# TRELLIS — card for AI agents

Microsoft's image-to-3D model that outputs meshes, 3D Gaussians or radiance fields from a single image.

Category: AI 3D model generators · maker Microsoft · checked 2026-11-10

## Price
- Price: Free
- Free tier: Open weights, free to download.

## License
MIT (models and most code); some submodules use other licenses — https://github.com/microsoft/TRELLIS/blob/main/LICENSE

- commercial: **conditions** (Commercial use with conditions)
  > i.e., for research and/or evaluation purposes only.
  (https://github.com/JeffreyXiang/diffoctreerast/blob/master/LICENSE)
- attribution: **no** (No credit needed)
  > The above copyright notice and this permission notice shall be included in all
  (https://github.com/microsoft/TRELLIS/blob/main/LICENSE)
- in app: **conditions** (Ship inside your app with conditions)
  > The following submodules may have different licenses
  (https://github.com/microsoft/TRELLIS)

## Rules to follow
- Model weights are MIT; keep the license notice when you redistribute them.
- The diffoctreerast renderer submodule is research and evaluation only; keep it out of commercial pipelines.
- Check the modified FlexiCubes submodule license before shipping.
- The nvdiffrast (NVIDIA, research or evaluation only) and mip-splatting (research licence) dependencies are not for commercial use.

## How to connect
- Download only
- Repository: https://github.com/microsoft/TRELLIS
- Website: https://huggingface.co/microsoft/TRELLIS-image-large

## Notes
Hugging Face API, Oct 25, 2026: 2,391,045 downloads, 715 likes; GitHub microsoft/TRELLIS: 14,104 stars.

## Numbers (AgentGid data)
- Gid Score: 37
- GitHub stars: 14,350

## Lines to paste into your prompt or AGENTS.md

```
- TRELLIS (AI 3D model generators). Microsoft's image-to-3D model that outputs meshes, 3D Gaussians or radiance fields from a single image. Price: free; commercial use with conditions; no credit needed; ship inside your app with conditions.
  Connect: download from https://huggingface.co/microsoft/TRELLIS-image-large (no API).
  Rule: Model weights are MIT; keep the license notice when you redistribute them.
  Rule: The diffoctreerast renderer submodule is research and evaluation only; keep it out of commercial pipelines.
  Rule: Check the modified FlexiCubes submodule license before shipping.
  Rule: The nvdiffrast (NVIDIA, research or evaluation only) and mip-splatting (research licence) dependencies are not for commercial use.
  License: https://github.com/microsoft/TRELLIS/blob/main/LICENSE (checked 2026-11-10; details: https://agentgid.com/assets/trellis/)
```

## Links
- AgentGid page: https://agentgid.com/assets/trellis/
- Is it free for commercial use? https://agentgid.com/assets/trellis/commercial-use/
- Source: https://huggingface.co/microsoft/TRELLIS-image-large
- Source: https://github.com/microsoft/TRELLIS
- Source: https://github.com/microsoft/TRELLIS/blob/main/LICENSE
- Source: https://github.com/JeffreyXiang/diffoctreerast/blob/master/LICENSE
- Source: https://github.com/nv-tlabs/FlexiCubes/blob/main/LICENSE.txt
- Source: https://huggingface.co/api/models/microsoft/TRELLIS-image-large


## TRELLIS: 3D models from a single picture

TRELLIS is a free AI model from Microsoft that turns one picture of an object into a 3D model. You can save the result as a ready 3D file, a point-cloud style scene file or a spinning preview video. It runs on a Linux computer with a large graphics card, and some of its helper parts are licensed for research only.

### Where to use it
- A 3D object from one photo: Give it a single picture of an object, such as a lamp or a figurine. You get a 3D model with colors as a GLB file for games, 3D viewers or Blender.
- Spinning preview videos: Render a short video that turns around the new object, for a product page, a pitch or a quick review with your team.
- Scenes for 3D splat viewers: Save the object as a cloud of soft 3D points in a PLY file, which splat viewers on the web can show with fine detail.
- Variants of the same object: Make several versions of one object, or change one part of it, for example to try different looks for a game item.
- 3D from a text idea: Start from a sentence instead of a photo. The makers advise making a picture first with an image model, then turning that picture into 3D.

### How to start
- On Linux with an NVIDIA card and the CUDA Toolkit, get the code with its submodules.

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

- Create the conda environment and install the parts you need with the setup script.

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

- Load the image model by its Hugging Face name, or from a local folder you downloaded.

```
TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")
```


### License
The models and most of the code are MIT, so they are free to use and share with the notice. Some helper parts used for rendering and for the 3D file export are for research or evaluation only, so check them before paid use.

### Good to know
- It needs an NVIDIA graphics card with at least 16 GB of memory and was tested only on Linux.
- The GLB export uses NVIDIA's nvdiffrast, whose license allows research or evaluation only. Get NVIDIA's consent or swap it out before selling the results.
- The diffoctreerast renderer that comes with it may not be used for commercial purposes without the licensor's consent.
- The text-to-3D models give plainer results than the picture model, because of limited training data.
- Microsoft has since released TRELLIS.2, a larger 4 billion parameter model that needs a card with 24 GB.

Checked 2026-11-10 against: https://huggingface.co/microsoft/TRELLIS-image-large, https://huggingface.co/microsoft/TRELLIS-image-large/raw/main/README.md, https://huggingface.co/api/models/microsoft/TRELLIS-image-large, https://github.com/microsoft/TRELLIS, https://github.com/microsoft/TRELLIS/blob/main/README.md, https://github.com/microsoft/TRELLIS/blob/main/LICENSE, https://github.com/microsoft/TRELLIS/blob/main/setup.sh, https://github.com/microsoft/TRELLIS/blob/main/trellis/utils/postprocessing_utils.py, https://github.com/JeffreyXiang/diffoctreerast/blob/master/LICENSE, https://github.com/nv-tlabs/FlexiCubes/blob/main/LICENSE.txt, https://github.com/NVlabs/nvdiffrast/blob/main/LICENSE.txt, https://github.com/autonomousvision/mip-splatting/blob/main/LICENSE.md, https://github.com/microsoft/TRELLIS.2/blob/main/README.md

### Mistakes to avoid
- Don't: Cloning the code without --recurse-submodules → Do: git clone --recurse-submodules https://github.com/microsoft/TRELLIS.git (Several renderers come in as submodules and must be compiled; without them the pipeline does not install.)
- Don't: Running on a card without flash-attn support (such as a V100) with default settings → Do: Install xformers and set ATTN_BACKEND=xformers before running (The default attention backend is flash-attn.)
- Don't: Using the text models for best quality → Do: Make a picture with an image model first, then use TRELLIS-image-large (The makers say the text models are less creative and detailed.)
- Don't: Treating the whole install as MIT and shipping GLB export in a paid product → Do: List each installed part's license (nvdiffrast, diffoctreerast, mip-splatting) and get consent or replace the research-only ones (to_glb imports nvdiffrast, and nvdiffrast's license allows non-commercial use only.)
- Don't: Leaving SPCONV_ALGO at auto for a single run → Do: Set os.environ['SPCONV_ALGO'] = 'native' when you run it only once (The auto setting runs a benchmark first, which only pays off over many runs.)

### Questions
- What files do I get? A GLB file with a textured mesh, a PLY file with 3D Gaussians, and preview videos of each form.
- Can I sell the models I make? The TRELLIS models are MIT. The GLB export imports nvdiffrast, which is for research or evaluation only, so check that part or replace it before paid use.
- Does it run on a Mac or Windows? It is tested only on Linux with an NVIDIA card; Windows setups are discussed by the community but not fully tested.
- Can I use several photos of the same object? Yes, the picture model accepts several images, but this works without extra training and may not give the best result for every set.
- Which model should I download? TRELLIS-image-large (1.2B parameters); the makers recommend the picture models over the text ones, and it also holds all the shared parts.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
