least 16 GB of memory; the repository says it has been verified on A100 and A6000 GPUs. source
Hardware requirement
The repository says an NVIDIA GPU with at least 16 GB of memory is necessary. source
License
TRELLIS models and most code use the MIT License, while some submodules have separate licenses. source
License and use
The project page says its materials are for academic and research purposes and are not intended for commercial use; the repository says the models and most code use the MIT License, with some submodules under different licenses. source
Model scale
The project provides pretrained models with up to 2 billion parameters, trained on a dataset of 500,000 3D assets. source
Model sizes
The project describes models with up to 2 billion parameters trained on a dataset of 500,000 3D objects. source
Model variants
The repository lists image-to-3D and text-to-3D pretrained models ranging from 342 million to 2 billion parameters. source
Operating system
The code is tested on Linux, and the repository describes Windows setup as not fully tested. source
Operating systems
The code is tested on Linux; Windows setup is described as not fully tested. source
Outputs
The model can produce Radiance Fields, 3D Gaussians, and meshes. source
Purpose
TRELLIS is a research model that generates 3D assets from text or image prompts. source
Quality
The project says its models generate detailed geometry and texture, while noting that photorealistic real-world object generation is limited. source
Representations
It can produce radiance fields, 3D Gaussians, and meshes from a shared Structured Latent representation. source
Research limits
The project page says the model is strongest at artistic-style assets and has limited ability to generate photorealistic real-world objects. source
Security reporting
The project directs vulnerability reports to the Microsoft Security Response Center rather than public GitHub issues. source
Support
The repository invites users who encounter installation issues to open an issue or contact the project maintainers. source
Support
The repository uses GitHub Issues for bug reports and feature requests, and says support is limited to the listed resources. source
Text-generation caveat
The repository recommends generating an image first and then using the image-conditioned model because text-conditioned models are less creative and detailed due to data limitations. source
Web demo
The repository provides a Gradio web demo that users can install and run locally. source

Checked against the maker's pages on 7 Oct 2026.