3D generation

  ✓ C:\work\chair-hunyuan3d-2mini.glb
  38,912 faces · 1.9 MB · 21.4s total · 14.2s shape · cuda

The first run downloads the model's weights. After that, generation never touches the network. Run arcflare genon its own to see every model and whether it's ready.

Make one the default with arcflare gen default hunyuan3d-2mini-turbo.

These are PyTorch pipelines, not GGUFs, so llama.cpp can't run them. ArcFlare keeps a separate Python environment in ~/.arcflare/gen. arcflare gen setup <model> does four things:

• creates ~/.arcflare/gen/venv from a Python 3.10+ it finds (py -3, python3, python)
• clones the model's code into ~/.arcflare/gen/repos
• installs its requirements, minus torch. A pinned torch would replace your working GPU build with whatever PyPI resolves to
• imports the model's code on your actual device and lists anything missing

Torch: you choose the build

The right torch depends on your GPU vendor, driver and OS. A wrong guess is a multi-GB download that then runs on the CPU. So setup only installs torch when you name the build:

AMD on Windows: borrow a working torch

There is no official ROCm torch for Windows. If another tool on your machine already has a working one, ArcFlare can borrow it:

• The other environment's packages become importable from ArcFlare's venv, read-only, and nothing is ever installed into it.
• The venv is built from the same Python version, because torch is compiled for one Python version and a 3.13 venv can't load a 3.12 torch.
• torchvision has to come from the same place as torch. The regular PyPI torchvision fails next to a ROCm build with operator torchvision::nms does not exist. If the donor environment lacks it, setup says so. Install the matching torchvision into that environment, from wherever its torch came from.

Already have an environment with everything installed? Use it directly, with no venv at all:

Textures

This builds Hunyuan's two texture extensions from CUDA source, so it needs an NVIDIA GPU, the CUDA toolkit and a C++ compiler. Shape generation works without them on any GPU torch supports.

Check it again later

Prints the interpreter, torch version, device, free VRAM and whether the model's code imports.

The image is the model. Image-to-3D reproduces exactly what the picture shows, and every flaw in it becomes geometry.

• One whole object, centred, nothing cropped at the edges.
• A three-quarter view shows more of the shape than a straight-on one.
• Flat, even light. Hard shadows get baked into the shape.
• Plain background. It's removed automatically, but a cut-out PNG with transparency skips that step.
• Iterate with the turbo variant or TripoSR, then regenerate the keeper with full Hunyuan3D at --octree 380.
• Text-to-3D is a picture first: if the mesh is wrong, the generated image usually was too. A photo you picked beats one generated for you.

The machine server exposes generation as three tools, so arcflare agent, Claude Code or any MCP client can make meshes:

A first run downloads weights and a CPU run takes minutes, but an MCP call has a time limit. So long jobs keep running and hand back an id rather than timing out. Jobs stop when the client disconnects, because a GPU job nobody is waiting for still holds VRAM.

❯ make a mesh from ref/teapot.png and render it in Blender

  ⚙ generate_3d ref/teapot.png
  └ mesh ok · hunyuan3d-2mini · 19.8s
  ⚙ blender_run import the glb, add a camera and a light
  ⚙ blender_render scene.blend

Setting HY3DGEN_MODELS or HF_HUB_CACHE yourself points ArcFlare at weights you already downloaded.