1. Workflow Overview

This is a PulID Flux-based face swapping workflow with two modes:

1. Text-to-Image: Generate 3D avatars from text descriptions

2. Image-to-Image: Convert real faces to 3D cartoon style
   Core Pipeline: Face detection → Style transfer → Image composition

2. Core Models

3. Key Components

3.1 Custom Nodes

- `PulID Flux Nodes`: Install from GitHub repo
- `JoyCaption`: Requires LLM model (~4GB VRAM)
- `FluxGuidance`: Dynamic CFG control module

3.2 Critical Nodes

• ApplyPulidFlux (ID:1): Core face-swapping node

• SamplerCustomAdvanced (ID:35/16): Enhanced sampler with noise control

• ConcatTextOfUtils (ID:50): Dynamic prompt concatenation

4. Workflow Structure

Group 1: Text-to-Image

• Input: Text prompt + random seed

• Process: CLIP encoding → 3D generation → Face fusion

• Output: 1024x768 3D avatar

Group 2: Image-to-Image

• Input: Photo + optional mask

• Process: Face detection → Style transfer

• Output: 3D-style converted image

5. I/O Specifications

Input Parameters:

{
  "text_prompt": "3D,cartoon,school uniform", # Required  
  "seed": -1,                        # Random seed  
  "reference_image": "optional.jpg"   # For img2img mode  
}

Output: PNG with metadata

6. Important Notes

1. Hardware Requirements:

   • Minimum 8GB VRAM (12GB for Llama)

   • CUDA acceleration required

2. Troubleshooting:

   • Face detection failure → Check InsightFace model

   • Style mismatch → Adjust Lora strength (0.7-1.2 recommended)

3. Optimization Tips:

   • Use attn_mask for precise face replacement

   • Disable live preview for batch processing