ComfyUI-SBTools

Version 1.0.0

Custom node collection for ComfyUI. Background removal and color analysis tools.

Nodes

Node	Category	Description
BiRefNet (RemoveBG)	SBTools/Image	Advanced background removal with 5 model variants
Find Unused Color	SBTools/Image	Find colors maximally different from image for chroma keying

Installation

1. Navigate to your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes

2. Clone this repository:

git clone https://github.com/Amatsukast/ComfyUI-SBTools.git

3. Install dependencies:

cd ComfyUI-SBTools
pip install -r requirements.txt

4. Restart ComfyUI

Usage

All nodes are located under SBTools/Image category in ComfyUI.

BiRefNet (RemoveBG)

Advanced background removal using BiRefNet models. Supports multiple model variants optimized for different use cases.

Available Models

Model	Best For	Resolution	Notes
BiRefNet-general	Everyday use	1024×1024	Fast, balanced performance
BiRefNet-HR	High quality	2048×2048	Best detail preservation
BiRefNet-portrait	People/portraits	1024×1024	Trained on human subjects
BiRefNet_dynamic	Aspect ratio preservation	Variable (256-2304px)	No image distortion
BiRefNet_toonout	Outline extraction	1024×1024	Creates toon-style outlines

Parameters

Required:

• image - Input image
• model - Model selection

Optional:

• mask_blur (0-64) - Softens mask edges. Use for smoother transitions.
• mask_offset (-20 to +20) - Expands (+) or shrinks (-) the mask boundary
• invert_output - Swaps foreground/background
• background - Alpha (transparent) or Color (custom color)
• background_color - Custom background color (hex code, e.g., #222222)

Outputs

• IMAGE - Processed image with transparent or colored background
• MASK - Black & white mask data
• MASK_IMAGE - Visualization of the mask (RGB)

Which Model Should I Use?

For most cases: Use BiRefNet-general

• Fast processing
• Good quality
• Works well for general objects

For high-quality results: Use BiRefNet-HR

• 4x larger processing area
• Better detail preservation
• Ideal for final outputs

For people/portraits: Use BiRefNet-portrait

• Optimized for human subjects
• Better hair/skin detection
• Trained on portrait datasets

For wide/tall images: Use BiRefNet_dynamic

• Preserves aspect ratio (no squashing/stretching)
• Automatically scales to fit
• Best for non-square images

For creative effects: Use BiRefNet_toonout

• Extracts outlines instead of removing background
• Creates line art/edge detection
• Useful for stylized effects

Tips

Processing Resolution:

• Square models (general, HR, portrait, toonout) resize images to a fixed square
• Dynamic model preserves aspect ratio by resizing the longest side only

Mask Refinement:

• Start with mask_blur: 0 and mask_offset: 0
• Use mask_blur: 2-5 for softer, more natural edges
• Use mask_offset: +2 to +5 if edges are too tight
• Use mask_offset: -2 to -5 if edges include too much background

Performance:

• Smaller images process faster
• BiRefNet-general is fastest
• BiRefNet-HR takes ~4x longer but produces better results

---

Find Unused Color

Finds a color in your image that is maximally different from all existing colors. Useful for chroma keying, masking, or creating selection areas.

Use Cases

• Chroma key backgrounds - Find a color that won't interfere with your subject
• Mask generation - Create temporary backgrounds for selection tools
• Color coding - Identify safe colors for overlays or annotations

Parameters

Required:

• image - Input image to analyze

Optional:

• min_distance (0-255, default: 30) - Minimum color distance required
  • Higher values = more different from existing colors
  • Lower values = easier to find but may be similar to image colors
• sample_size (1000-50000, default: 5000) - Number of pixels to sample
  • Higher values = more accurate but slower
  • Lower values = faster but may miss similar colors

Outputs

• hex_color - Color in hex format (e.g., #00FF00)
• R - Red component (0-255)
• G - Green component (0-255)
• B - Blue component (0-255)

How It Works

1. Samples random pixels from your image
2. Tests common pure colors (green, blue, magenta, cyan, yellow, red)
3. Returns the color with the maximum distance from all sampled pixels
4. If no candidate meets the criteria, performs a coarse grid search

Tips

For chroma keying:

• Use min_distance: 40-60 for safe separation
• Pure green (0, 255, 0) is usually selected for typical images

For quick results:

• Use sample_size: 5000 (default)
• Increase to 10000-20000 for images with complex color palettes

Distance metric:

• Distance is calculated as Euclidean distance in RGB space
• A distance of 30 means the color differs by ~30 units per channel on average

---

Model Storage

BiRefNet models are automatically downloaded and stored in:

ComfyUI/models/sbtools/BiRefNet/

Models are downloaded on first use and cached for future sessions.

Credits

BiRefNet Node is based on ComfyUI-RMBG by AILab-AI

BiRefNet Models by ZhengPeng7 - HuggingFace | GitHub

License

Code License

This project is licensed under GNU General Public License v3.0 (GPL-3.0).

• BiRefNet node is based on ComfyUI-RMBG by AILab-AI (GPL-3.0)
• Find Unused Color node is original work by Amatsukast

See LICENSE for details.

Model License

BiRefNet models by ZhengPeng7 are licensed under Apache License 2.0.

• Model repository: ZhengPeng7/BiRefNet
• You can use the models freely for commercial and non-commercial purposes

Summary

• ✅ You can use, modify, and distribute this code
• ✅ You can use the models commercially
• ⚠️ Modifications must also be GPL-3.0
• ⚠️ You must provide source code when distributing

---

Changelog

[1.0.0] - 2026-04-13

Added

• BiRefNet (RemoveBG) node with 5 model variants
  • BiRefNet-general: 1024px, balanced performance
  • BiRefNet-HR: 2048px, high quality
  • BiRefNet-portrait: 1024px, optimized for human subjects
  • BiRefNet_dynamic: variable resolution with aspect ratio preservation
  • BiRefNet_toonout: 1024px, outline extraction
• Find Unused Color node for chroma key color detection
• Automatic model download from HuggingFace
• Custom model storage in models/sbtools/BiRefNet/
• Support for transparent and colored backgrounds
• Mask refinement options (blur, offset, invert)

Optimizations

• Reduced model selection from 11 to 5 carefully chosen variants
• Improved aspect ratio handling for BiRefNet_dynamic
• Optimized processing resolution based on input size

Technical

• Modular package structure
• English-only codebase
• Comprehensive README documentation
• Version tracking in __init__.py

---

Note: This is a personal custom node collection. Use at your own discretion.