1. Workflow Overview

This workflow specializes in facial photo retouching with:

• Automatic defect repair: Oiliness, acne, scars

• Texture preservation: Enhances skin details without over-smoothing

• Localized processing: Only modifies facial areas

Core Models:

• TruePortrait XL Apex: Realistic portrait base model

• Skin Texture Lora (weight 0.6): Optimizes pores and skin quality

• YOLOv8m-seg: Face detection model

2. Key Components

Required Custom Nodes:

1. Impact Pack (via ComfyUI Manager):

   • Provides FaceDetailer node

2. Ultralytics Detector:

   • Download person_yolov8m-seg.pt

3. Segment Anything (SAM):

   • Required model: sam_vit_b_01ec64.pth

Critical Parameters:

• SEGSRangeFilterDetailerHookProvider:

  • Limits processing area width ≤512px

• Detailer sampling:

  • 30 steps, CFG=5, euler sampler

3. Workflow Structure

Group Logic:

1. Input Group:

   • Node: LoadImage (e.g., oily_face_sample.jpg)

2. Detection Group:

   • Nodes: UltralyticsDetectorProvider + SAMLoader

   • Function: Precise face segmentation

3. Retouching Group:

   • Core Node: FaceDetailer

   • Key Settings:

     Denoise strength: 0.4  
     Mask expansion: 10px  
     Refinement passes: 3  

4. Output Group:

   • Node: SaveImage

   • Result: Retouched PNG with natural skin texture

4. Input & Output

Input Requirements:

• Formats: JPG/PNG

• Optimal Resolution: ≥1024px (short edge)

Output Features:

• Removes oiliness/acne/scars

• Preserves realistic skin pores

5. Usage Notes

1. Hardware:

   • Recommended ≥8GB VRAM (SAM is resource-intensive)

2. Troubleshooting:

   • Background artifacts: Adjust SEGSRangeFilter width threshold

   • Over-smoothed skin: Reduce Lora weight (default 0.6→0.4)

3. Customization:

   • Swap Lora for styles (e.g., "Midnight|Flux Xiaohongshu Photography_V1")