Custom Workflows - jmpijll/discomfy GitHub Wiki

Custom Workflows Guide 🎨

Learn how to create and integrate custom ComfyUI workflows with DisComfy.

📋 Overview

DisComfy supports custom ComfyUI workflows, allowing you to:

  • Use specialized AI models
  • Create unique generation styles
  • Implement custom processing pipelines
  • Integrate community workflows

✅ Workflow Requirements

Must-Have Properties

Every workflow node MUST have:

  1. class_type - The ComfyUI node type
  2. inputs - Dictionary of node inputs (can be empty)

Valid Node Structure

{
  "1": {
    "inputs": {
      "seed": 12345,
      "steps": 30
    },
    "class_type": "KSampler",
    "_meta": {
      "title": "KSampler"
    }
  }
}

Common class_type values:

  • KSampler - Sampling node
  • CLIPTextEncode - Text encoding
  • LoadImage - Image loading
  • VAEDecode - VAE decoding
  • SaveImage - Image output
  • And many more...

🚀 Creating Custom Workflows

Step 1: Design in ComfyUI

  1. Open ComfyUI in your browser
  2. Build your workflow using nodes
  3. Test thoroughly - ensure it works!
  4. Note important nodes:
    • Where prompts go (CLIPTextEncode)
    • Sampling nodes (KSampler)
    • Dimension nodes (EmptyLatentImage)
    • Output nodes (SaveImage)

Step 2: Export Workflow

CRITICAL: Use API Format!

❌ WRONG: Regular "Save"

File → Save

✅ CORRECT: API Format

File → Save (API Format)

Why API Format?

  • Regular save includes UI positions and metadata
  • API format includes only node structure
  • DisComfy requires API format for validation

Step 3: Validate Workflow

DisComfy v1.3.1+ automatically validates workflows and provides helpful error messages:

Workflow 'my_workflow' has 2 invalid node(s):
  - Node #1: Missing required 'class_type' property
  - Node #5: Missing required 'inputs' property

Common issues:
  1. Each node must have a 'class_type' property
  2. Each node must have an 'inputs' property
  3. Workflow must be exported as API format

To fix: In ComfyUI, use 'Save (API Format)'

Manual Validation:

# Check JSON syntax
python -m json.tool workflows/my_workflow.json

# Check for class_type
grep -o '"class_type"' workflows/my_workflow.json | wc -l
# Should match node count

Step 4: Save to DisComfy

# Copy to workflows directory
cp my_workflow.json discomfy/workflows/

Step 5: Configure in DisComfy

Edit config.json:

"workflows": {
  "my_custom": {
    "file": "my_workflow.json",
    "name": "My Custom Workflow",
    "description": "Specialized workflow for X",
    "enabled": true
  }
}

🔧 Integrating Workflows

Basic Integration

For workflows that work with standard parameters:

"workflows": {
  "artistic_style": {
    "file": "artistic_style.json",
    "name": "Artistic Style",
    "description": "Creates artistic style images",
    "enabled": true
  }
}

DisComfy will automatically update:

  • Prompts (CLIPTextEncode nodes)
  • Seeds (RandomNoise, KSampler)
  • Steps (BasicScheduler, KSampler)
  • Dimensions (EmptyLatentImage, EmptySD3LatentImage)

Advanced Integration

For custom parameters, add update logic to image_gen.py:

def _update_custom_workflow_parameters(
    self,
    workflow: Dict[str, Any],
    prompt: str,
    custom_param: float,
    seed: Optional[int] = None
) -> Dict[str, Any]:
    """Update custom workflow parameters."""
    updated_workflow = json.loads(json.dumps(workflow))
    
    if seed is None:
        seed = random.randint(0, 2**32 - 1)
    
    for node_id, node_data in updated_workflow.items():
        class_type = node_data.get('class_type')
        
        # Standard updates
        if class_type == 'KSampler':
            node_data['inputs']['seed'] = seed
        
        # Custom node updates
        elif class_type == 'CustomNode':
            node_data['inputs']['custom_param'] = custom_param
    
    return updated_workflow

📝 Common Workflow Patterns

Image Generation Workflow

Required nodes:

  1. Model loading (CheckpointLoader, UNETLoader)
  2. CLIP encoding (CLIPTextEncode) for positive/negative
  3. Latent creation (EmptyLatentImage)
  4. Sampling (KSampler)
  5. VAE decode (VAEDecode)
  6. Save (SaveImage)

Example structure:

CheckpointLoader → CLIPTextEncode (positive)
                → CLIPTextEncode (negative)
                → KSampler → VAEDecode → SaveImage
EmptyLatentImage →

Image Editing Workflow

Required nodes:

  1. Image loading (LoadImage)
  2. Model loading
  3. CLIP encoding for edit prompt
  4. Image encoding (VAEEncode)
  5. Sampling
  6. VAE decode
  7. Save

Example structure:

LoadImage → VAEEncode → KSampler → VAEDecode → SaveImage
                     ↑
CheckpointLoader → CLIPTextEncode

Upscaling Workflow

Required nodes:

  1. Image loading (LoadImage)
  2. Upscale node (ImageUpscaleWithModel)
  3. Optional: Refining with sampling
  4. Save

Example structure:

LoadImage → ImageUpscaleWithModel → SaveImage
         Optional refinement:
         → VAEEncode → KSampler → VAEDecode →

🎯 Node Identification

DisComfy updates nodes by class_type. Here's how to identify them:

Prompt Nodes (CLIPTextEncode)

DisComfy looks for nodes with title metadata:

{
  "6": {
    "inputs": {
      "text": "prompt here"
    },
    "class_type": "CLIPTextEncode",
    "_meta": {
      "title": "CLIP Text Encode (Prompt) - Positive"
    }
  }
}

Title keywords:

  • "Positive" → Updated with user prompt
  • "Negative" → Updated with negative prompt

Sampling Nodes

KSampler:

{
  "inputs": {
    "seed": 12345,      // ← Updated
    "steps": 30,        // ← Updated
    "cfg": 7.0          // ← Updated
  },
  "class_type": "KSampler"
}

RandomNoise (Flux):

{
  "inputs": {
    "noise_seed": 12345  // ← Updated
  },
  "class_type": "RandomNoise"
}

Dimension Nodes

EmptyLatentImage:

{
  "inputs": {
    "width": 1024,      // ← Updated
    "height": 1024,     // ← Updated
    "batch_size": 1     // ← Updated
  },
  "class_type": "EmptyLatentImage"
}

LoRA Nodes

LoraLoaderModelOnly:

{
  "inputs": {
    "lora_name": "style.safetensors",  // ← Updated
    "strength_model": 1.0               // ← Updated
  },
  "class_type": "LoraLoaderModelOnly"
}

🔍 Debugging Workflows

Common Issues

Issue 1: Workflow validation fails

Error: Node #1: Missing required 'class_type' property

Solution: Re-export using "Save (API Format)"

Issue 2: Prompts not applied

Generated image doesn't match prompt

Solution: Check CLIPTextEncode nodes have title metadata

Issue 3: Wrong dimensions

Generated image has wrong size

Solution: Verify EmptyLatentImage nodes exist

Issue 4: LoRAs not working

LoRA not applied to generation

Solution: Check LoraLoaderModelOnly node exists

Debug Logging

Enable debug logging to see workflow processing:

// config.json
"logging": {
  "level": "DEBUG"
}

Check logs for:

DEBUG - Loaded workflow: my_custom
DEBUG - Updated workflow parameters: prompt='...', size=1024x1024
DEBUG - Found 5 KSampler nodes
DEBUG - Updated node 3: seed=12345

Testing Workflow

  1. Test in ComfyUI first - Ensure it works standalone
  2. Start with simple parameters - Don't use LoRAs initially
  3. Check logs - Look for errors or warnings
  4. Test incrementally - Add features one at a time

📚 Example Workflows

Simple Flux Workflow

{
  "10": {
    "inputs": {
      "vae_name": "FLUX1/ae.safetensors"
    },
    "class_type": "VAELoader"
  },
  "11": {
    "inputs": {
      "clip_name1": "ViT-L-14.safetensors",
      "clip_name2": "t5-xxl.safetensors",
      "type": "flux"
    },
    "class_type": "DualCLIPLoader"
  },
  "6": {
    "inputs": {
      "text": "prompt here",
      "clip": ["11", 0]
    },
    "class_type": "CLIPTextEncode",
    "_meta": {
      "title": "Positive Prompt"
    }
  },
  "25": {
    "inputs": {
      "noise_seed": 12345
    },
    "class_type": "RandomNoise"
  },
  "13": {
    "inputs": {
      "noise": ["25", 0],
      "sampler": ["16", 0],
      "sigmas": ["17", 0]
    },
    "class_type": "SamplerCustomAdvanced"
  }
}

Image Edit Workflow

{
  "41": {
    "inputs": {
      "image": "input.png"
    },
    "class_type": "LoadImage"
  },
  "6": {
    "inputs": {
      "text": "edit instructions",
      "clip": ["11", 0]
    },
    "class_type": "CLIPTextEncode",
    "_meta": {
      "title": "Positive"
    }
  },
  "3": {
    "inputs": {
      "seed": 12345,
      "steps": 20,
      "cfg": 2.5,
      "positive": ["6", 0],
      "latent_image": ["47", 0]
    },
    "class_type": "KSampler"
  }
}

💡 Best Practices

Workflow Design

  1. Keep it simple - Complex workflows are harder to debug
  2. Use clear titles - Add descriptive _meta.title to nodes
  3. Test thoroughly - Ensure it works in ComfyUI first
  4. Document nodes - Comment which nodes do what
  5. Version control - Keep old versions for reference

Node Organization

  1. Group related nodes - Use similar node IDs
  2. Name consistently - Use clear, descriptive titles
  3. Avoid hardcoded values - Use nodes for configurable values
  4. Test edge cases - Try extreme values

Integration

  1. Start simple - Basic parameters first
  2. Add features incrementally - One at a time
  3. Validate input - Check parameters make sense
  4. Handle errors gracefully - Don't crash on bad input
  5. Document usage - Add examples and notes

🔗 Resources

ComfyUI Documentation

DisComfy Examples

  • Check workflows/ directory for working examples
  • See image_gen.py for parameter update logic
  • Review CHANGELOG.md for workflow updates

🎯 Next Steps

🎨 Ready to create? Export your ComfyUI workflow and integrate it!