Hunyuan3D 2.1 without CUDA (Mac and AMD): 3D generator — what it does and what you need to run it

Hunyuan3D 2.1 without CUDA is a build that lets you run the Tencent Hunyuan 3D 2.1 3D object generation model on computers without NVIDIA GPUs. The author adapted the code to run on Apple Silicon via Metal Performance Shaders and on AMD via ROCm, adding a web interface and PBR texture support.

What it does

The generator creates a 3D model from a single image in PNG or JPG format. The output is a mesh with textures that can be saved in one of seven formats: obj, glb, ply, stl, fbx, dae, or 3mf. As described by the author, the model now outputs PBR textures — that is, maps suitable for physically correct rendering.

The web interface is localized into three languages: English, Chinese, and Russian. The build includes installer scripts and a compatibility layer that replaces CUDA dependencies with MPS or ROCm calls.

What you need to run it

You will need about 30 GB of disk space for the model weights. The description states that the shape weights take roughly 15 GB, the colorization weights about 6.5 GB, the facebook/dinov2-giant model about 4.5 GB, plus the RealESRGAN checkpoint.

Platforms supported as described by the author: Apple Silicon (via MPS) and AMD (via ROCm). Python 3.10—3.12, git, and a working PyTorch build for ROCm are required. The installer script does not overwrite an existing PyTorch installation, since the correct wheel depends on the architecture.

The code is open source, but the license is non-standard — the authors do not specify a name, suggesting you check it manually in the repository. The repository was created on April 18, 2026, the last code change was on August 9, 2026, and the latest release is v0.1.0.

Who it is for

The tool is aimed at those who do not have an NVIDIA GPU with CUDA but do have a Mac on Apple Silicon (M1, M2, M3, M4) or an AMD GPU with ROCm support. It lets you quickly get a draft 3D model from a photo or sketch without renting cloud resources or assembling dependencies manually.

It is not suitable for Windows users with NVIDIA GPUs — for them, the original Tencent project makes more sense. Also, do not expect high speed on integrated Mac graphics or weak AMD cards: 30 GB of weights and heavy computation require modern hardware and patience during download.

Vladimir Talyzin's build removes the CUDA barrier for Mac and AMD owners, packaging the Hunyuan3D 2.1 model into a localized web interface with installers. At the same time, the size of the downloaded weights and the non-standard license are factors worth evaluating before installation.