SAM3_SmartInpainter

# SAM3 Smart Inpainter for ComfyUI

> 🎯 One node to find faces (or any object) and repaint them — using text prompts.  
> ✅ Tested with **SAM3 only** (`sam3.pt`)

---

## ✨ What This Node Does

This node combines **segmentation** + **inpainting** in one step:

1. You type a prompt like `"face"`, `"eyes"`, `"scarf"`
2. The node finds that area using the **SAM3** AI model
3. It repaints *only that area* using your diffusion model
4. Blends the result back into your original image

✅ Great for: face detail, fixing eyes, changing accessories, removing small objects  
✅ Works with: any ComfyUI text-to-image workflow  
✅ Model supported: **SAM3** (`sam3.pt` from HuggingFace)

---

## 📦 What You Must Install

### Step 1: Install a SAM3 Wrapper (Required)

This node needs **one** of these custom nodes to provide the `sam3` Python package:

#### Option A (Recommended): ComfyUI-Easy-Sam3
```bash
cd custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Sam3.git

Option B: ComfyUI-RMBG

cd custom_nodes
git clone https://github.com/1038lab/ComfyUI-RMBG.git

Step 2: Install Python Requirements

A requirements.txt file is included. Install it using the method that matches your ComfyUI setup:

---

🔧 Installation Guide

For ComfyUI Portable (Windows python_embeded)

# 1. Go to ComfyUI root folder
cd F:\ComfyUI

# 2. Install this node
cd custom_nodes
git clone https://github.com/majidfida/SAM3_SmartInpainter.git

# 3. Install requirements using ComfyUI's embedded Python
..\python_embeded\python.exe -m pip install -r SAM3_SmartInpainter\requirements.txt

# 4. Install SAM3 wrapper (if not already installed)
git clone https://github.com/yolain/ComfyUI-Easy-Sam3.git

For ComfyUI with Virtual Environment (Linux / Mac / Windows venv)

# 1. Activate your venv
# Windows:
venv\Scripts\activate
# Linux/Mac:
source venv/bin/activate

# 2. Go to custom_nodes and install this node
cd custom_nodes
git clone https://github.com/majidfida/SAM3_SmartInpainter.git

# 3. Install requirements
pip install -r SAM3_SmartInpainter/requirements.txt

# 4. Install SAM3 wrapper (if not already installed)
git clone https://github.com/yolain/ComfyUI-Easy-Sam3.git

requirements.txt Content

opencv-python>=4.5.0
omegaconf>=2.1.0
hydra-core>=1.1.0

---

🗂️ Where to Put Your SAM3 Model

Download sam3.pt from:
🔗 https://huggingface.co/Majidfida/Sam3/tree/main

Then place it in:

ComfyUI/models/sam3/sam3.pt

---

🔧 How to Use (Simple Steps)

Basic Face Inpaint Workflow:

[Load Image] → [SAM Smart Inpainter] → [Save Image]
                   ↑
         Connect: MODEL, CLIP, VAE, POSITIVE, NEGATIVE

Important Parameters:

| Parameter | Recommended Value | What It Does | |-----------|------------------|--------------| | prompt | "face" or "eyes" | What to find and repaint | | text_threshold | 0.25 – 0.35 | Lower = finds smaller/blurry things | | min_mask_area | 0.001 (0.1%) | Allow tiny masks (for far-away faces) | | crop_padding | 64 – 128 | Give extra space around the mask | | mask_dilation | 6 – 10 | Make thin masks thicker (prevents holes) | | inpaint_resolution | 768 or higher | Better detail in the repainted area | | denoise | 0.4 – 0.6 | How much to change the original (lower = more subtle) | | blend_mode | "feather" | Smooth edges when pasting back |

Tips for Small or Distant Faces:

• Set min_mask_area → 0.0005
• Set text_threshold → 0.25
• Set crop_padding → 128
• Keep inpaint_resolution at 768 or higher

💡 Tip: Inpaint Multiple Faces (Man + Woman)

Want to repaint two different faces in one image? Just chain multiple nodes:

[Load Image]
       ↓
[SAM Smart Inpainter #1] → prompt: "woman face"
       ↓ (output_image)
[SAM Smart Inpainter #2] → prompt: "man face"
       ↓
[Save Image]

Each node finds and repaints only its target. Connect the output_image of the first node to the image input of the second. Use specific prompts like "woman face", "man beard", "girl eyes" to avoid overlap.

---

⚙️ Advanced: Use With Any Workflow

This node works like a "smart crop + inpaint" block. You can plug it into:

• Text-to-image pipelines
• Image-to-image upgrades
• Face restoration chains
• Object removal/editing flows

Just connect:

• MODEL, CLIP, VAE from your checkpoint loader
• positive / negative from your prompt encoders
• image from your input

---

❗ Important Notes

🔹 I will not monitor this repository.
I may not reply to issues, messages, or pull requests. This node is shared as-is for the community to use and improve.

🔹 SAM3 only.
This node has only been tested with sam3.pt. SAM2 support is not included.

🔹 For face inpainting mainly.
This node is designed mainly for faces, but you can type any prompt: "hat", "logo", "tree" — it will try to find and repaint it.

🔹 Got errors?

• Make sure you installed ComfyUI-Easy-Sam3 or ComfyUI-RMBG
• Make sure sam3.pt is in ComfyUI/models/sam3/
• Make sure you ran pip install -r requirements.txt
• Check the console log — the node prints helpful debug info

---

🙏 Credits

• SAM3 model: Meta AI Research / Facebook (HuggingFace)
• ComfyUI-Easy-Sam3: @yolain
• ComfyUI-RMBG: @1038lab
• This node: Community contribution

---

💡 Simple rule: If the mask is too small → lower min_mask_area.
If the repaint looks blurry → raise inpaint_resolution.
If edges look harsh → increase mask_blur or use blend_mode: feather.

Happy inpainting! 🎨✨