1. Workflow Overview

• Purpose: Convert anime/2D images into photorealistic portraits with automatic face/hand refinement.

• Core Models:

  • majicMIX realistic v7: Base photorealism model

  • ControlNet (Canny): Preserve original composition

  • UltralyticsDetector (YOLOv8): Face/body detection

  • ESRGAN_4x: Image upscaling

2. Key Nodes

3. Workflow Structure

• Group 1: Input Processing

  • Input: Anime image (e.g., initial_song.jpg)

  • Output: Line art + tags (blue_hair, dress, etc.)

• Group 2: Image Generation

  • Uses ControlNet with KSamplerAdvanced (40 steps, Euler, CFG=8)

• Group 3: Refinement

  • FaceDetailer for face/hands (manual mask editing available)

• Group 4: Upscaling

  • 2x ESRGAN upscale to ≥2048px

4. Input/Output

• Inputs:

  • Image: Recommended 1024x1536px

  • Negative prompt: Auto-includes (worst quality:2), lowres

• Output: PNG portrait saved to ComfyUI/output

5. Notes

• VRAM: ≥12GB required (peak usage during refinement)

• Dependencies:

  • SAM model sam_vit_b_01ec64.pth in models/sam

  • Chinese prompts auto-translated via BaiduTranslateNode

• Troubleshooting:

  • Line art failure → Adjust image contrast

  • Distorted hands → Enable Hand Repair Component group