Home Resources AI 3D generators
TRELLIS
Microsoft's open-source models that turn images into 3D

- Best for
- Technically minded developers with a large NVIDIA GPU who want a free, self-hosted image-to-3D model
- Makes
- Textured 3D meshes with PBR materials, exported as GLB
- Made by
- Microsoft
- Where you use it
- Linux
- Pricing
- Free and open source (MIT); you run it on your own hardware
- Free plan
- Yes
- Open source
- Yes (MIT)
- Export
- Exports GLB files you can import into Blender or a game engine
- Status
- Research
TRELLIS is a family of open-source 3D generation models published by Microsoft on GitHub. This entry covers the current version, TRELLIS.2, whose model weights and paper appeared in December 2025, and notes where the original differs.
The original TRELLIS came with a December 2024 paper, which its repo describes as a CVPR 2025 Spotlight. It creates 3D assets from text or image prompts and can output radiance fields, 3D Gaussians or meshes, using models of up to 2 billion parameters. TRELLIS.2 is a single 4-billion-parameter image-to-3D model built on a sparse voxel format its authors call O-Voxel. Its GitHub README says it copes with open surfaces such as clothing and leaves, models base color, roughness, metallic and opacity, and exports a GLB with PBR materials. The authors report roughly 3 seconds per asset at 512ยณ resolution and about 60 seconds at 1536ยณ on an NVIDIA H100 GPU. The repo also includes code for texturing an existing shape and for training.
Code and weights are MIT licensed, so the cost is your own hardware: the README says TRELLIS.2 is tested only on Linux and needs an NVIDIA GPU with at least 24GB of memory (16GB for the original). Microsoft also hosts a Hugging Face demo that runs in a browser. For game developers it is a way to generate 3D props from images on their own machines; like any generated mesh, results should be checked before they go into an engine. AIGamer has not tested it.
Limitation: TRELLIS.2 is tested only on Linux and needs an NVIDIA GPU with at least 24GB of memory
FAQ
Is TRELLIS free to use?
Yes. TRELLIS and TRELLIS.2 release their code and weights under the MIT license, though some dependencies have their own licenses. You need your own NVIDIA GPU, or you can try the Hugging Face demo.
What is the difference between TRELLIS and TRELLIS.2?
The original TRELLIS (2024) takes text or images and outputs radiance fields, 3D Gaussians or meshes, with models of up to 2 billion parameters. TRELLIS.2 (December 2025) is a 4-billion-parameter image-to-3D model with PBR materials and GLB export.
What GPU do I need for TRELLIS.2?
The README says an NVIDIA GPU with at least 24GB of memory, verified on A100 and H100 cards, and the code is tested only on Linux.
Can TRELLIS make 3D models from text?
The original TRELLIS includes text-to-3D models, but its authors recommend making an image with a text-to-image model first and then using the image model. TRELLIS.2 is image-to-3D only.
Sources
Compiled by AIGamer from TRELLIS's own pages and the sources below, checked 2 November 2026. We haven't tested TRELLIS hands-on.
- GitHub: microsoft/TRELLIS.2 README
- GitHub: microsoft/TRELLIS.2 LICENSE
- GitHub: microsoft/TRELLIS README
- arXiv: Native and Compact Structured Latents for 3D Generation (Xiang et al., 2025)
- arXiv: Structured 3D Latents for Scalable and Versatile 3D Generation (Xiang et al., 2024)
- Hugging Face: microsoft/TRELLIS.2-4B model card
- Hugging Face: TRELLIS.2 demo Space
Image: Microsoft (GitHub)
Written in our own words; quotes are marked and credited. How we report