Hunyuan3D-2-WinPortable

YanWenKunGPL-3.0更新放缓

Windows 上开箱即用的混元 3D 2.0/2.1 整合包,免配置环境即可从图片或文字生成带纹理的 3D 模型。

📦Portable package for running Hunyuan3D 2.0/2.1 on Windows. | 混元 3D 2.0/2.1 整合包

v4-cu129发布于 2025-08-25 01:14 UTC1,373star129fork同步于 2026-11-07 07:32 UTC

怎么上手

由 README 提炼,执行前请确认
下载 cu129 或 cu126 的 .7z.001 与 .002 放同一目录,解压 001 到浅路径如 C:\AI\HY3D2;先运行 UPDATE.bat,再双击 RUN.bat,浏览器打开 http://localhost:8080/
硬件要求按 README 说明

需 NVIDIA 显卡且驱动版本 ≥576.57;几何生成最低约 3 GB 显存,贴图生成最低约 6 GB 显存,系统内存 ≥24 GB,显存越少越依赖内存卸载。Blackwell 架构需用 CUDA 12.9 包,Volta/Pascal/Maxwell 只能用 CUDA 12.6 包。

需要 CUDA
下载:源码 zip·GitHub Releases(安装包)本站不镜像安装包

项目说明

由 README 自动翻译提炼

在 Windows 上免配置运行混元 3D 2.0/2.1 官方 Gradio 演示的整合包

Windows 上开箱即用的混元 3D 2.0/2.1 整合包,免配置环境即可从图片或文字生成带纹理的 3D 模型。

核心能力

  • 集成 mmgp 显存优化,让消费级显卡也能运行
  • 提供 CUDA 12.9 与 12.6 两套整合包,按显卡架构选择
  • 内置启动器,可勾选启用贴图生成并自动编译依赖
  • 支持 2.0、2.0 官方原版、2.1 及 API 多种程序模式
  • 模型文件按需自动下载,生成结果存于各自缓存目录

容易踩的坑

解压路径过深会触发 260 字符 MAX_PATH 限制报错,务必用浅目录;贴图生成需另装 CUDA Toolkit 12.9.1(仅选 Libraries and Compilers)和 VS Build Tools 2022 的 C++ 桌面开发负载,否则只能出白模;首次启动会按需下载模型并可能编译依赖,耗时较长;建议开启开发者模式以符号链接避免模型目录重复膨胀。

内容由模型根据 README 生成,可能不完整或不准确,以官方文档为准。看原文 README

最近发版

5 个版本v4-cu129
  1. v4-cu1292025/08/25 CUDA 12.9最新正式版2025-08-25 · 2025-08-25 01:14 UTC

    README | 中文说明

    A major update includes:

    1. Hunyuan3D-2.1;
    2. Full support for Blackwell (50 series) GPUs;
    3. A new GUI launcher.

    Tip

    • If you're upgrading from the old version, copy/move your HuggingFaceHub folder to the new one.

    Versions

    • Python 3.12.11
    • CUDA 12.9
    • PyTorch 2.8.0
  2. v4-cu1262025/08/25 CUDA 12.62025-08-25 · 2025-08-25 01:12 UTC
    查看更新说明

    README | 中文说明

    This package supports Hopper, Ada Lovelace, Ampere, Turing, Volta, Pascal, and Maxwell GPUs.

    Versions

    • Python 3.12.11
    • CUDA 12.6
    • PyTorch 2.8.0
  3. v3.32025/07/12pre-release2025-07-16 · 2025-07-16 19:19 UTC
    查看更新说明

    A pre-release with latest PyTorch 2.7.1 + CUDA 12.8 and xFormers 0.0.31.post1 installed. Also with triton installed.

    You need a GPU with at least Turing architecture (RTX 20 / GTX 16 series) to use this package. For Volta, Pascal and Maxwell GPUs, please use the v2 package instead.

    Notes

    1. You need to download 2 files (.7z.001 and .7z.002), then open the 001 file and extract it. You don't need to open the 002 file again.

    2. If you're upgrading from the previous version, copy/move your HuggingFaceHub folder to the new one.

    3. You may want to install the latest CUDA Toolkit (12.9.1) if possible. As a reminder, it's recommended to uninstall the older versions of CUDA Toolkit first, and only Libraries and Compilers are needed to install

    Versions

    • Python 3.12.11
    • CUDA 12.8
    • PyTorch 2.7.1
  4. v3.22025/05/29pre-release2025-05-28 · 2025-05-28 21:21 UTC
    查看更新说明

    A pre-release for Blackwell (RTX 50 series) and Ada Lovelace (RTX 40 series) GPUs with triton installed.

    The usage is the same as the current README.

    Notes

    1. You need to download 2 files (.7z.001 and .7z.002), then open the 001 file and extract it. You don't need to open the 002 file again.

    2. If you're upgrading from the previous version, copy/move your HuggingFaceHub folder to the new one.

    3. You may want to install the latest CUDA Toolkit (12.9) if possible. As a reminder, it's recommended to uninstall the older versions of CUDA Toolkit first, and only Libraries and Compilers are needed to install

    Versions

    • Python 3.12.10
    • CUDA 12.8
    • PyTorch 2.7.0
  5. v3.12025/05/02pre-release2025-05-02 · 2025-05-02 11:59 UTC
    查看更新说明

    A pre-release for Blackwell (RTX 50 series) GPUs.

    The usage is the same as the current README.

    Notes

    1. You need to download 2 files (.7z.001 and .7z.002), then open the 001 file and extract it. You don't need to open the 002 file again.

    2. If you're upgrading from the old version, copy/move your HuggingFaceHub folder to the new one.

    3. For Blackwell users, you may want to install the latest CUDA Toolkit (12.9) if possible. As a reminder, it's recommended to uninstall the older versions of CUDA Toolkit first, and only Libraries and Compilers are needed to install

    Known issue(s)

    • When running in MV mode, the web page will hide "MultiView Prompt" and "MultiView to 3D Gallery" by default. You need to click to show them first. This could be a Gradio compatibility issue.

    Versions

    • Python 3.12.10
    • CUDA 12.8
    • PyTorch 2.7.0

用户点评

登录用户的真实使用记录,可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译

    [.text-center] English | link:README.zh.adoc[中文说明]

    Hunyuan3D 2.0/2.1 Windows Portable Pack

    image::docs/hy3d2-mv.webp["Screenshot"]

    [cols="3*", frame=none, grid=none] |=== a|image::docs/hy3d2-turbo-tex.webp[] a|image::docs/hy3d21-shape.webp[] a|image::docs/hy3d21-tex.webp[] |===

    A dedicated package for running https://github.com/Tencent-Hunyuan/Hunyuan3D-2[Hunyuan3D 2.0] and https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1[Hunyuan3D 2.1] Gradio Web App (official demo) on Windows.

    Integrated https://github.com/deepbeepmeep/mmgp[mmgp] from https://github.com/deepbeepmeep/Hunyuan3D-2GP[Hunyuan3D-2GP] made by https://github.com/deepbeepmeep[DeepBeepMeep], allowing the models to run on consumer-grade GPUs.

    Instructions - Preparation

    1. Basic Requirements

    ** NVIDIA GPU with driver version ≥576.57 (i.e., after June 2025). ** ≥ 3 GB VRAM for geometry generation (with maximum optimization for 2.1 model) ** ≥ 6 GB VRAM for texture generation (with maximum optimization for 2.0 model) ** ≥ 24 GB System RAM (the less VRAM you have, the more RAM is needed for model offloading)

    2. Download the Pack Based on Your GPU Architecture

    [cols="1,1,1,1,1,1,1,1", options="header"] |=== | GPU Architecture | Blackwell | Ada Lovelace | Ampere | Turing | Volta | Pascal | Maxwell

    | Example GPU | RTX 5090 | RTX 4090 | RTX 3090 | RTX 2080 | TITAN V | GTX 1080 | GTX 980

    | CUDA 12.9 | ✔️ | ✔️ | ✔️ | ✔️ | ❌ | ❌ | ❌

    | CUDA 12.6 | ❌ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ | ✔️ |===

    __ Note: This CUDA compatibility limitation comes from https://github.com/pytorch/pytorch/releases/tag/v2.8.0[PyTorch] and is not related to NVIDIA CUDA Toolkit. __

    3. (Optional) Enable Developer Mode

    image::docs/developer-mode-dialog.en.webp["Developer Mode"]

    Search for "Developer settings" in the Start Menu, or go to Settings -> System -> For developers, and enable "Developer Mode".

    This mode enables the file system's symbolic link (symlink) feature, which prevents the Hugging Face Hub from repeatedly downloading model files and causing the model storage directory to become excessively large.

    4. (Optional) Install Tools for Texture Generation

    If you do not install the following tools, the program can still generate mesh ("shape" or "white model"), but it will not be able to generate textures.

    . Install https://developer.nvidia.com/cuda-12-9-1-download-archive?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_network[CUDA Toolkit 12.9.1]

    ** During installation, you only need to select https://github.com/YanWenKun/Comfy3D-WinPortable/raw/refs/heads/main/docs/cuda-toolkit-install-selection.webp[Libraries and Compilers].

    ** Other versions (e.g., https://developer.nvidia.com/cuda-13-0-0-download-archive[13.0] ) should theoretically work fine, and there is no need to reinstall if you already have one.

    . Install Visual Studio Build Tools 2022 ( https://visualstudio.microsoft.com/visual-cpp-build-tools/[Download page] | https://aka.ms/vs/17/release/vs_BuildTools.exe[Installer] )

    ** In the installer, select the "Desktop development with C++" workload.

    Instructions - Launcher

    image::docs/launcher.en.png["Launcher"]

    . It is recommended to run UPDATE.bat first.

    . Double-click RUN.bat to run the launcher.

    . If you check "Enable Texture Generation", the necessary dependencies for this feature will be automatically compiled and installed on the first launch, which may take longer.

    . Model files will be downloaded on demand when the program starts.

    . After the message running on http://0.0.0.0:8080 appears, open http://localhost:8080/ in your browser.

    . You can force quit the program by closing the command line window.

    Instructions - Program Selection

    . Hunyuan3D 2.0: Corresponds to the Hunyuan3D-2 folder. Generated files are in its gradio_cache subdirectory.

    . Hunyuan3D 2.0 Official Code: Corresponds to the Hunyuan3D-2-vanilla folder. Generated files are in its gradio_cache subdirectory. ** Does not use mmgp optimizations. It requires more VRAM but generates faster. ** Available model references: ** https://huggingface.co/tencent/Hunyuan3D-2/tree/main ** https://huggingface.co/tencent/Hunyuan3D-2mini/tree/main ** https://huggingface.co/tencent/Hunyuan3D-2mv/tree/main ** Third-party fine-tuned models should theoretically work but have not been tested.

    . Hunyuan3D 2.1: Corresponds to the Hunyuan3D-2.1 folder. Generated files are in its save_dir subdirectory.

    . API 2.0: Corresponds to the Hunyuan3D-2 folder, includes partial mmgp VRAM optimizations.

    . API 2.1: Corresponds to the Hunyuan3D-2.1 folder, does not include mmgp VRAM optimizations. ** Available model references: ** https://huggingface.co/tencent/Hunyuan3D-2.1/tree/main

    Usage Tips

    . To view 3D files like glb and obj, you can use https://github.com/cnr-isti-vclab/meshlab/releases[MeshLab].

    . After checking "Use Turbo Model", you can still select Fast or Standard mode in the web UI to improve generation quality.

    . The "Remove Background" feature uses rembg with its default settings, which might leave faint white edges. If your image already has a clean (alpha/transparent) background, there is no need to check this option.

    . "Developer Mode" can be turned off at any time.

    Advanced Information

    . The "Text-to-3D" feature uses the https://huggingface.co/Tencent-Hunyuan/HunyuanDiT-v1.1-Diffusers-Distilled[HunyuanDiT-v1.1-Diffusers-Distilled] text-to-image model.

    . For the specific meaning of each mmgp optimization level, please refer to the https://github.com/deepbeepmeep/mmgp#usage[mmgp documentation] .

    . For process control and batch generation, consider using https://www.comfy.org[ComfyUI]. Refer to its https://docs.comfy.org/tutorials/3d/hunyuan3D-2[example workflow], and be aware of its functional limitations.

    . For more advanced usage, it is recommended to write Python scripts with the help of AI and run them using this portable pack. The entire generation process has many adjustable parameters, and scripts can offer a more streamlined and flexible approach.

    . This pack does not include bpy (Headless Blender) because the packaging CI could not install it in its environment. Its main purpose in the code is to convert 3D file formats, but its absence has not affected the software's usability.

    Related Repositories

    The following code repositories were bundled in this pack:

    [cols="2"] |=== |Official 2.0 Repository |https://github.com/Tencent-Hunyuan/Hunyuan3D-2

    |2.0 with mmgp integration |https://github.com/YanWenKun/Hunyuan3D-2

    |2.1 with mmgp integration and Windows adaptation |https://github.com/YanWenKun/Hunyuan3D-2.1

    |Scripts for legacy versions |https://github.com/YanWenKun/Hunyuan3D-2-WinPortable-Scripts

    |Scripts for the latest version |https://github.com/YanWenKun/Hunyuan3D-2-WinPortable-Scripts-GUI |===

    Acknowledgements

    More

    Explore my other projects: