TRELLIS.2 Review: MIT License, Hardware Needs and Output

TRELLIS and TRELLIS.2 are different projects

People searching “trellis 3d” can land on either. The original TRELLIS (v1) also uses the MIT license, accepts text or image prompts, and needs an NVIDIA GPU with at least 16 GB, verified on A100 and A6000, Linux only. TRELLIS.2, with a 4B-parameter model published on Hugging Face in December 2025, is described as built for high-fidelity image-to-3D and for generating PBR textures for a given 3D shape. The TRELLIS.2 README and card describe single-image input; we did not see text-prompt input documented, so for text-to-3D you would generate an image first, which the v1 README itself recommends.

License: MIT, with two caveats

The README says the model and code are under the MIT License, and the LICENSE file is “MIT License, Copyright (c) Microsoft Corporation.” The Hugging Face tag is license:mit. MIT contains no territory limit, user-count threshold or restriction on output, which separates it from the Hunyuan3D community license. The caveats:

• The dependencies nvdiffrast and nvdiffrec are each “governed by its own License.” Both carry non-commercial limits: nvdiffrast’s license says it “only may be used or intended for use non-commercially”, meaning research or evaluation, and nvdiffrec’s has the same limit. Whether running them to make assets for sale is covered is a question for a lawyer; the license page has the text.
• The MIT license covers the model and code; it does not say what training data the model saw or what rights exist in the images you feed it. If you feed it copyrighted concept art, that is your question to answer, not MIT’s.

For game assets generated locally, that leaves you with no vendor account, no plan to cancel, and no free-tier attribution; the exposure is copyright in your inputs and the unsettled legal position of AI output, covered in Hunyuan3D and TRELLIS license terms. This is not legal advice. Third-party hosted TRELLIS services have their own terms, which the MIT license does not govern. 3D AI Studio lists TRELLIS 2 on its model menu, but its per-generation cost for that model was not stated in the pages we read; see the 3D AI Studio review.

Hardware: the real gate

From the README and card: at least 24 GB of GPU memory, verified on A100 and H100, currently tested only on Linux, CUDA Toolkit 12.4 recommended, and default PyTorch 2.6.0 with CUDA 12.4. In practice that means a card with at least 24 GB, or a rented cloud GPU. It does not say the model will fail on less; it says what is necessary and tested, and behavior below 24 GB or on Windows is not documented officially. Our planned local run test targets exactly that gap, and no number for consumer-card speed or measured VRAM exists on this site yet. Install is by setup.sh into a conda environment named trellis2; example.py runs it, python app.py launches the demo, and example_texturing.py or app_texturing.py handles texturing. No ComfyUI integration is mentioned in the README or card.

Resolution, speed and the memory cost of each

The repository documents three resolution settings and H100 timings: 512 cubed in about 3 seconds (2 shape, 1 material), 1024 cubed in about 17 seconds (10 and 7), and 1536 cubed in about 60 seconds (35 and 25). Those are vendor numbers on an H100, a data-center card, so they say nothing about a 24 GB consumer card. The README also says the default attention backend is flash-attn, that GPUs without support (it names V100) need xformers and the ATTN_BACKEND environment variable, and that installation “can take a long time and may need to be done one dependency at a time.” The GLB export example uses a 4096 texture and a decimation_target of 1,000,000 faces, which is far above a game-ready budget.

What users report below the 24 GB line

The README says 24 GB is necessary; the issue tracker shows what happens near that line. These are individual reports, not Microsoft statements, and none had a visible maintainer reply when read.

• Issue #188, “OOM on smaller GPUs” (opened Sep 7, 2026): the reporter ran TRELLIS.2-4B with the default 1024_cascade pipeline on an NVIDIA A10G with 23 GB and hit out-of-memory errors in the shape decoder or in Mesh.fill_holes(). Their analysis was that sampling tensors stay resident and that the raw decoded mesh can exceed 10 million vertices at 1024 resolution. Freeing those tensors and emptying the CUDA cache before fill_holes() cut peak use from about 21 GB to about 11 GB with unchanged output, by their report.
• Community reply on that issue (Oct 14, 2026): another user reported completing the pipeline on a 16 GB RTX 5070 Ti by running late decoder stages in spatial tiles and changing some export steps, and linked patches in a personal repository. They reported a single tile matches untiled output exactly and smaller tiles differ slightly (maximum feature difference 0.016 at 1024 with 128-voxel tiles). These are unofficial patches we have not run.
• Other recent titles: #169 (Jun 11, 2026) scattered surface holes on non-CUDA platforms; #162 (May 20) black and reflective spots on light images; #161 and #160 (May) failures loading the DINOv3 extractor and a missing sparse_structure_decoder; #154 (Apr 25) a cuDNN initialization error. The list showed 131 open issues. We read titles for these, not the threads.

What it means for a job: if you have a 24 GB card, one reporter’s 23 GB A10G ran out of memory at default settings, which suggests headroom is thin, and a 16 GB card has a community route that is not part of the release. If you need Windows, none of the recent titles we saw mention it, which is an absence, not support.

Output and what survives the next step

The documented output is sample.glb, described as a “PBR-ready 3D asset in GLB format,” exported OPAQUE by default, plus a preview video. The model card is candid about limits: raw meshes “may occasionally contain small holes or minor topological discontinuities,” and the base model “has not been aligned with human preferences.” That is the limitation not in the headline. The model card adds that strictly watertight output, for example for printing, needs the provided post-processing scripts, and the README notes that the alpha channel in GLB export is stored in the texture but not active until you connect it manually in your 3D software. For a resin miniature, that means holes are expected and a watertight mesh is a separate step. Expect to run a repair pass; see repairing holes and floaters in AI meshes. Default opaque export would also matter for any asset that needs transparency, such as foliage or glass. We have no independent quality evidence and do not rely on competitors’ showcase claims.

Who should use it and who should skip it

• Use if: you own a 24 GB NVIDIA card or will rent one, you want an MIT license, and you can tolerate Linux and a repair pass.
• Skip if: you have a laptop GPU, run Windows only, or need a rigged, engine-ready character from a single click.
• Consider alternatives if: you are under 24 GB. Hunyuan3D 2.1 lists 10 GB for shape only (29 GB for shape plus texture), with territory limits.

Alternatives

The Hunyuan3D vs TRELLIS.2 comparison addresses license against VRAM; the open-source generators list shows the wider set. Our methodology lists the planned checks; none has been run.

More generators are listed in the AI 3D generator reviews hub.