TRELLIS adds text-conditioned 3D models.

An archive brief on the release of text models and training code. This is 3D asset generation, not CAD.

3DDownloadable

By Privaj Scout · Brief published Sep 10, 2026 · Sources checked Sep 10, 2026

SCOUT’S TAKE

Direct text input is convenient, but Microsoft recommends generating an image first and using its image-conditioned model for better detail.

See the comparison

Text variants: 342M–2B
Output: 3D assets
Documented VRAM: 16 GB+ NVIDIA
License: MIT

What changed.

The March 25 update added training code and text-conditioned models.
The text family ranges from 342M to 2B parameters.
Outputs can include meshes, radiance fields and 3D Gaussians.

Microsoft repository · dated updates and requirements ↗

SIDE BY SIDE

Text or image input?

Models serve different input workflows; larger parameter counts do not establish better output.

TRELLIS image-large compared with TRELLIS text family
What matters	TRELLIS image-large	TRELLIS text family
Input	Image	Text
Listed parameters	1.2B	342M / 1.1B / 2B
Maker guidance	Recommended for detail	Less detailed in its tests

Before you run it.

The repository specifies Linux and an NVIDIA GPU with at least 16 GB memory. Those are documented requirements, not measurements for every model and output setting.

What the evidence shows

No matched speed test is attached. Do not interpret a generated mesh as an editable, dimensionally constrained CAD design.

Read another briefDeepSeek-V4.1-Flash