This basic guide explains how to use AI tools to create high quality 3d generated models using Comfyui and Hunyuan3D-2. The installer follows the standard manual install steps to deploy this software on windows using a windows bat file. (scroll down this page to section Installation files to download the bat file which you can run from any dir. But before you run the installer read the guide and do the prerequisites and hardware check.)

The installer sets up a free free, open-source ComfyUI Portable (portable means it has an embedded python so you dont have to manage any technical python code setup on your pc) install with Hunyuan3D-2 plugin on your local Windows PC. You dont have to do any git, python or technical configuration. The only setup you do is in the local ComfyUI webapp desktop to do final config and update.
Its possible to install on linux/mac with nvidia Nvidia using the exact same steps but the script is tested and works for Windows. A converted Ubuntu linux install script is also provided but its untested.

Using this setup, you can quickly generate game-ready placeables, creatures, or tileset parts from either a 2D image or a text prompt, producing fully textured 3D models with UV mapping. The AI intelligently predicts the 3D shape and even generates unseen textures based on the image input. (Check out the example screenshots of a gorilla’s 2D image 2d images: Cobra statue, Red Dragonborn and gorilla on a white background and its resulting 3D model in Blender.)the resulting generated models by Hunyuan3D-2 exported intoBlender.)

Default Hunyuan-3d-2 workflow using ComfyUI:

Generating a 3D model with textures typically takes 1 to 3 minutes with Nvidia gpu, depending on your chosen quality settings and PC hardware. If you have an a 6 GB vram or better Nvidia GPU, you’ll achieve the faster and higher-quality results. However, if you don’t have an Nvidia card, you can use your CPU (slow choice many mins generating) or try an online service.

...

Hunyuan3D-2 a free Online 3D Model Generation (Hugging Face hosted) 

If you prefer an online option, Hugging Face offers a free 3D model generator powered by Hunyuan3D-2, accessible here which has a simple but limted interface:
🔗 Hunyuan3D-2 on Hugging Face Screenshot of the web app below shows converting a 2d image of cobra head (top left) to a 3d model shown side and back: 

Below shows source 2d image ready for conversion to 3d. A 2d cobra image placeable statue generated from AI descripton from Google gemini imagin3:using Google Gemini Imagen3. 

Online tool Next shows the free online tool Hunyuan3D-2 with imported 2d image (or you could give it an AI text chat description for what you want it to 3d create) and then the generated 3d model with texture button pressed and result. 

The result is very similar to the 2d image, whats missing is the gold star on its body. By generating again with another seed number (random starter value) then other variations can be produced.  

Click to zoom in:

This online version has a fixed face count of 40K, and while the results can be impressive, they lack the fine-tuned control available in ComfyUI. (40k faces single mesh is too much for NWN).
This simplified generator can also be installed locally but for the effort and features you are better off installing ComfyUI.

...

GLB 3d model file then imported perfectly into blender with UV texture setup automatically:

Local ComfyUI vs. Online Services

If you want full control over settings like polygon count, texture resolution, and quality, a local installation of ComfyUI is the best choice. Some third-party providers also offer hosted ComfyUI instances with a pay-per-use model.

On Hugging Face’s free tier, you can generate three models per day before needing to purchase credits. However, if you have a 6GB+ VRAM Nvidia GPU and at least 24GB RAM, you can generate higher-quality 3D models directly on your PC within minutes.

ComfyUI: AI-Powered 3D Model Generation from 2D Images or Text

ComfyUI and its plugin AI model generators like Hunyuan3D-2 allow you to generate 3D models with textures and UV maps from either a 2D image or a text description. The AI intelligently reconstructs the missing parts of the image, creating a fully detailed 3D model ready for import into Blender.

...

https://discord.com/channels/1315602970867662878/1315608784135131156

NWNVault Discord

Join this discord to discuss ComfyUI in the NWNVault workflow-and-tools channel

https://discord.com/channels/255017439371329537/511127010962046986

ComfyUI example workflow is a accessed as web baed browser app with python installed 3d generator server. that connects to a python server application. So it can be installed locally on same pc or as remote service or for easy sharing or deployment. 
These screen shots shows the ComfyUI desktop and workflow and of the Cobra statue 2d image input top left. The setup workflow is provided and so you just load the image and tweak some settings and then press queue button to generate the 3d model.  
The workflow spans across three pages and you can drag/move/zoom the canvas to see the workflow. It starts top left and the steps flows to the bottom right where the completed 3d model is shown.

Image Added

Image Removed

Features & Capabilities

• Polygon Control: Generate models with exact faces count you want from low poly (5K faces) to high poly (250K+ faces).
• Texture Quality: Adjustable texture resolution, supporting up to 2K upscaled and detailed textures.
• Extensive open source plugin eco system: 2d AI image generators, Advanced upscalers, PBR texture generation and many more. 

Sample generate AI 2d image Red dragonborn to 3d red dragonborn model

2d image created of Red dragonborn in A pose created using google AI Imagen3 from simple text description and result generated 3d model: 

Sample generated AI 2d Image to 3D Gorilla Model

2d gorilla image in A pose with white background created by google Imagen3:

3d model generated by Comfyui Hunyuan3D-2 imported into Blender:

To demonstrate the capabilities of ComfyUI you can import the generated this sample file in Blender . The generated 3d model comes with UV map and texture already setup so import is easyof the generated gorilla. 

...

The screen shot of Blender 3D gorilla model shown has 20K faces, which was achieved by splitting the original model into two meshes (top and bottom halves, each with 10K faces) by cropping the 2d image into two parts from the belt: part below is waist and legs, part above is torso, arms and head. This generated two 3d model meshes and was then merged as one model in Blender. This technique is useful optimization for Neverwinter Nights, where the per-mesh face limit is 10K. Also breaking the model into meshes with less to create per part makes it more detailed and easier for the 3d generator limited texture memory.

Ease of Use & Batch Processing

📌 No programming or workflow setup is required— open the provided workflow and simply tweak a few parameters to match your hardware, and you can generate a fully textured 3D model in 1 to 3 minutes.
📌 Batch processing is supported, allowing you to generate multiple models at once with different images or settings. 

Getting the Best Results from Your 2D Images

• For characters use a A-pose or T-pose facing front not at an angle and filling the whole picture frame to the edge 
• Ensure a sharp, high-quality image with a plain white background
• Avoid weapons or accessories obstructing the body
• For placeables, use a clear, full-frame image with a white background not too many placeable items or parts in one scene. 
• Fine tune settings in ComfyUI. Once you have a good workflow and settings it can be used over and over.
• Upscale the 2d image to 2k or 4k. Split 2d image into parts. Generate each part at 10k faces (nwn limit per mesh). Import parts into blender, rescale and compose into multi mesh model. 
• Fine detail. (eg heads) Upscale image to 4k and cut out the face image and scale face to fill most of frame. Leave behind the hair as that take a lof of verticies. Generate the face model at 10k faces or less, import into blender and rescale and compose to replace the low poly head with the new 10k higher res head.    

Fine-Tuning Settings for Best Results

To achieve high-quality 3D models and textures, it’s important to adjust a few key parameters:

Note: as you increase the parameter numbers (quality settings) it causes more work for the cpu and memory, so a 1 to 3 min model generation on default settings with boosted settings for highest quality it can take 5 to 10 mins per model. 

Batch mode: Dont work on a single model at a time. Use your time effiecently by working on multiple images and models to get the best quality and a batch workflow going. While the GPU is busy in batch mode generatig 20 models you could be editing the 2d images into parts, upscaling or cutting out heads as higher res parts, inspecting the many model generated already to select the best one, creating walkmeshes, composing multiple model meshes into a single higher quality model. (eg the gorilla example)

Its best to use higher settings and submit multiple times the same image so you can try different output quality to find the best models generatable. By not waiting for the current run to finish you just submiting again with either a new image or different parameters and it queues up and runs automatically and you can do many images/models in a batch run at the same time. When you queue it may group the generated models and so do stages with multiple models to make generation more effiecent. You can monitor the batch run via the log or on the top left there is a batch queue toggle button to show the queue and results. You can right click and choose load workflow on the preview image it will load the workflow with image and settings from a previous submission. 

ComfyUI queue sumbission makes it very efficient to try multiple settings (recommeded to get the best results) with long generation time (5 to 10 mins per model) as Comfyui batch mode runs in the background to convert to models. It can run for hours and then you can inspect results and can decide what you want to do with the models or resubmit with new settings in another batch run. While your GPU and pc is busy you can do the work of 2d image management or do light blender work to inspect the models but you cant really use do gpu heavy tasks while its doing batch mode.

Workflow node properties tuneable:

✅ Num Faces – Controls polygon count
✅ Octree Rendering – Helps with structure refinement
✅ Steps Count – Affects model detail level (aim for 50 to 100)
✅ Guidance scale – Affects how much the 2d image will be used exactly to generate model/texture or how much freedom AI can get to make up its own design. able to set between 0 and 10. 0.75 default which means 75% follow the image. Higher values tend to produce better looking models. lower values result is more like playdo mess. It works fine at 9.5 or 9.8 if you want to follow the image exactly. 
✅ Texture Size – Defines texture resolution
✅ Delight Settings – Adjusts texture lighting and color balance. Can make a big difference so worth adjusting per starter image. default 0.75 try range between 0.50 to 0.90 for whats works for your image brighness.  
✅ Upscaler 4x – Improves texture quality (OFF by default; should be turned ON). A 6x better quality ai upscaler extra plugin is recommended.

✅ Camera Settings – You can add more camera angles by adding in the degree numbers in the comma separated list. Default is 6 camera positions (front, back, left, right, top, bottom stored as 90 degree angles). if you add in extra 45 degree angles you get more coverage which helps to paint the model better. note: more angles takes more vram. You see the extra camera position and image produced when you run the workflow.
✅ Primitive reference image size (Texture Memory Renderer) – Recommended increase to 786MB or 1024MB (Default is 512MB, but at 1024MB, the AI may generate unintended details. try it with 1024 and different seeds to shake off the small chance of painting hallucination, or reduce memory to 786mb)

Extending Your Workflow

ComfyUI is open-source, meaning new features, plugins, and enhancements—such as other 3D or 2D AI generators, PBR texture processing and advanced upscalers—are continually being developed.

Installing ComfyUI with Hunyuan3D Workflow (Automatic Installer)

System Requirements

📌 Operating System: Windows 10 or 11 (Linux and Mac also but this install script is Windows bat file)
📌 Disk Space: 50GB free space in your Windows home folder. Ideally fast SSD. Can be moved to another drive after install is done. 
📌 RAM: 24GB minimum (32GB recommended for multitasking)
📌 GPU: Nvidia GPU with at least 6GB VRAM (8GB–12GB recommended for best performance)

...

• If you have less than 8GB VRAM, you can enable the low VRAM mode when launching ComfyUI.
• By default, ComfyUI automatically detects available VRAM.
• The installation downloads large AI model files (~50GB total), so a fast internet connection is recommended.
• Installation time: ~2 hours (depending on internet speed).

...

Installation Process (new Windows installer uptodate)

Follow this new guide to install the latest version of comfyui. It also has an easier process and more sophisticated windows installer compared to the bat installer i created below. It automates the same install process but uses a newer version of Comfyui and Hunyuan 3d-2.

The steps I did for a fresh install is to run the installer let to download fully and wait for the message 'all startup tasks have been completed'. then https://github.com/YanWenKun/ComfyUI-Windows-Portable Download the two zip files. Download the models and place in the dir as noted after you run the installer that creates the empty model dirs.

1) copy all the files in dir ExtraScripts to one dir up so you can run_nvidia_gpu.bat to start ComfyUI. (you cant start comfyui with run_nvidia_gpu.bat in the ExtraScripts dir but it works one dir up)

2) Manager button. Press Update Comfyui and restart then refresh the webpage.

3) install two custom nodes using Manager button and Custom Nodes Manager: ComfyUI Hunyuan-3D-2 and ComfyUI-ComfyUI-Hunyuan3DWrapper.

4) In CMD terminal go to root dir and run: .\python_standalone\python.exe -m pip install "Comfyui\custom_nodes\comfyui-hunyan3dwrapper\wheels\custom_rasterizer-0.1.0+torch260.cuda126-cp312-cp312-win_amd64.whl"

5) From Comfyui web interface menu (top left) Workflow/Open to load workflow in the ComfyUI\custom_nodes\comfyui-hunyan3dwrapper\example_workflows\hy3d_example_01.json.

6) Top left on the now loaded worflow open image button and choose the penguin found in the example_workflows dir from the last step you did.

7) Press the blue button RUN located at the bottom of the workflow. Wait a few mins for models to download and the 3d penguin should appear bottom right.

Installation Process (old version)

ComfyUI installation involves five main steps. Follow the steps in order and do not skip any to ensure a successful setup. The install guide details via screenshots every step.

💡 Good news! The installer handles all the heavy lifting, so no manual setup is required for windows but you do need to do the ComfyUI gui setup of plugins which are shown with screenshots and explanation. 

Installation Files
 

You need to download and install CUDA and git and then run two batch files. 

...

Install Guide

Follow this google doc or download the pdf that has screenshots and detailed how to install step by step:

Widget Connector
url https://docs.google.com/document/d/1QAadTW3AFNQybGDaSGo3kQa3psN1GsHXHUzeDZ_D0XU/edit?usp=sharing

Overview of steps you follow to do ComfyUI Install:

1) Download and Install Nvidia CUDA drivers
https://developer.nvidia.com/cuda-downloads

...

2)Download and Install git software. https://git-scm.com/downloads/win Use blue link Click here to download. Run the installer. The setup will prompt you with several options; it’s fine to use the default settings for all of them. You will not have to use Git but the installer needs it.  

3)Create Github.com account. This is so the installer can use your account login to download the software it needs when running.

...

You can import GLB model into Blender with UV textures already setup.

Linux Installer

Untested Linux install script converted from Widows bat to Ubuntu Debian compatible bash shell script: