Guides
What is TRELLIS? Microsoft's open-source image and text to 3D model, explained
TRELLIS turns images and text into 3D assets. What it outputs, its MIT licence, the 16 GB GPU it needs, and how to use it without installing it.
3 min read
Quick answer
What is TRELLIS?
It is an open-source 3D generation model from Microsoft Research that turns an image or a text prompt into a 3D asset. It is released under the MIT licence, can output Gaussian splats, radiance fields or meshes, and exports textured GLB files. Running it yourself needs an NVIDIA GPU with at least 16 GB of memory.
The short version
TRELLIS comes from the paper Structured 3D Latents for Scalable and Versatile 3D Generation (Xiang et al.), a CVPR 2025 spotlight, with code on Microsoft's GitHub. Its idea is to generate one shared 3D representation first, a structured latent, and then decode it into whichever output you want.
That is why one model can hand you three different kinds of 3D:
- 3D Gaussians, saved as PLY: soft, photoreal, but not a mesh. Gaussian splatting vs meshes explains the difference.
- Radiance fields, for rendering.
- Meshes, which can be baked into a textured GLB with the repo's post-processing tools. This is the output that matters for games, Blender and printing.
What you need to run it yourself
From the repository's own README:
- An NVIDIA GPU with at least 16 GB of memory. It was tested on A100 and A6000 cards.
- Linux. Other systems are not tested.
- Some patience with CUDA dependencies.
Most gaming laptops fall short of that. The code and models are MIT licensed (a few submodules have their own licences), which is unusually permissive for a 3D model and one reason it is widely built on.
Ways to use TRELLIS without a 16 GB GPU
- A Hugging Face Space. There are public demos you can try in the browser. Free GPU time there is limited per day, and queues are common. Image to 3D on Hugging Face covers the catch.
- A hosted API, such as Replicate, billed per run.
- An app that runs it for you. TRELLIS is the default 3D model in Vi3W: upload or generate an image, and the 3D step runs TRELLIS for 20 XP (Fast or Balanced) or 30 XP (High Quality), on every plan including the free one. Vi3W then retopologises the result and exports GLB, FBX or OBJ.
What TRELLIS is good and less good at
It works best on a single, whole object on a clean background, like a prop, a piece of furniture or a stylised character. Like every single-image model, it has to guess the back. Thin parts (straps, antennas, fingers) and see-through materials come out less reliably than solid shapes.
Its raw mesh is a dense triangle mesh, which is normal for generators. If you plan to animate or edit the model, you'll want retopology, either your own or the generator's.
TRELLIS, Hunyuan 3D and Tripo
They're often named together. Hunyuan 3D is Tencent's family of open 3D models, with its own licence terms, and Tripo is VAST's commercial model. They differ in licence, cost and the kinds of objects they handle best. How to choose a 3D model in Vi3W puts all of them in one table with what each costs to run.
Try it
Open Vi3W, upload an image, and the first 3D model runs on TRELLIS. No GPU required.
Last reviewed