Comparison · Generators
Hunyuan3D vs TRELLIS.2: License, VRAM and Fit
Hunyuan3D 2.1 vs TRELLIS.2 for self-hosting: territory and user-cap license terms, documented VRAM needs, inputs and outputs, and what speed claims can and cannot show.
The deciding axis is where you are allowed to run each model and what hardware it needs. TRELLIS.2 is MIT-licensed and asks for an NVIDIA GPU with at least 24 GB of memory; Hunyuan3D 2.1 can run in smaller stages (10 GB for shape alone) but its Community License excludes the European Union, the United Kingdom and South Korea. Which one makes better meshes is not settled by anything we can cite.
What would change this answer
If you are inside the excluded territories, or expect over one million monthly active users, the license outweighs everything else and TRELLIS.2’s MIT license is the more open one, subject to the non-commercial limits on its nvdiffrast and nvdiffrec dependencies. If you have a 12 to 16 GB card, TRELLIS.2’s stated minimum rules it out as documented, and only Hunyuan’s shape stage (10 GB) has a published path on that hardware; its texture stage (21 GB) and full pipeline (29 GB) do not, though the older Hunyuan3D 2.0 README documents 16 GB for shape and texture together (its license was not compared here). A same-GPU run with inspectable outputs would change the quality discussion; it is planned in the open-source local run and has not been done.
Documented differences
| Item | Hunyuan3D 2.1 (Tencent) | TRELLIS.2 (Microsoft) |
|---|---|---|
| Weights and code license | Tencent Hunyuan 3D 2.1 Community License (GitHub shows "Other") | MIT for model and code; dependencies nvdiffrast and nvdiffrec "governed by [their] own License" |
| Territory | Worldwide excluding the EU, UK and South Korea; use outside the territory "is unlicensed and unauthorized" | No territory restriction in the MIT text |
| Scale cap | More than 1 million monthly active users in the calendar month before the version release date requires requesting a license from Tencent | None stated |
| Output rights | Section 6(d): "Tencent claims no rights in Outputs You generate." Section 5(b) bars using outputs to improve another AI model | MIT text does not address outputs separately |
| Documented VRAM | 10 GB shape, 21 GB texture, 29 GB both; a --low_vram_mode flag exists | "At least 24GB of memory is necessary"; verified on A100 and H100 |
| Operating system | README: "Hunyuan3D 2.1 supports Macos, Windows, Linux." Tested with Python 3.10, PyTorch 2.5.1+cu124. The texture-stage build runs bash compile_mesh_painter.sh, so on native Windows you need a bash-capable shell | "Currently tested only on Linux"; GPUs without flash-attn support (the README names V100) need xformers installed manually and ATTN_BACKEND set to xformers |
| Inputs | Image to 3D shape plus PBR texture pipeline; tags also list text-to-3D | Single image to 3D; PBR texturing for a given shape (example_texturing.py) |
| Output files | Untextured mesh from shape pipeline; textured PBR mesh from paint pipeline; formats not specified in the 2.1 README. Parameters: 3.3B shape model, 2B paint model | sample.glb, described as PBR-ready, exported opaque by default (alpha is kept in the texture but you connect it manually); README example exports at texture_size=4096 |
| Stated limitations | None quoted in the README sections we read | Model card: raw meshes "may occasionally contain small holes or minor topological discontinuities" |
| Latest open weights | 2.1 (June 2025) and Omni; hosted 3.0 and 3.1 exist but no open 3.x weights were found | TRELLIS.2-4B, 4B parameters (Hugging Face, December 2025) |
License: the sharper difference
The license is the clearest, checkable gap. Tencent’s Community License is territorial and carries a user-count threshold, while Microsoft’s is a plain MIT license with two dependency caveats. The territory wording is explicit enough to quote: the grant covers “the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea.” Anyone distributing a game or tool from, or to, those regions should read the full text and, if it matters commercially, ask a lawyer; this is not legal advice. Our license-by-license notes sit in open-weights licenses for Hunyuan3D and TRELLIS.
The MIT license also does not say the dependencies are free for commercial use. nvdiffrast and nvdiffrec have their own terms, and both limit use to non-commercial research or evaluation; see the license page. Check them before using the pipeline for paid work.
Hardware
Here the documents point opposite ways. Tencent’s README gives staged numbers (10, 21 and 29 GB) and a low-VRAM flag; Microsoft gives one minimum of 24 GB and a Linux-only note. That means a 24 GB card is the documented floor for TRELLIS.2, and unsupported configurations may work but are not documented. We did not find an official statement on consumer cards below 24 GB for either model.
GPU by GPU: what the documents support
| Your GPU memory | Documented path | Source |
|---|---|---|
| About 10 to 20 GB | Hunyuan3D 2.1 shape stage only (10 GB), then texture elsewhere. Hunyuan3D 2.0 documents 6 GB for shape and 16 GB for both. TRELLIS.2: below its stated minimum | Hunyuan READMEs; TRELLIS.2 README |
| 21 to 28 GB | Hunyuan 2.1 shape and texture run separately (10 and 21 GB), not together (29 GB). TRELLIS.2 meets its 24 GB minimum from 24 GB up | Both READMEs |
| 29 GB and up | Hunyuan 2.1 shape plus texture in one run; TRELLIS.2 | Both READMEs |
The 21 to 28 GB row is an inference from the stage figures, not a statement either README makes; whether the stages can be run in sequence on one card without unloading is not documented. Neither README says anything about cards between 10 and 24 GB beyond the figures above, and the Hunyuan README shows the --low_vram_mode flag without saying what it trades away.
Setting up each: what the READMEs make you do
Hunyuan 2.1’s install is a Python 3.10 environment with PyTorch 2.5.1 (CUDA 12.4 wheels), pip install -r requirements.txt, a pip install -e . of the custom rasterizer, a bash-script build of the mesh painter, and a Real-ESRGAN checkpoint download. TRELLIS.2 recommends conda and a setup.sh whose example command installs flash-attn, nvdiffrast, nvdiffrec, cumesh, o-voxel and flexgemm. Hunyuan’s texture stage is a separate model that runs on a shape, and TRELLIS.2’s README points to example_texturing.py for generating PBR textures for a given 3D shape (its roadmap also lists shape-conditioned texturing code, and we could not tell whether that item is done). Whether either accepts a mesh you cleaned up or modeled elsewhere is not shown in what we read, and neither README shows results on non-generated meshes.
Speed and quality claims
TRELLIS.2’s own README gives times on an H100: about 3 seconds at 512 resolution, 17 at 1024 and 60 at 1536 (shape and material stages listed separately). Hunyuan’s README, in the sections we read, gives no timing. Speed numbers circulating online beyond that are not comparable. The site trellis2.app, which sells a hosted TRELLIS 2 service from $19 a month and discloses that, publishes about 3 seconds at 512 resolution and 60 seconds at 1536 on an H100, against 10 to 30 seconds for Hunyuan reported by the vendor on unspecified hardware. That site itself says these figures are not directly comparable, and that its quality statements come from official showcases and community feedback rather than controlled tests. It says TRELLIS 2 resolves harder-surface detail more sharply and that community feedback often credits Hunyuan 2.x with steadier structure on character-like subjects. Treat both as seller and community impressions, not measurements. For a hosted-queue view, 3D AI Studio lists both models and has its own timing and credit claims, but it is a credit seller and we have not verified its numbers.
When to choose each
Choose TRELLIS.2 when you are on a 24 GB-class Linux machine, want MIT terms and need a single image-to-GLB pipeline. Its documented weakness is the occasional holes in raw meshes, so plan for cleanup.
Choose Hunyuan3D 2.1 when you are outside the excluded territories, want the staged pipeline, and either have a GPU with the 29 GB the README gives for shape plus texture or only need the 10 GB shape stage. Its Hugging Face tags list text-to-3D, but the README documents image input. Its weakness is the territory and user-cap terms, and the full pipeline needs more memory than TRELLIS.2, not less.
If you would rather pay for a hosted route, see the comparison hub or Meshy alternatives. Reviews: Hunyuan3D, TRELLIS.2. The wider field is in open-source AI 3D generators; our approach is in the methodology.
Still unknown
- Measured peak VRAM and time for either model on a consumer GPU.
- Whether TRELLIS.2 runs on Windows (its README says Linux only) and whether Hunyuan’s Windows claim holds for the texture-stage build.
- How the non-commercial limits on nvdiffrast and nvdiffrec apply to assets generated with them.
- Whether an open Hunyuan 3.x release exists.