Vi3W
Skip to content

Guides

Local image-to-3D AI: which models run on your own GPU, and how to run them

Run image-to-3D AI locally: TRELLIS.2, TRELLIS, Hunyuan3D, TripoSG and TripoSR by VRAM and licence, plus ComfyUI, GGUF builds and Blender add-ons.

5 min read

A potted plant model building itself out of small glowing cubes above a graphics card lying on a desk
Illustration. Running locally means the GPU on your desk does the work, and sets the limits.

Quick answer

Can you run image-to-3D AI on your own computer?

Yes, if you have an NVIDIA GPU with enough memory. TRELLIS.2 (MIT) wants 24 GB, TRELLIS v1 16 GB, TripoSG 8 GB and TripoSR about 6 GB. The easiest route in late 2026 is ComfyUI, which made TRELLIS.2 a core node in v0.34. Check each model's licence before using the output commercially.

At a glance

Local image-to-3D models, from each project's own README and licence file, checked November 21, 2026.
ModelLicenceVRAM in the READMEEasiest way to runOutput
TRELLIS.2MIT24 GB+ NVIDIA (community builds aim lower)ComfyUI v0.34+ core nodesPBR GLB
TRELLIS (v1)MIT16 GB+ NVIDIAOfficial repo; community ComfyUI nodes3D mesh
Hunyuan3D 2.1Tencent Hunyuan community licence (not EU, UK, South Korea)10 GB shape, 21 GB texture, 29 GB bothOfficial repo; ComfyUI wrapperTextured mesh
Hunyuan3D 2.0 / 2miniSame licence family6 GB shape, 16 GB with textureOfficial repo (macOS, Windows, Linux)Mesh
TripoSGMIT8 GB+Official repoMesh
TripoSRMITAbout 6 GBOfficial repoMesh
Stable Fast 3DStability Community Licence (free under $1M revenue)About 6 GBOfficial repo; Mac via MPS experimentalTextured mesh

Why run image to 3D locally

Running image-to-3D on your own machine has real advantages. Your images never leave your computer. There are no credits, queues or monthly limits. You can batch a folder of product photos overnight. And you can chain the model into other tools, like a background remover before and a retopology step after.

It also has real costs. You need an NVIDIA GPU with enough VRAM, you'll spend time on setup, and when something breaks, you're the support desk. This guide is about image input specifically. For a broader tour of local 3D generators, see local AI 3D model generator.

Step 1: match the model to your VRAM

VRAM is the deciding number, so start there. Every figure below comes from the project's own README.

  • 24 GB or more: TRELLIS.2, Microsoft's 4B-parameter model. MIT. PBR GLB output with base colour, roughness, metallic and opacity. Officially tested only on Linux.
  • 16 GB: TRELLIS v1, MIT, still maintained in 2026. Or Hunyuan3D 2.0 with texture.
  • 10 GB: Hunyuan3D 2.1 shape only, with --low_vram_mode. Texture needs 21 GB.
  • 8 GB: TripoSG (VAST, MIT, 1.5B).
  • About 6 GB: TripoSR (MIT), Stable Fast 3D, or Hunyuan3D 2.0 shape only.

Two licence notes. Hunyuan3D's community licence isn't valid in the EU, the UK or South Korea. Stable Fast 3D is free under US$1M annual revenue and asks you to register for commercial use. More in open-source 3D AI licences, and on cards in GPU for AI 3D generation.

Meta's SAM 3D Objects is also local and open, but needs Linux and a 32 GB GPU, and outputs a Gaussian splat by default.

Step 2: pick how you'll run it

ComfyUI (easiest for TRELLIS.2)

Since ComfyUI v0.34 (blog post dated August 31, 2026), TRELLIS.2 and Pixal3D are core nodes. They need no custom nodes or compiled CUDA extensions, because post-processing was rewritten in PyTorch and SciPy. They also dropped nvdiffrast and nvdiffrec, which carried non-commercial licences. New nodes include Load, Preview and Save 3D, mesh post-processing, UV unwrap, and normal and AO baking. Official weights are at Comfy-Org/TRELLIS.2 in bf16, plus an int8 diffusion model. The blog gives no VRAM figure. Walkthrough: TRELLIS 2 in ComfyUI.

A desktop computer with a glowing graphics card behind a glass side panel and a small 3D mushroom house floating above it
Illustration. A desktop tower with a big card is still the easiest place to run these models.

Community ComfyUI wrappers

visualbruno/ComfyUI-Trellis2 (MIT) is tested on Windows 11 and ships Windows wheels. PozzettiAndrea/ComfyUI-TRELLIS2 (MIT) is another option. Hunyuan3D has kijai's ComfyUI-Hunyuan3DWrapper.

The official repositories

Each project's GitHub repo has its own install steps. They're the reference, but expect Linux-first instructions and CUDA builds. Step by step: run TRELLIS locally.

GGUF and C++ builds

For smaller cards, quantised TRELLIS.2 builds exist: ilintar/trellis2-gguf for trellis.cpp, a C++/GGML runtime with CUDA and Vulkan, also usable from Lemonade; and Aero-Ex/Trellis2-GGUF for ComfyUI GGUF nodes. They're community projects, and VRAM claims for them are unverified.

Inside Blender

StableGen (GPL-3.0) is a Blender add-on that uses ComfyUI's TRELLIS.2 nodes to make meshes, and SkinTokens to rig them. FishWoWater/trellis_blender is another plugin.

Windows and AMD

Windows works through the ComfyUI routes above. For AMD cards, community ROCm builds exist, such as WakeUpMorty/Trellis2-Rocm-Windows-Installer. Expect more troubleshooting than on NVIDIA.

Step 3: give it a good image

Local models are just as fussy as hosted ones. One object, centred, plain background, even lighting, no heavy shadows. Remove the background first. Tips: best photos for image to 3D.

Step 4: clean up the result

What comes out is a dense mesh. Import the GLB into Blender, check scale and normals, then decimate or retopologise for games. If TRELLIS.2 output looks wrong where it should be transparent, that's expected: it exports in OPAQUE mode by default, keeping alpha in the texture, so you wire it to the material's alpha yourself.

Local or hosted: a quick decision

Go local when:

  • your images are confidential, like unreleased products or client work under NDA
  • you'll generate in volume, and per-model credits would add up
  • you want to chain steps in ComfyUI and keep the whole pipeline on your machine
  • you already own a 16 GB or 24 GB NVIDIA card

Stay hosted when:

  • you're on a Mac or a laptop without a big NVIDIA GPU
  • you want models that aren't open, such as Meshy, Tripo, Rodin or Hunyuan 3D 3.x
  • you need rigging, animation or retopology on the same model without assembling the tools yourself
  • your time is worth more than the setup

Many people do both: they try ideas on a hosted tool, then move the repeatable work to a local pipeline once they know which model suits their objects.

How to compare local models fairly

Run the same image through two models at the same settings. Compare the back of the model, the wireframe, the texture maps and the file size. Note your settings, because resolution changes both quality and VRAM use.

When your GPU isn't enough

If your card is too small, or you're on a Mac, Vi3W runs TRELLIS (the free default), TRELLIS 2, Hunyuan 3D and 20-odd other image-to-3D models in the browser. Then you can rig, retexture or retopologise the result and download it as GLB, FBX, OBJ, STL, PLY or glTF. It's hosted, not local. Try Vi3W.

Frequently asked questions

image to 3d ai local

Pick the model your VRAM allows: TRELLIS.2 on 24 GB, TRELLIS v1 on 16 GB, TripoSG on 8 GB, TripoSR on about 6 GB. Run it through ComfyUI or the project's own repo, then clean the mesh in Blender.

best image to 3d model for comfyui

TRELLIS.2 is the most convenient, because it's been a core ComfyUI node since v0.34 (August 2026), with no custom nodes or compiled extensions, plus UV unwrap and baking nodes. Pixal3D became native at the same time. Hunyuan3D has community wrappers.

image to 3d ai free download

The weights are free to download from GitHub and Hugging Face. For TRELLIS.2 in ComfyUI, the official repackaged weights are at Comfy-Org/TRELLIS.2, including an int8 diffusion model. Read the licence file that comes with whatever you download.

Last reviewed