Innovate Futures @ Benji

ComfyUI Hunyuan 3D 2.0 Multi View - 1 Image Create 3D Mesh Object No Problem!

Added 2025-03-25 13:42:30 +0000 UTC

In this blog post, I’ll guide you through my ComfyUI workflow for generating 3D objects using Hunyuan 3D 2.0 MV (Multi-View). This native integration allows you to create detailed 3D meshes from multi-angle images without relying on external custom nodes. Let’s dive into how this powerful tool works and how you can optimize it for the best results.

Workflow Overview

The Hunyuan 3D 2.0 MV workflow in ComfyUI enables the generation of 3D mesh objects from multi-view reference images. Unlike traditional single-view 3D reconstruction, this method uses multiple angles (front, side, back, etc.) to improve accuracy and detail.

Key Features

Multi-View Conditioning – Uses multiple images (front, side, back) to generate a more accurate 3D mesh.
Native ComfyUI Support – No need for custom nodes; works directly with the latest ComfyUI update.
Fast & Lightweight – The FP16 model is only 4.9GB, making it efficient for local GPUs.
High-Resolution Output – Default resolution of 3072px, adjustable for better quality.
3D Mesh Export – Saves generated objects in .obj format for use in Blender or other 3D software.

Step-by-Step Workflow

1. Setting Up Hunyuan 3D 2.0 MV in ComfyUI

Ensure you have the latest ComfyUI version (git pull in the terminal).
Download the Hunyuan 3D 2.0 MV model from Hugging Face (FP16 SafeTensor file).
Place the model in ComfyUI/models/checkpoints/.

2. Loading the Multi-View Workflow

In ComfyUI, navigate to the 3D tab and select Hunyuan 3D 2.0 MV.
The default workflow includes:
- Clip Vision Encoders (for multi-view conditioning).
- K Sampler (steps: 20, CFG: 7.5).
- Hunyuan 3D VAE Decode (specialized for 3D mesh reconstruction).
- Mesh Output Node (exports the 3D object).

3. Preparing Multi-View Input Images

For best results:

Use clean, isolated images (remove backgrounds to avoid noise).
Capture at least three angles (front, side, back).
(Optional) Use WAN 2.1’s 360° Rotation LoRA to generate multi-view references from a single image.

4. Running the Generation

Load images into the respective Clip Vision slots (front, side, back).
Adjust resolution (default: 3072px) if needed.
Click Queue Prompt to generate the 3D mesh.
The output is saved in ComfyUI/output/mesh/ as an .obj file.

Multi-View vs. Single-View: Key Differences

Feature Multi-View (Hunyuan 3D 2.0 MV) Single-View (Standard Hunyuan 3D 2.0)

Input Requirements Multiple angles (front, side, back) Single image (front view only)

Accuracy High (captures full 3D structure) Low (guesses side/back details)

Use Case Detailed 3D models for animation Quick previews, low-detail objects

Output Quality Sharp edges, correct proportions Blurry sides, floating artifacts

Optimizing the Workflow

1. Remove Backgrounds Before Processing

Use AI segmentation tools (e.g., RemBG) to isolate objects.
Prevents unwanted elements (grass, trees) from corrupting the 3D mesh.

2. Use 360° Rotation for Reference Images

Generate multi-view angles with WAN 2.1’s 360° LoRA.
Helps capture all sides without manual photography.

3. Post-Processing in Blender

Import the .obj file for:
- Texturing (use AI-generated color transfers).
- Rigging & Animation (if needed).

4. Style Transfer with WAN 2.1

Record a rotating 3D preview in ComfyUI.
Use Depth Map Control LoRA to re-colorize the mesh in video form.

Conclusion

The Hunyuan 3D 2.0 MV workflow in ComfyUI is a game-changer for AI-assisted 3D modeling. By leveraging multi-view inputs, you can generate high-fidelity 3D meshes with accurate proportions—far surpassing single-view methods.

Key Takeaways:
✅ Multi-view = Better Accuracy (always remove backgrounds!).
✅ Use 360° LoRA for easy reference image generation.
✅ Export to Blender for texturing/animation.
✅ Combine with WAN 2.1 for dynamic style transfers.

Ready to try it out? Download the Hunyuan 3D 2.0 MV model and start creating! For more advanced workflows, check out my tutorials on WAN 2.1 video enhancements and Video2Video using Depth Lora