Sign In

UltraShape for ComfyUI — High-Fidelity 3D Mesh Refinement Nodes

3

UltraShape for ComfyUI — High-Fidelity 3D Mesh Refinement Nodes

If you’re generating **coarse 3D meshes in ComfyUI** (for example with **Hunyuan3D-2.1**) and want to push them to **production-quality geometry**, this project wraps **UltraShape 1.0** into native ComfyUI nodes — fully integrated, batch-friendly, and workflow-ready.

This is a **community ComfyUI wrapper** around the official UltraShape implementation from PKU-YuanGroup.

---

## 🔥 What Is UltraShape?

**UltraShape 1.0** is a **two-stage diffusion-based mesh refinement model** designed to transform **coarse 3D meshes** into **high-detail, clean geometry**.

It focuses on:

* Adding **high-frequency geometric detail**
* Improving **surface smoothness**
* Enhancing **mesh topology**
* Preserving the original object structure

UltraShape is especially powerful when paired with **image-to-3D pipelines** like Hunyuan3D.

**Typical pipeline:**

```
Input Image

Hunyuan3D (coarse mesh)

UltraShape (refined mesh)

GLB / OBJ / Production Asset
```

---

## 🧩 What This ComfyUI Node Pack Adds

This repository exposes UltraShape as **first-class ComfyUI nodes**, making it easy to drop into existing 3D workflows.

### Included Nodes

* **LoadUltraShapeModel**
* **UltraShapeRefineFromHunyuan**
* **UltraShapeRefine (from mesh path)**

### Key Features

✅ Seamless integration with Hunyuan3D
✅ Supports **batch processing**
✅ Works with existing ComfyUI 3D nodes
✅ No scripting required
✅ Docker image available with all dependencies preinstalled

---

## 🖼 Example Workflow

**Hunyuan3D → UltraShape refinement → Export**

> 📌 Replace this image with your actual workflow screenshot

```

screenshot.png


```

You can also download the exact workflow JSON used here:

👉 **Example Workflow JSON**
`3d_hunyuan3d-ultrashape-v2.1.json`

---

## 🧠 How UltraShape Refines Meshes

UltraShape operates on an existing mesh and:

* Converts it into a voxel representation
* Runs a **two-stage diffusion refinement**
* Reconstructs the surface via **marching cubes**
* Outputs a significantly more detailed mesh

Unlike remeshing tools, UltraShape **adds real geometric detail**, not just smoothing.

---

## ⚙️ Installation

### Option 1 — ComfyUI Manager (Recommended)

1. Install **ComfyUI Manager**
2. Open Manager → Search for **UltraShape**
3. Click **Install**
4. Restart ComfyUI

---

### Option 2 — Manual Install

```bash
cd ComfyUI/custom_nodes
git clone --recursive https://github.com/hahihula/ComfyUI-UltraShape.git
```

If already cloned without submodules:

```bash
cd ComfyUI-UltraShape
git submodule update --init --recursive
```

Restart ComfyUI afterward.

---

## 🐳 Docker (One-Command Setup)

A prebuilt Docker image is available with everything configured:

```bash
docker pull hahihula/comfyui-ultrashape:latest

docker run --gpus all -p 8188:8188 \
-v $(pwd)/models:/opt/comfyui/models \
-v $(pwd)/output:/opt/comfyui/output \
hahihula/comfyui-ultrashape:latest
```

**Included:**

* ComfyUI v0.6.0
* PyTorch 2.8.0 + CUDA 12.8
* Flash Attention
* All UltraShape dependencies
* Optimized for RTX 3090+

---

## 📥 Model Download

UltraShape requires a separate model checkpoint:

```bash
mkdir -p ComfyUI/models/ultrashape/

wget https://huggingface.co/infinith/UltraShape/resolve/main/ultrashape_v1.pt \
-O ComfyUI/models/ultrashape/ultrashape_v1.pt
```

---

## 🧪 Usage in ComfyUI

### Basic Node Setup

1. **LoadImage**
2. **Generate coarse mesh** (Hunyuan3D-2.1)
3. **LoadUltraShapeModel**
4. **UltraShapeRefineFromHunyuan**
5. **SaveGLB**

```
LoadImage

Hunyuan3D

LoadUltraShapeModel → UltraShapeRefineFromHunyuan → SaveGLB
```

---

## 🎛 Important Parameters

### UltraShapeRefineFromHunyuan

* **num_inference_steps** (default: 50)
Higher = more detail, slower

* **octree_resolution** (default: 1024)
Controls mesh detail & VRAM usage

* **scale** (default: 0.99)
Mesh normalization (important!)

* **remove_bg**
Useful for noisy image inputs

---

## 🚀 Performance & VRAM Tips

| Setting | Effect |
| ----------------------- | --------------------------------- |
| octree_resolution = 512 | Faster, less detail |
| steps = 30–40 | Good quality / speed balance |
| VRAM | 12GB minimum, 16–20GB recommended |

**RTX 3090 @ 1024 resolution:**
~9–10 minutes per mesh (50 steps)

---

## 🧱 Example Output

> 📌 Insert before/after mesh preview here

```
[ INSERT EXAMPLE MESH RENDER / VIEWER IMAGE ]
```

---

## 🐛 Common Issues & Fixes

**`No module named 'ultrashape'`**
→ Run `git submodule update --init --recursive`

**Out of Memory**
→ Lower `octree_resolution` or steps

**Broken mesh output**
→ Try `scale = 0.95` or regenerate coarse mesh

---

## 🔗 Links & Credits

* **ComfyUI Wrapper**
[https://github.com/hahihula/ComfyUI-UltraShape](https://github.com/hahihula/ComfyUI-UltraShape)

* **UltraShape (Official)**
PKU-YuanGroup / UltraShape-1.0

* **Paper**
*UltraShape 1.0: High-Fidelity 3D Shape Generation via Scalable Geometric Refinement*

* **Hunyuan3D-2.1**
Tencent Hunyuan

---

## 🙏 Acknowledgements

* PKU-YuanGroup for UltraShape
* Tencent for Hunyuan3D
* ComfyUI community for the ecosystem

---

## ⭐ Final Notes

This is a **community-created ComfyUI wrapper**, not an official release from the UltraShape authors.
If you find it useful, consider **starring the repo** and sharing your results on Civitai.

Happy meshing 🧊🧠

3