1. Workflow Overview

• Purpose
  This workflow is designed for highly customizable portrait regeneration, allowing one-click modifications to nationality, age, art style, details, and facial expressions. Key features:

  • Generate new portraits based on input images (preserving composition).

  • Adjust attributes via prompts and LoRAs (e.g., "Belarusian 20-years-old").

  • Control realism/artistic style and fine details (skin texture, pupils, etc.).

• Core Models

  1. Stable Diffusion (Flux.1-Dev fp8): Base model with integrated CLIP & VAE.

  2. LoRAs:

     • F.1 Asian Portrait Photography_v1.0: Optimized for Asian faces.

     • Other LoRAs (disabled): Can be loaded dynamically via UI.

  3. WD14Tagger: Auto-tagging input images for prompt generation.

2. Key Nodes

• PortraitMaster_CN: Central panel for nationality, age, and detail control.

• CLIPTextEncodeFlux: Uses T5XXL for advanced text encoding.

• KSampler: Generates images with euler sampler (20 steps).

• WD14Tagger|pysssss: Analyzes input images to generate tags (e.g., "white shirt").

• Dependencies

  • Models:

    • Place F.1 Compact in models/unet/.

    • LoRAs in models/loras/.

  • Plugins:

    • Install pysssss nodes (e.g., WD14Tagger) via ComfyUI Manager.

3. Workflow Structure

• Group 1: Input Processing

  • Input: Portrait image (e.g., 小红书网页版.jpg).

  • Output: Tags from WD14Tagger + manual prompts.

• Group 2: Parameter Tuning

  • Configure nationality, age, and skin details via PortraitMaster_CN.

• Group 3: Image Generation

  • Output: 1024x1536 portrait.

4. Inputs & Outputs

• Inputs:

  • Image: Drag & drop to LoadImage.

  • Prompts: e.g., raw photo, realistic, belarusian 20-years-old.

  • Resolution: 1024x1536 (set in EmptyLatentImage).

• Output:

  • High-realistic portrait (PNG).

5. Notes

• Troubleshooting:

  • Missing F.1 Compact? Download from HuggingFace.

• Performance:

  • Requires ≥8GB VRAM; fp8 models recommended.

• Compatibility:

  • CLIPTextEncodeFlux requires Flux.1-specific CLIP models.