What This Review Covers
This page reviews open-weight image-to-3D releases: models whose weights you can download and run on your own hardware. It is not a review of hosted services — for those, use the workflow comparison instead. If you only want to know what Microsoft's release is and how this site relates to it, the Microsoft TRELLIS.2 explainer answers that directly.
Every product fact below comes from the project's own repository or documentation, linked inline. This page does not publish an unrecorded benchmark, does not rank the projects by output quality, and does not borrow numbers from a different release.
The Four Releases Side By Side
| Project | Licence published in the repository | What it ships |
|---|---|---|
| TRELLIS.2 (Microsoft) | MIT | 4B model weights, inference and training code |
| Hunyuan3D 2.1 (Tencent) | Tencent Hunyuan 3D 2.1 Community License | 3.3B shape + 2B texture weights, training code |
| TripoSR (Tripo AI + Stability AI) | MIT | Source code, pretrained model, demo script |
| Stable Fast 3D (Stability AI) | Stability AI Community License | Weights on a gated Hugging Face repo, code, ComfyUI nodes |
TRELLIS.2: The Microsoft Release
Microsoft's official repository describes TRELLIS.2 as a 4-billion-parameter image-to-3D model built on a field-free sparse voxel structure it calls O-Voxel, which the project credits with handling open surfaces (clothing, leaves), non-manifold geometry, and internal enclosed structures without a lossy conversion step. Published output resolutions are 512³, 1024³, and 1536³, with full PBR materials — base color, roughness, metallic, and opacity — and the repository is MIT-licensed.
Setup is the demanding part. The project states the code is currently tested only on Linux, needs an NVIDIA GPU with at least 24 GB of memory, and has been verified on A100 and H100 cards. The documented environment is Python 3.8 or later with PyTorch 2.6.0 on CUDA 12.4, flash-attn as the default attention backend, and xformers as the fallback for GPUs that do not support it. Before installing, read the current README rather than an older tutorial: the earlier Microsoft TRELLIS project is a different release, and its model names, flags, and output modes do not transfer.
The pretrained checkpoint is published as microsoft/TRELLIS.2-4B on Hugging Face with a demo Space, and Microsoft's own inference timings (about 3 s / 17 s / 60 s at the three resolutions) are measured on an H100, so treat them as model-runtime figures rather than end-to-end production latency. Export goes through the project's O-Voxel post-processing into GLB, with parameters for decimation target and texture size; note that exported GLBs are opaque by default and the alpha channel must be connected manually if you need transparency.
Hunyuan3D 2.1: Weights, Training Code, and a Territorial Licence
Tencent's Hunyuan3D 2.1 repository publishes two components: Hunyuan3D-Shape-v2-1 (3.3B parameters, image to shape) and Hunyuan3D-Paint-v2-1 (2B parameters, texture synthesis). The project describes it as the first release of its line to ship full model weights and training code, and its texture pipeline models PBR materials rather than only RGB colour.
The README states memory needs of about 10 GB VRAM for shape generation, 21 GB for texture generation, and 29 GB for both, and documents macOS, Windows, and Linux support with Python 3.10 and PyTorch 2.5.1. Tencent also operates a hosted workspace at 3d.hunyuan.tencent.com, so the same brand appears on both the open-weight and hosted sides of a shortlist.
The licence is the constraint to read before planning a deployment. The Tencent Hunyuan 3D 2.1 Community License excludes the European Union, the United Kingdom, and South Korea from its licensed territory; products above 1 million monthly active users need a separate licence from Tencent; and anyone offering the model as a service must disclose the real provider and state that Tencent is not affiliated with or endorsing it.
TripoSR: The Lightweight Baseline
TripoSR is the official codebase for a feedforward single-image reconstruction model developed by Tripo AI and Stability AI. The repository ships source code, pretrained weights, and an interactive demo under the MIT licence, and documents a default footprint of about 6 GB VRAM for a single input image — the lightest of the four to stand up. The project states that it can reconstruct a model in under 0.5 seconds on an A100; that is the vendor's own published figure on named hardware, not an independent measurement, and it is not the time you will spend getting a usable asset into a destination tool.
Output is a mesh with vertex colours by default, with an optional baked texture pass and a configurable texture resolution. The repository does not document a remeshing step, so if topology matters to your pipeline, plan to retopologise separately.
Stable Fast 3D: TripoSR With UV Unwrapping and Lighting Estimates
Stable Fast 3D (SF3D) is Stability AI's successor in the same family: the README states it is based on TripoSR but explicitly optimised to produce meshes without reconstruction artefacts alongside textures with UV unwrapping, and it predicts material parameters so assets can be dropped into a renderer. It exports GLB, uses roughly the same 6 GB VRAM footprint, and exposes a texture resolution plus remesh options (none, triangle, or quad) with a target vertex count. The repository also ships official ComfyUI nodes.
Two practical differences matter. First, the weights sit behind a gated Hugging Face repository, so access must be requested and a token configured before the first run. Second, the licence is the Stability AI Community License: free for research and non-commercial use, and free for commercial use while the organisation stays below 1,000,000 USD in annual revenue, with registration required for commercial use and a separate enterprise licence above that threshold. The README marks Mac (MPS) and Windows support as experimental, with a CPU fallback.
Hardware Is The Real Filter
Sort the four by what you can actually run:
- Linux, 24 GB+ NVIDIA GPU: TRELLIS.2, the only one of the four whose repository publishes explicit 512³–1536³ output modes.
- Up to 29 GB VRAM on macOS, Windows, or Linux: Hunyuan3D 2.1, with a two-stage shape-plus-texture pipeline.
- About 6 GB VRAM: TripoSR, and Stable Fast 3D, which additionally documents experimental Mac and Windows support and a CPU fallback.
Add the licence to the hardware column: an MIT or gated-community model is not interchangeable with a licence that excludes your territory or charges for scale.
What None Of Them Will Do For You
A single image does not contain the back, the underside, the true dimensions, or the internal structure of an object. Every model here estimates those regions, so inspect the result from several angles before trusting it. None of the four certifies an asset for a specific destination either: topology, UVs, normals, scale, material channels, and wall thickness still need checking in the tool where the asset will live, and a mesh that looks correct in a viewer can still fail a slicer or an engine import.
Generated assets are drafts. Budget for retopology, UV repair, or scaling work, and never use a single-image reconstruction as dimensional evidence for a functional part.
How To Choose
Run the same small test on the model your hardware and licence allow, using your own images, and record failures as well as successes. The comparison protocol covers the accepted-asset calculation and the fields to capture per job.
If you do not want to run any of this locally, the hosted TRELLIS.2 workflow on this site takes one image, tracks the provider task, and returns a GLB — start with a 512 job and inspect the file before spending more credits.










