Open Source Image to 3D Model: How It Works | TRELLIS.2

Two Different Things Share This Name

An "open source image to 3D model" can mean the model itself or a service built around it. Keeping those apart is the whole point of this page, because the licence, the hardware, and the support story belong to different parties.

• The model: TRELLIS.2 is an image-to-3D model published by Microsoft Research. It is described as a 4-billion-parameter flow-matching transformer with an O-Voxel sparse voxel representation, and it produces geometry together with PBR material data. The upstream repository is github.com/microsoft/TRELLIS.2, and that repository is the source of truth for the code, the weights, the licence, and the self-hosting requirements.
• The service: trellis-2.com is an independent, third-party hosted interface around that model. It is not Microsoft, it is not the official project site, and it is not endorsed by Microsoft. Accounts, jobs, credits, the viewer, and support belong to this service. The independent explainer goes through the distinction in more detail.

What Open Weights Do Not Automatically Give You

Open weights are a licensing and access property, not a service level.

• Weights do not include a GPU. Self-hosting needs suitable hardware or rented compute.
• Weights do not include an interface, job queue, viewer, or file conversion.
• Weights do not remove the physics of the problem: one image still cannot reveal hidden surfaces or real dimensions.
• Weights do not promise the same version forever. A local setup pins a revision; a hosted service upgrades on its own schedule, so a local default and a hosted default may not match.

Read the repository licence yourself before a commercial decision. This page will not restate licence terms it has not verified, and the editorial policy explains why product pages here stay close to what can be checked.

Self-Hosting vs This Hosted Service

If you need to modify the model, train on your own data, or keep everything on a private machine, self-hosting is the honest answer. If you need a mesh from a picture today, without a GPU, the hosted route is the direct one. The comparison page covers how open-weight and vendor-managed tools differ in practice, and the ComfyUI setup guide is the place to start if you want the local install path.

How to Try the Model Without Installing Anything

1. Sign in and open the image-to-3D generator.
2. Upload one image. Clear subjects with defined boundaries work best.
3. Pick a resolution. 512³ is the cheapest evaluation level; the credit cost is displayed before you submit. All hosted runs require credits; no new first-task allowance is granted.
4. Wait for the tracked job. A standard run often completes in about one to three minutes, and failed or canceled jobs return their reserved credits.
5. Inspect the GLB in the browser, download it, and convert it to glTF, OBJ, STL, PLY, or VOX if another program needs a different container.

Common Mistakes

1. Calling this website the open-source project. It is an independent service around the model.
2. Assuming open source means unlimited free generation. Hosted generation is a credit product, and self-hosting costs hardware instead.
3. Ignoring the licence. Open weights still carry terms; check them in the repository.
4. Comparing a local build with this service as if they were the same release. Revisions, defaults, and settings can differ.
5. Expecting open weights to fix input limits. Hidden geometry stays estimated either way.

Frequently Asked Questions

Is TRELLIS.2 open source?

Microsoft Research published the TRELLIS.2 model, and its repository is public. The repository is the authority on the exact licence and on what self-hosting requires.

Is trellis-2.com the official open-source project?

No. It is an independent third-party service that runs the model behind its own interface, credits, and support. It is not Microsoft and does not speak for the project.

Can I run an open-source image-to-3D model locally?

Yes, if you can supply the hardware and the setup. That path keeps you in control of the revision and the data, and it moves the maintenance to you.

Do I need a GPU to use this service?

No. Generation runs on a managed provider queue, and the browser handles inspection and format conversion.

Which output does the model produce?

A GLB starting asset with geometry and a PBR material set, including an alpha channel. Other formats come from the free converter or from your own 3D software.

Try the Hosted Route First

If you want to judge the model before committing to hardware, run one 512³ job on one clear image and inspect the result.

Open the Image-to-3D Generator

See what the model is and is not