mcpbytes-image-to-3d

Turn one image of an object into a textured 3D model (GLB) on your own NVIDIA GPU, from the command line or from any MCP client (Claude, Cursor, VS Code, …). It is a harness around Tencent's Hunyuan3D 2.1, whose code it vendors. The harness cuts the object out, has Hunyuan3D model its shape and paint PBR textures, and then finishes the model the way MCPBytes' hosted create_3d_model_from_image tool does at economy quality: welded, cleaned of loose debris, standing on the ground at unit size, with a compressed web copy and previews from six sides. It can also send jobs to the hosted tool when you have no GPU.

Provided by MCPBytes.com. Tencent is not affiliated with, associated with, sponsoring or endorsing it. Hunyuan3D 2.1's licence does not apply in the European Union, the United Kingdom or South Korea: see Licence.

Example

A real, unedited result of this package (--detail full, seed 42, on an NVIDIA A40): the input image and the finished model from six sides, as written to model_preview.png.

Input	Result

Requirements

• Linux or WSL2 with an NVIDIA GPU of 32 GB or more and a CUDA 12.8 driver (R570 or newer). --detail picks the settings from the GPU's memory; you can override it:

  --detail	Views painted	Faces	Peak GPU memory*	Time per model*
  full (picked for 48 GB cards and larger)	8 at 768 px	120,000	41.2 GB	386 s
  reduced (picked for smaller ones)	6 at 512 px	40,000	28.2 GB	203 s

  *Measured on an NVIDIA A40 (48 GB), including about two minutes of model loading. Your numbers will differ.

• The CUDA 12 toolkit (nvcc) and a C++ compiler (g++). Hunyuan3D has two native extensions, its texture rasterizer and its mesh inpainter; they are compiled for your GPU the first time and cached.

• About 19 GB of disk for the model files, downloaded once, and about 16 GB for the Python packages (PyTorch with CUDA).

• uv (or Python 3.11 or 3.12 with pip).

• No NVIDIA GPU (an Apple Silicon Mac, for example)? Use --cloud with an MCPBytes API key; see below.

Install and run

uvx --python 3.12 --torch-backend=auto mcpbytes-image-to-3d chair.jpg

--python 3.12: Hunyuan3D's libraries need Python 3.11 or 3.12, and uv would otherwise pick its newest Python. --torch-backend=auto installs the CUDA build of PyTorch that matches your driver. With pip (on Python 3.11 or 3.12): pip install mcpbytes-image-to-3d --extra-index-url https://download.pytorch.org/whl/cu128.

The first run downloads the pinned files listed below (about 19 GB, resumable) and compiles the extensions (111 s in all on a RunPod A40; your download speed decides). mcpbytes-image-to-3d --download does both without modelling anything. Results go to chair-3d/ next to where you run it, or to a new or empty folder you name: mcpbytes-image-to-3d chair.jpg models/chair. The command prints a JSON summary of the run on stdout.

Input: a PNG, JPEG or WebP of one object, whole and in focus. A transparent background is used as the cut-out; any other background is removed first. The back and underside are invented, so look at model_preview.png before you use a model.

Options: --seed 0-4294967295 and --detail auto|full|reduced. Seeds are derived the same way as in the hosted tool. The same seed, image, settings and GPU repeated a model byte for byte in our tests.

MCP server

Add this to your MCP client's configuration (for example claude_desktop_config.json, .cursor/mcp.json):

{
  "mcpServers": {
    "image-to-3d": {
      "command": "uvx",
      "args": ["--python", "3.12", "--torch-backend=auto", "mcpbytes-image-to-3d", "--mcp"]
    }
  }
}

Run uvx --python 3.12 --torch-backend=auto mcpbytes-image-to-3d --download once first. A 19 GB download would outlast any client's timeout, so the tool asks for it instead of starting it. The server provides one tool, create_3d_model_from_image, which takes the absolute path of an image, an optional seed and an optional output_dir. It writes the files, returns their summary and shows the six-view preview so the model can look at it. Files go to ~/mcpbytes/image-to-3d/<timestamp>/ unless the call passes output_dir, or you set MCPBYTES_OUTPUT_DIR. Hunyuan3D loads on the first call and stays in memory; --detail applies to the server too.

Cloud mode (optional)

--cloud sends the job to MCPBytes GPUs instead and charges prepaid credits (prices at mcpbytes.com/pricing). Create a key at console.mcpbytes.com/keys:

export MCPBYTES_API_KEY=...
uvx --python 3.12 mcpbytes-image-to-3d chair.jpg --cloud
uvx --python 3.12 mcpbytes-image-to-3d --mcp --cloud     # the MCP server, backed by the hosted tool

The model files are the same; generation.json is the hosted service's, and there is no provenance.json. The run prints a request key: if it is interrupted, run it again with --idempotency-key <key> to resume the same job without a second charge. Nothing is sent to MCPBytes unless you pass --cloud.

Models

Nothing is bundled. These files are downloaded at pinned commits or checksums and licensed to you by their publishers:

• tencent/Hunyuan3D-2.1: the shape and texture models (Tencent Hunyuan 3D 2.1 Community License Agreement)
• facebook/dinov2-giant: the image encoder of the texture model (Apache-2.0)
• Real-ESRGAN x4plus: upscales the painted views (BSD-3-Clause)
• U²-Net, as distributed by rembg: background removal (Apache-2.0)

Hugging Face files go into its cache, so HF_HOME, HF_ENDPOINT, HF_TOKEN and HF_HUB_OFFLINE work as usual. The other two files and the compiled extensions go into ~/.cache/mcpbytes/image-to-3d/ (or $XDG_CACHE_HOME, or MCPBYTES_CACHE_DIR). See NOTICE for attributions.

Development

uv run --with pytest pytest

The tests need no GPU. The vendored code and what we changed in it are described in src/mcpbytes_image_to_3d/vendor/hunyuan3d/README.md. This repository is also the packages/image-to-3d submodule of MCPBytes/tools, the workspace it shares with mcpbytes-core.

Licence

Our code is GNU AGPL v3.0 only, with an additional permission to combine it with Hunyuan3D 2.1. The vendored Hunyuan3D code in src/mcpbytes_image_to_3d/vendor/hunyuan3d/ and the Hunyuan3D weights are Tencent's, under the Tencent Hunyuan 3D 2.1 Community License Agreement, which binds your use of them: it does not apply in the European Union, the United Kingdom or South Korea, it carries an Acceptable Use Policy, and outputs must not be used to improve any other AI model. The models you generate are not covered by the AGPL, and Tencent claims no rights in them. See NOTICE.