Skip to content

Latest commit

 

History

History
221 lines (177 loc) · 6.84 KB

File metadata and controls

221 lines (177 loc) · 6.84 KB

Master Validation Workflow - Complete Control Guide

The Master Validation Workflow combines all quality validation methods into a single, UI-controllable workflow. This gives you complete control over every aspect of generation directly from the ComfyUI interface.

🎛️ Workflow Overview

Single workflow name: {persona-id}_UI_MASTER_CONTROL

Generates: Up to 8 output versions (4 methods × 2 versions each)

  • Original quality versions
  • 2x upscaled versions with enhanced detail

🔧 UI Control Panel Reference

Node 2: LoRA Strength Control

🎛️ strength_model: 0.65    # Persona feature strength
🎛️ strength_clip: 0.65     # Prompt influence strength

Recommended values:

  • 0.5-0.6: Very natural, subtle persona features
  • 0.65-0.7: Balanced realism and recognition
  • 0.75-0.8: Strong persona features, less natural
  • 0.85+: Maximum features, may look artificial

Node 7: Resolution Control

🎛️ width: 1152             # Image width
🎛️ height: 1152            # Image height  
🎛️ batch_size: 1           # Number of images per method

Recommended sizes:

  • 1024×1024: Standard quality, faster
  • 1152×1152: Higher detail, slower
  • 1280×1280: Maximum detail, much slower

Nodes 3,4,5: Prompt Style Controls

Node 3 - Basic Prompt (Simple generations)

🎛️ text: "portrait of {trigger_word}"

Node 4 - Photorealistic Prompt (⭐ RECOMMENDED)

🎛️ text: "RAW photograph, {trigger_word}, photorealistic, highly detailed, sharp focus, natural lighting, 85mm lens..."

Node 5 - Ultra-Realistic Prompt (Maximum realism)

🎛️ text: "award-winning professional photograph of {trigger_word}, hyperrealistic, extreme detail, natural skin pores..."

Node 6: Negative Prompt Control

🎛️ text: "cartoon, anime, illustration, painting..."  # Add more terms as needed

Nodes 10-13: Sampling Method Controls

Method 1 - Ultra Conservative (Node 10) - Most Natural

🎛️ seed: 12345              # Change for variations
🎛️ steps: 45               # SET TO 1 TO DISABLE METHOD
🎛️ cfg: 5.5                # Lower = more natural
🎛️ sampler_name: "dpmpp_2m_sde"
🎛️ scheduler: "karras"
🎛️ positive: ["5", 0]       # Which prompt to use (3,4, or 5)

Method 2 - Professional (Node 11) - Balanced

🎛️ seed: 23456              # Different seed
🎛️ steps: 40               # SET TO 1 TO DISABLE METHOD
🎛️ cfg: 6.0                # Balanced setting
🎛️ sampler_name: "dpmpp_2m"
🎛️ positive: ["4", 0]       # Photorealistic prompt

Method 3 - Alternative (Node 12) - Backup

🎛️ steps: 35               # SET TO 1 TO DISABLE METHOD
🎛️ cfg: 6.5                
🎛️ sampler_name: "euler_a"

Method 4 - High Detail (Node 13) - Maximum Quality

🎛️ steps: 50               # SET TO 1 TO DISABLE METHOD
🎛️ cfg: 6.0                
🎛️ sampler_name: "dpmpp_2m_sde"

Nodes 40-43: Upscaling Controls

🎛️ upscale_model: "RealESRGAN_x2plus.pth"    # 2x upscaling
🎛️ upscale_model: "RealESRGAN_x4plus.pth"    # 4x upscaling  
🎛️ upscale_model: ""                         # DISABLE upscaling

🎯 Usage Scenarios

Scenario 1: Quick Quality Test

  1. Disable all methods except Method 1 (set steps=1 for others)
  2. Set LoRA strength to 0.65
  3. Use photorealistic prompt (Node 4)
  4. Generate and evaluate

Scenario 2: Strength Comparison

  1. Enable only Method 1 and Method 2
  2. Set different LoRA strengths: Node 2 = 0.6, duplicate and set to 0.7
  3. Same seed for both
  4. Compare results

Scenario 3: Full Validation Run

  1. Enable all 4 methods
  2. Use different seeds (12345, 23456, 34567, 45678)
  3. Generate 8 versions total
  4. Compare and identify best method

Scenario 4: Prompt Style Testing

  1. Enable Method 1 only
  2. Change the positive connection: Try [3,0], [4,0], [5,0]
  3. Compare prompt styles
  4. Use best-performing prompt

🔄 Workflow Optimization Process

Step 1: Initial Test

✓ Enable Method 1 only (steps=45, others=1)
✓ LoRA strength = 0.65
✓ Photorealistic prompt (Node 4)
✓ Generate and evaluate

Step 2: Strength Optimization

✓ Try LoRA strengths: 0.6, 0.65, 0.7
✓ Keep same seed for comparison
✓ Find most natural-looking strength

Step 3: Method Comparison

✓ Enable all methods with optimal LoRA strength
✓ Use same seed across all methods
✓ Identify best sampling method

Step 4: Final Optimization

✓ Use best method + strength + prompt combination
✓ Test with different scenarios/poses
✓ Document optimal settings for future use

📊 Output Interpretation

File Naming Convention:

  • {persona_id}_METHOD_1_ULTRA_CONSERVATIVE - Most natural approach
  • {persona_id}_METHOD_2_PROFESSIONAL - Balanced detail/realism
  • {persona_id}_METHOD_3_ALTERNATIVE - Different sampling approach
  • {persona_id}_METHOD_4_HIGH_DETAIL - Maximum feature emphasis
  • _2X suffix - 2x upscaled versions

Quality Assessment:

Rate each output on:

  • Natural skin texture (not plastic/smooth)
  • Realistic lighting (natural shadows)
  • Facial feature accuracy (matches training data)
  • Expression authenticity (human-like)
  • Overall photographic quality (camera-like)

Power User Controls

Advanced Prompt Engineering in UI:

# Node 4 - Modify for specific scenarios:
"RAW photograph, {trigger_word}, [SCENARIO], photorealistic, highly detailed, sharp focus, [LIGHTING], 85mm lens, depth of field, natural skin texture, detailed facial features, realistic, professional photography"

# Examples:
[SCENARIO]: "business meeting", "casual portrait", "outdoor environment"
[LIGHTING]: "natural lighting", "studio lighting", "golden hour lighting"

Advanced Sampling Tweaks:

# For more natural look:
cfg: 5.0-6.0, scheduler: "karras"

# For more detail:  
cfg: 6.5-7.0, scheduler: "exponential"

# For consistency:
Use same seed across all enabled methods

Batch Processing:

# Node 7 - batch_size: 4
# Generates 4 versions of each enabled method
# Total outputs: 4 × enabled_methods × 2 (upscaled)

🚀 One-Command Setup

To set up everything for maximum persona realism:

# Complete setup with master workflow
./scripts/setup_full_validation.sh persona-saramillie

This creates the master workflow plus all supporting tools for comprehensive persona quality validation and optimization.

📖 Related Guides