Best Practices for Custom Nodes

Follow these best practices to create reliable, performant, and user-friendly custom nodes for ComfyUI.

Code Organization

File Structure

Organize your extension with a clear directory structure:

my-custom-nodes/
├── __init__.py              # Extension entry point
├── nodes/
│   ├── __init__.py
│   ├── image_processing.py  # Image nodes
│   └── text_processing.py   # Text nodes
├── web/                     # Frontend extensions
│   └── my_extension.js
├── example_workflows/       # Example workflows
│   └── example.json
├── locales/                 # Translations
│   └── en/
│       └── main.json
├── requirements.txt         # Python dependencies
└── README.md

Module Naming

Use descriptive, unique names to avoid conflicts:

class MyExtension_ImageBlur(io.ComfyNode):
    @classmethod
    def define_schema(cls) -> io.Schema:
        return io.Schema(
            node_id="MyExtension_ImageBlur",
            display_name="Image Blur",
            category="MyExtension/Image",
            ...
        )

Performance Optimization

Use Lazy Inputs

Implement lazy evaluation for expensive operations:

class ExpensiveNode(io.ComfyNode):
    @classmethod
    def define_schema(cls) -> io.Schema:
        return io.Schema(
            node_id="ExpensiveNode",
            inputs=[
                io.Image.Input("image"),
                io.Boolean.Input("enable_processing"),
                io.Int.Input(
                    "iterations",
                    lazy=True,  # Only evaluate if needed
                    default=10
                ),
            ],
            outputs=[io.Image.Output()],
        )
    
    @classmethod
    def check_lazy_status(cls, image, enable_processing, iterations):
        # Only request iterations if processing is enabled
        if enable_processing:
            return ["iterations"]
        return []
    
    @classmethod
    def execute(cls, image, enable_processing, iterations):
        if not enable_processing:
            return io.NodeOutput(image)
        
        # iterations is guaranteed to be available here
        for _ in range(iterations):
            image = process(image)
        
        return io.NodeOutput(image)

Implement Fingerprinting

Avoid re-execution when inputs haven’t meaningfully changed:

import hashlib

class ImageLoader(io.ComfyNode):
    @classmethod
    def fingerprint_inputs(cls, image_path):
        # Only reload if file actually changed
        import os
        if os.path.exists(image_path):
            stat = os.stat(image_path)
            return f"{image_path}:{stat.st_mtime}:{stat.st_size}"
        return image_path

Memory Management

Be mindful of GPU memory:

import torch

class MemoryEfficientNode(io.ComfyNode):
    @classmethod
    def execute(cls, large_tensor):
        # Process in smaller chunks
        batch_size = 8
        results = []
        
        for i in range(0, len(large_tensor), batch_size):
            batch = large_tensor[i:i+batch_size]
            result = process_batch(batch)
            results.append(result.cpu())  # Move to CPU
            
            # Clear GPU memory
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
        
        return io.NodeOutput(torch.cat(results).cuda())

User Experience

Provide Helpful Tooltips

Add clear, descriptive tooltips:

io.Float.Input(
    "strength",
    min=0.0,
    max=2.0,
    default=1.0,
    tooltip="Controls the strength of the effect. "
            "Values > 1.0 increase intensity, < 1.0 decrease it."
)

Use Appropriate Input Types

Choose the right input type and display mode:

io.Float.Input(
    "blur_radius",
    min=0.0,
    max=100.0,
    step=0.1,
    display_mode=io.NumberDisplay.slider,  # Visual feedback
    tooltip="Blur radius in pixels"
)

Sensible Defaults

Provide good default values:

io.Schema(
    node_id="ImageResize",
    inputs=[
        io.Image.Input("image"),
        io.Int.Input(
            "width",
            default=512,  # Common size
            min=64,
            max=4096,
            step=64
        ),
        io.Int.Input(
            "height",
            default=512,
            min=64,
            max=4096,
            step=64
        ),
        io.Combo.Input(
            "interpolation",
            options=["bicubic", "bilinear", "nearest"],
            default="bicubic"  # Best quality
        ),
    ],
    outputs=[io.Image.Output()],
)

Error Handling

Validate Inputs

Check inputs and provide clear error messages:

class SafeNode(io.ComfyNode):
    @classmethod
    def execute(cls, image, scale_factor):
        if scale_factor <= 0:
            raise ValueError(
                f"scale_factor must be positive, got {scale_factor}"
            )
        
        if image.shape[0] == 0:
            raise ValueError("Input image batch is empty")
        
        # Validate image dimensions
        if image.shape[1] < 8 or image.shape[2] < 8:
            raise ValueError(
                f"Image too small: {image.shape[1]}x{image.shape[2]}. "
                f"Minimum size is 8x8"
            )
        
        result = scale_image(image, scale_factor)
        return io.NodeOutput(result)

Handle Edge Cases

Consider edge cases in your implementation:

@classmethod
def execute(cls, images, mask=None):
    # Handle optional mask
    if mask is None:
        mask = torch.ones_like(images[:, :, :, 0])
    
    # Handle dimension mismatch
    if mask.shape[1:] != images.shape[1:3]:
        import torch.nn.functional as F
        mask = F.interpolate(
            mask.unsqueeze(1),
            size=(images.shape[1], images.shape[2]),
            mode='bilinear'
        ).squeeze(1)
    
    # Handle empty batch
    if images.shape[0] == 0:
        return io.NodeOutput(images)
    
    result = apply_mask(images, mask)
    return io.NodeOutput(result)

Documentation

Docstrings

Document your nodes with clear docstrings:

class ImageBlend(io.ComfyNode):
    """
    Blends two images together using various blend modes.
    
    Supports multiple blend modes including normal, multiply, screen,
    and overlay. The blend strength can be controlled with the opacity
    parameter.
    
    The images must have the same dimensions. If they differ, the second
    image will be automatically resized to match the first.
    """
    
    @classmethod
    def define_schema(cls) -> io.Schema:
        ...

Example Workflows

Provide example workflows in example_workflows/:

{
  "last_node_id": 3,
  "nodes": [
    {
      "id": 1,
      "type": "LoadImage",
      "outputs": {"IMAGE": {"links": [1]}}
    },
    {
      "id": 2,
      "type": "MyExtension_ImageBlur",
      "inputs": {"image": {"link": 1}},
      "widgets_values": [5.0]
    }
  ]
}

README

Include comprehensive README:

# My Custom Nodes

## Installation

```bash
cd ComfyUI/custom_nodes/
git clone https://github.com/username/my-custom-nodes
cd my-custom-nodes
pip install -r requirements.txt

Nodes

Image Blur

Blurs an image using Gaussian blur. Inputs:

image: Input image
radius: Blur radius (0-100)

Outputs:

image: Blurred image

## Testing

### Test Your Nodes

Create unit tests for your nodes:

```python
import torch
import pytest
from .nodes import ImageBlur

def test_image_blur():
    # Create test image
    image = torch.rand(1, 512, 512, 3)
    
    # Execute node
    result = ImageBlur.execute(image=image, radius=5.0)
    
    # Verify output
    assert result[0].shape == image.shape
    assert result[0].min() >= 0.0
    assert result[0].max() <= 1.0

def test_edge_cases():
    # Test with minimal image
    tiny = torch.rand(1, 8, 8, 3)
    result = ImageBlur.execute(image=tiny, radius=1.0)
    assert result[0].shape == tiny.shape
    
    # Test with large radius
    image = torch.rand(1, 256, 256, 3)
    result = ImageBlur.execute(image=image, radius=100.0)
    assert result[0].shape == image.shape

Versioning and Compatibility

Version Your Extension

Use semantic versioning:

__version__ = "1.2.0"

class MyExtension(ComfyExtension):
    async def on_load(self) -> None:
        print(f"Loading My Extension v{__version__}")

Maintain Backwards Compatibility

When updating nodes, preserve compatibility:

class ImageProcessor(io.ComfyNode):
    @classmethod
    def define_schema(cls) -> io.Schema:
        return io.Schema(
            inputs=[
                io.Image.Input("image"),
                # New parameter with default
                io.Combo.Input(
                    "mode",
                    options=["fast", "quality"],
                    default="fast"  # Maintains old behavior
                ),
            ],
            ...
        )

Common Patterns

Progress Reporting

Report progress for long operations:

from comfy_api.latest import ComfyAPISync

class LongRunningNode(io.ComfyNode):
    @classmethod
    def execute(cls, image, iterations):
        api = ComfyAPISync()
        results = []
        
        for i in range(iterations):
            result = process_iteration(image, i)
            results.append(result)
            
            # Update progress
            api.execution.set_progress(
                value=i + 1,
                max_value=iterations,
                preview_image=result  # Show preview
            )
        
        return io.NodeOutput(torch.stack(results))

Batch Processing

Handle batch inputs efficiently:

@classmethod
def execute(cls, images):
    # Process entire batch at once (efficient)
    results = batch_process(images)
    
    # Or process individually if needed
    # results = [process_single(img) for img in images]
    # results = torch.stack(results)
    
    return io.NodeOutput(results)

Always test your nodes with different batch sizes to ensure correct behavior.

Security Considerations

Validate File Paths

Never trust user-provided paths:

import os

def safe_load_file(file_path, allowed_dir):
    # Resolve to absolute path
    abs_path = os.path.abspath(file_path)
    abs_allowed = os.path.abspath(allowed_dir)
    
    # Check if path is within allowed directory
    if not abs_path.startswith(abs_allowed):
        raise ValueError(f"Access denied: {file_path}")
    
    return abs_path

Sanitize Inputs

Validate and sanitize all user inputs:

import re

@classmethod
def execute(cls, filename, content):
    # Sanitize filename
    safe_name = re.sub(r'[^a-zA-Z0-9_.-]', '', filename)
    if not safe_name:
        raise ValueError("Invalid filename")
    
    # Limit content size
    max_size = 10 * 1024 * 1024  # 10MB
    if len(content) > max_size:
        raise ValueError("Content too large")
    
    # Process safely
    ...

Follow these best practices to create nodes that are fast, reliable, and user-friendly!

Custom Nodes

ComfyUI Manager

Best Practices for Custom Nodes

Code Organization

File Structure

Module Naming

Performance Optimization

Use Lazy Inputs

Implement Fingerprinting

Memory Management

User Experience

Provide Helpful Tooltips

Use Appropriate Input Types

Sensible Defaults

Error Handling

Validate Inputs

Handle Edge Cases

Documentation

Docstrings

Example Workflows

README

Nodes

Image Blur

Versioning and Compatibility

Version Your Extension

Maintain Backwards Compatibility

Common Patterns

Progress Reporting

Batch Processing

Security Considerations

Validate File Paths

Sanitize Inputs

Build docs developers (and LLMs) love

Custom Nodes

ComfyUI Manager

​Code Organization

​File Structure

​Module Naming

​Performance Optimization

​Use Lazy Inputs

​Implement Fingerprinting

​Memory Management

​User Experience

​Provide Helpful Tooltips

​Use Appropriate Input Types

​Sensible Defaults

​Error Handling

​Validate Inputs

​Handle Edge Cases

​Documentation

​Docstrings

​Example Workflows

​README

​Nodes

​Image Blur

​Versioning and Compatibility

​Version Your Extension

​Maintain Backwards Compatibility

​Common Patterns

​Progress Reporting

​Batch Processing

​Security Considerations

​Validate File Paths

​Sanitize Inputs

Build docs developers (and LLMs) love

Code Organization

File Structure

Module Naming

Performance Optimization

Use Lazy Inputs

Implement Fingerprinting

Memory Management

User Experience

Provide Helpful Tooltips

Use Appropriate Input Types

Sensible Defaults

Error Handling

Validate Inputs

Handle Edge Cases

Documentation

Docstrings

Example Workflows

README

Nodes

Image Blur

Versioning and Compatibility

Version Your Extension

Maintain Backwards Compatibility

Common Patterns

Progress Reporting

Batch Processing

Security Considerations

Validate File Paths

Sanitize Inputs