TRELLIS.2: open-source image to 3D
TRELLIS is Microsoft's open 3D generation model, and TRELLIS.2 is its 4-billion-parameter successor from December 2025. It builds the object in a sparse voxel structure that holds shape and material together, so it handles thin parts, open surfaces and transparent materials better than most. It is MIT licensed, which makes it the most permissive of the strong open models.
| Made by | Microsoft |
|---|---|
| Released | December 2025 (TRELLIS in December 2024) |
| Licence | MIT |
| Commercial use | Yes |
| GPU memory | 24 GB (TRELLIS: 16 GB) |
| Output | Mesh with PBR materials, including transparency |
TRELLIS and TRELLIS.2
The first TRELLIS, from December 2024, introduced a structured latent: a sparse 3D grid where each occupied cell carries features for both shape and appearance. From that one representation it could decode a mesh, Gaussian splats or a radiance field. It became one of the most downloaded open 3D models on Hugging Face.
TRELLIS.2, released in December 2025, scales the generator to 4 billion parameters and replaces the representation with what Microsoft calls O-Voxels. They store geometry and full PBR materials directly, so the output can have open surfaces, non-manifold parts and transparency without a lossy conversion step.
Speed and hardware
Microsoft's model card lists generation time by resolution on its test GPU: about 3 seconds at 512³, 17 seconds at 1024³ and 60 seconds at 1536³. The README asks for an NVIDIA GPU with at least 24 GB of memory, verified on A100 and H100. The original TRELLIS needs 16 GB.
- Try it in a browser: Microsoft's TRELLIS.2 Space on Hugging Face.
- Run it locally: Linux, CUDA, and the setup script in the GitHub repo.
- Licence: MIT, for the code and the weights. Commercial use is allowed.
Where it struggles
Like every single-image model, TRELLIS.2 has to invent the back of the object. Symmetric products come out well; anything with a distinct back, text or a logo is a guess. Textures are generated, not copied from the photo, so colours and small print can drift from the real product. Check the result against the product before you publish it.
No GPU, no setup?
2dto3D runs a production pipeline for product photos: it generates the hidden views first, then builds a textured model at real size, with AR, an embed snippet and export to 7 formats. About 5 minutes per model, from $2, nothing to install.
Your first model is $2, refunded in full within 14 days if you do not like it. Try it with a sample photo
Frequently asked
Is TRELLIS free for commercial use?
Yes. Both TRELLIS and TRELLIS.2 are released under the MIT licence, which allows commercial use of the code, the weights and the generated models.
What GPU do I need for TRELLIS.2?
An NVIDIA GPU with at least 24 GB of memory, per the official README. That means an RTX 4090 or 5090, or a cloud A100 or H100. The original TRELLIS needs 16 GB.
Can TRELLIS make 3D models from text?
The original TRELLIS has text-to-3D checkpoints. TRELLIS.2 is image-to-3D. For text, generate an image first with any image model, then pass the image to TRELLIS.2.
Other open models and datasets
Facts checked 2026-11-06 against the model cards, licence files and READMEs.