Skip to content

Free, open-source 3D sculpting for Linux, macOS and Windows: dense meshes, remeshing, painting, built-in AI (image/text to 3D), Send to Unreal, game-ready LODs, Gaussian splats and OpenUSD.

5.9 KB
Raw Blame

ComfyUI workflows for SculptForge

Ready-made ComfyUI workflows that SculptForge's Generate 3D window can run on your own ComfyUI server (image or text in, mesh out). They are saved in ComfyUI's API format, so you can also queue them yourself with POST /prompt, or drag one onto the ComfyUI canvas to look at and change it.

Workflow Input What it does Needs Runs on
sculptforge-inflate.json image Inflates a sketch or silhouette into a rounded, closed base mesh with SculptForge, voxel-remeshes it into an even, watertight sculpting surface (removing the crease where the front and back halves meet), then exports a GLB. The ComfyUI-SculptForge node pack and SculptForge installed. No models. Any CPU, about a second
hunyuan3d-image-to-3d.json image Hunyuan3D 2.1 image-to-3D with ComfyUI's built-in nodes: a detailed, untextured mesh saved as GLB. hunyuan_3d_v2.1.safetensors (7.4 GB). A GPU (Tencent quotes about 10 GB of VRAM for shape generation)
text-to-3d.json text FLUX.1 [schnell] draws the object on a plain backdrop, BiRefNet cuts it out onto white, Hunyuan3D 2.1 turns it into a mesh. Also saves the concept image. flux1-schnell-fp8.safetensors (17.2 GB), birefnet.safetensors (444 MB), hunyuan_3d_v2.1.safetensors (7.4 GB). A GPU; ComfyUI moves each model in and out of VRAM as it is needed

Apart from the SculptForge nodes in the first one, all three use only nodes that ship with ComfyUI. They were checked against ComfyUI 0.42.0; older versions may lack Remove Background or Trim Text, so update ComfyUI if a node is reported missing.

Model files

Download each file into the folder shown, inside your ComfyUI install, then restart ComfyUI or press R in the editor to refresh the lists.

File Folder Used by Download
hunyuan_3d_v2.1.safetensors models/checkpoints/ hunyuan3d-image-to-3d, text-to-3d Comfy-Org/hunyuan3D_2.1_repackaged
flux1-schnell-fp8.safetensors models/checkpoints/ text-to-3d Comfy-Org/flux1-schnell
birefnet.safetensors models/background_removal/ text-to-3d Comfy-Org/BiRefNet
ComfyUI/
└── models/
    ├── checkpoints/
    │   ├── hunyuan_3d_v2.1.safetensors
    │   └── flux1-schnell-fp8.safetensors
    └── background_removal/
        └── birefnet.safetensors

The Hunyuan3D 2.1 checkpoint holds the shape model, its image encoder and its VAE, so it is the only file the image-to-3D workflow needs. To use Hunyuan3D 2.0 instead (smaller, 4.9 GB), download hunyuan3d-dit-v2_fp16.safetensors into models/checkpoints/, pick it in the loader, and set the latent resolution to 3072, steps to 20 and cfg to 8. FLUX.1 [schnell] and BiRefNet are released under licences that allow commercial use; check the Hunyuan3D 2.1 licence for where it applies.

The image-to-3D workflow works best with one object, whole and centred, on a plain light background (a white sketch page is ideal). For photos with busy backgrounds, copy the BiRefNet nodes (9 to 12) from text-to-3d.json in front of the Hunyuan3D part.

How SculptForge fills in a workflow

Any API-format workflow works with SculptForge if it follows these rules. The workflows here already do.

  • Image: every LoadImage node titled sf_image gets the picture SculptForge uploads. If no node has that title, the first LoadImage node does.
  • Text: the node titled sf_prompt gets the prompt in its text, prompt or string input, whichever it has. In text-to-3d.json that is a Trim Text node, so the prompt then gets the 3D-friendly framing ("a single object, centered, plain white background, ...") appended before it reaches FLUX.
  • Seed: a node titled sf_seed, and every input named seed or noise_seed, may get a fresh random seed.
  • Result: SculptForge downloads the first mesh file (.glb, .gltf, .obj, .ply, .stl, .fbx, .usdz) in the finished job's outputs. Other outputs, such as the concept image, are ignored.

To make your own: build the graph in ComfyUI, set node titles with right click > Title, end it with a mesh saver (Save 3D Model, or SculptForge Export), then save it with Workflow > Export (API).

Checking a workflow

Two scripts in tools/ check workflows against a running ComfyUI. Both use only Python's standard library (and curl).

# Every node exists on the server, every required input is set, every link points at a real output of the
# right type, values are in range, an output node exists, plus which nodes SculptForge will fill in.
tools/comfyui-validate-workflow.py integrations/comfyui/workflows/*.json
tools/comfyui-validate-workflow.py --server http://192.168.1.20:8188 --strict my-workflow.json

# Run one for real like SculptForge does: upload assets/demo/teddy-sketch.png, fill in the inputs, queue it,
# wait, download the mesh and check it.
tools/comfyui-smoke.sh integrations/comfyui/workflows/sculptforge-inflate.json
tools/comfyui-smoke.sh --prompt "a carved stone owl" --out /tmp/meshes integrations/comfyui/workflows/text-to-3d.json

Missing model files and an unloaded SculptForge pack are warnings in the validator (errors with --strict), since they are about the server rather than the workflow.