NodeWise Studio
RUDownload for Linux

A picture into a 3D model: Hunyuan3D 2

<!-- lesson: 3d-image -->

Lesson "3d-image", in the group 3D of the lesson list. Steps: 13.

What you will be able to do

You build a graph that turns one picture of an object into a 3D model you can open in other programs.

What you need

  • It goes on from A first picture from text. The card says so; it is a hint, not a lock.
  • Model files the lesson checks:
    • for ckpt_name of Image Only Checkpoint Loader (img2vid model): a file whose name has hunyuan3d.
  • Nodes: Image Only Checkpoint Loader (img2vid model), ModelSamplingAuraFlow, Load Image, CLIP Vision Encode, Hunyuan3Dv2Conditioning, Empty Hunyuan 3D v2 Latent, KSampler, Hunyuan 3D VAE Decode, Voxel to Mesh, Save 3D Model.
  • The app downloads and installs nothing. The step texts below say which files and folders, and which ComfyUI template names them.

The steps

The words of each step, as the lesson shows them. Node names are in bold.

  1. What comes out (reading)

    Hunyuan3D 2 looks at one picture of an object and builds its shape in 3D. The result is a mesh, a surface of many small triangles, without colours: a white model.

    The best picture shows one object whole, on a plain background, from a little above and to the side.

    The lesson builds the graph from an empty canvas: "File ▸ New project".

  2. Load the model

    Add a Image Only Checkpoint Loader (img2vid model) node. In ckpt_name choose hunyuan3d-dit-v2_fp16.safetensors (version 2.0) or hunyuan_3d_v2.1.safetensors (version 2.1) from models/checkpoints.

    This loader takes a model that reads pictures, not text: its outputs are the model, the picture reader CLIP_VISION and the decoder.

    If the file you need is not in the list, your ComfyUI does not have it. The app downloads nothing. The easiest way is to open, in ComfyUI itself, the template "HY 3D 2.0" or "HY 3D 2.1: Single View to Model": it shows which files are missing and where to put them. Then choose "Sync ▸ Sync now" in this app's menu, and the files appear in the lists.

  3. Set the noise shift

    Add a ModelSamplingAuraFlow node, connect the MODEL output of Image Only Checkpoint Loader (img2vid model) to its model input and set shift to 1.

    Both templates use 1. This model was trained for that value.

  4. Load the picture

    Add a Load Image node and choose the picture in its image field.

    A background with other objects becomes part of the model. Remove it first if you can.

  5. Let the model see the picture

    Add a CLIP Vision Encode node. Connect the CLIP_VISION output of Image Only Checkpoint Loader (img2vid model) to its clip_vision input and the picture from Load Image to its image input. Set crop to center.

    This node turns the picture into a prompt the model can follow.

  6. Turn it into a condition

    Add a Hunyuan3Dv2Conditioning node and connect the output of CLIP Vision Encode to its clip_vision_output input.

    It gives two outputs, positive and negative, ready for the sampler. Nothing is written by hand here: the picture is the whole prompt.

  7. Add the blank

    Add a Empty Hunyuan 3D v2 Latent node and set resolution to 3072.

    This is how finely the shape is worked out. Version 2.0 uses 3072, version 2.1 uses 4096.

  8. Add the sampler

    Add a KSampler node and connect four inputs:

    • model from ModelSamplingAuraFlow;
    • positive and negative from the outputs of the same name of Hunyuan3Dv2Conditioning;
    • latent_image from Empty Hunyuan 3D v2 Latent.

    Set steps to 20 and cfg to 8, as the template of version 2.0 does. For version 2.1 the template uses 30 and 5.

  9. Turn the result into a volume

    Add a Hunyuan 3D VAE Decode node. Connect the output of KSampler to its samples input and the VAE output of Image Only Checkpoint Loader (img2vid model) to its vae input.

    The result is a voxel grid: space cut into small cubes, each either inside the object or outside.

  10. Turn the volume into a mesh

Add a Voxel to Mesh node and connect the output of Hunyuan 3D VAE Decode to its voxel input. Set algorithm to surface net and threshold to 0.6.

surface net gives a smoother surface than basic. The threshold says where the inside ends: lower makes the object fatter, higher thinner and with holes.

  1. Save the model

Add a Save 3D Model node and connect the output of Voxel to Mesh to its mesh input.

  1. Where the file goes and what opens it (reading)

In ComfyUI the node writes a .glb file into the output/3d folder and shows the model on the canvas.

A .glb file opens in Blender (File ▸ Import ▸ glTF), in most 3D viewers and game engines, and in a web browser through any glTF viewer. The model has no colours yet: they are painted in another program or by a texturing graph.

The ComfyUI-3D-Pack pack and the TRELLIS.2 nodes of ComfyUI itself also build models with colours, but their graphs are much larger.

  1. The graph is ready (reading)

Save the graph with "File ▸ Save workflow for ComfyUI…", open the file in ComfyUI and run it there. This app runs nothing in ComfyUI itself.

If the model has holes or lumps: a cleaner picture helps most, then a larger resolution.

What the lesson checks

A reading step is done when you press Got it, next. A doing step ticks itself off when the canvas holds:

  • Step 2, Load the model:
    • a Image Only Checkpoint Loader (img2vid model) node.
    • ckpt_name of Image Only Checkpoint Loader (img2vid model) holds a file whose name has hunyuan3d (capitals, dots and dashes do not matter).
  • Step 3, Set the noise shift:
    • a ModelSamplingAuraFlow node.
    • MODEL of Image Only Checkpoint Loader (img2vid model) into model of ModelSamplingAuraFlow.
    • shift of ModelSamplingAuraFlow is 1.
  • Step 4, Load the picture:
    • a Load Image node.
    • image of Load Image: a value is chosen.
  • Step 5, Let the model see the picture:
    • a CLIP Vision Encode node.
    • CLIP_VISION of Image Only Checkpoint Loader (img2vid model) into clip_vision of CLIP Vision Encode.
    • IMAGE of Load Image into image of CLIP Vision Encode.
    • crop of CLIP Vision Encode is center.
  • Step 6, Turn it into a condition:
    • a Hunyuan3Dv2Conditioning node.
    • CLIP_VISION_OUTPUT of CLIP Vision Encode into clip_vision_output of Hunyuan3Dv2Conditioning.
  • Step 7, Add the blank:
    • a Empty Hunyuan 3D v2 Latent node.
    • resolution of Empty Hunyuan 3D v2 Latent is from 3072 to 4096; Do this for me sets 3072.
  • Step 8, Add the sampler:
    • a KSampler node.
    • MODEL of ModelSamplingAuraFlow into model of KSampler.
    • positive of Hunyuan3Dv2Conditioning into positive of KSampler.
    • negative of Hunyuan3Dv2Conditioning into negative of KSampler.
    • LATENT of Empty Hunyuan 3D v2 Latent into latent_image of KSampler.
    • steps of KSampler is from 20 to 30; Do this for me sets 20.
    • cfg of KSampler is from 5 to 8; Do this for me sets 8.
  • Step 9, Turn the result into a volume:
    • a Hunyuan 3D VAE Decode node.
    • LATENT of KSampler into samples of Hunyuan 3D VAE Decode.
    • VAE of Image Only Checkpoint Loader (img2vid model) into vae of Hunyuan 3D VAE Decode.
  • Step 10, Turn the volume into a mesh:
    • a Voxel to Mesh node.
    • VOXEL of Hunyuan 3D VAE Decode into voxel of Voxel to Mesh.
    • algorithm of Voxel to Mesh is surface net.
    • threshold of Voxel to Mesh is from 0.5 to 0.7; Do this for me sets 0.6.
  • Step 11, Save the model:
    • a Save 3D Model node.
    • MESH of Voxel to Mesh into mesh of Save 3D Model.

See also

This page is imported from the application repository's wiki (docs/wiki/lessons/Image-To-3D.md, revision b8e3f71). The same guide is built into the app: Help ▸ User guide.

To refresh it, run npm run import and then npm run build in the website folder.