TRELLIS
MicrosoftTRELLIS is Microsoft's family of 3D asset generation models built on structured 3D latents. The original models turn a text or image prompt into radiance fields, 3D Gaussians or textured meshes; TRELLIS.2, a 4B model, generates meshes with PBR materials from a single image at voxel resolutions up to 1536 cubed and can texture an existing mesh from a reference image.
Openness
5 high confidence- weights
- open(the TRELLIS image and text models and TRELLIS.2-4B download from Hugging Face without a gate)
- data
- open(TRELLIS-500K is published as metadata over its public source collections with the toolkits to rebuild it, and the TRELLIS.2 card states that the dataset is publicly released)
- code
- open(both repositories publish full training code (train.py and the trainers))
- license
- MIT(OSI
TRELLIS publishes everything needed to reproduce it: MIT-licensed weights, the full training code, and the TRELLIS-500K training set with the tools to rebuild it from its public sources. Only a few third-party rendering submodules sit under other licenses.
- https://huggingface.co/api/models?search=TRELLIS&author=microsoft&sort=downloads&direction=-1&limit=100 recorded 2026-09-26
Hugging Face listing: the five TRELLIS and TRELLIS.2 checkpoints, license mit, public.
- https://huggingface.co/microsoft/TRELLIS.2-4B/raw/main/README.md recorded 2026-09-26
TRELLIS.2 card: "This model is released under the MIT License. The code and dataset are publicly released to facilitate reproduction and further research."
- https://ungh.cc/repos/microsoft/TRELLIS.2/files/main/LICENSE recorded 2026-09-26
TRELLIS.2 LICENSE body: "MIT License / Copyright (c) Microsoft Corporation."
- https://ungh.cc/repos/microsoft/TRELLIS.2/files/main/README.md recorded 2026-09-26
TRELLIS.2 README: "We provide the full training codebase, enabling users to train TRELLIS.2 from scratch or fine-tune it on custom datasets."
- https://ungh.cc/repos/microsoft/TRELLIS/files/main/LICENSE recorded 2026-09-26
LICENSE body: "MIT License / Copyright (c) Microsoft Corporation."
- https://ungh.cc/repos/microsoft/TRELLIS/files/main/README.md recorded 2026-09-26
README: "We provide TRELLIS-500K, a large-scale dataset containing 500K 3D assets curated from Objaverse(XL), ABO, 3D-FUTURE, HSSD, and Toys4k"; "Release training code"; "The training code is centered around train.py".
Adoption
4 high confidenceAdoption is measured as Hugging Face downloads across the five TRELLIS checkpoints, almost all of them for TRELLIS-image-large and TRELLIS.2-4B.
- https://huggingface.co/api/models?search=TRELLIS&author=microsoft&sort=downloads&direction=-1&limit=100 recorded 2026-09-26
Listing sorted by downloads: TRELLIS-image-large 2118600, TRELLIS.2-4B 1724957 in the trailing 30 days.
Capability
5 medium confidenceTRELLIS is among the three highest-rated models on the public image-to-3D arena, closed services included, and the highest-rated with open weights. It places level with Meshy's hosted models there.
- https://3d-arena-3d-arena.hf.space/api/leaderboard?vote_type=render recorded 2026-09-26
Render-board entries "TRELLIS-3DGS" (score 1399, open_source true), "Meshy-6" (1359) and "TRELLIS" (1304).
- https://huggingface.co/spaces/3d-arena/3d-arena/raw/main/src/routes/About.svelte recorded 2026-09-26
3D Arena About page: "The 3D Arena leaderboard evaluates leading generative 3D models"; models take image inputs and produce mesh or splat outputs.
Verified 2026-09-26