1. Workflow Overview

This is a material/style transfer pipeline featuring:

• Advanced facial & color transfer via IPAdapterFaceIDKolors

• Structure preservation with ControlNet line art

• Enhanced detail retention using CLIP vision encoding + InsightFace

• High-quality output with target material properties

2. Core Models

3. Key Components

Required Custom Nodes:

1. IPAdapter Suite

   • Includes IPAdapterFaceIDKolors, IPAdapterNoise etc.

   • Install via ComfyUI Manager (search IPAdapter-Plus)

2. InsightFace Loader

   • Requires additional antelopev2 model file

4. Pipeline Stages

Stage 1: Preprocessing

• Inputs:

  • Source image: 花瓣素材_珠宝系列...jpg (material reference)

  • Target image: 97d3bd57...ycxIpG.jpg (content reference)

• Key Operations:

  • PrepImageForClipVision: Normalization

  • LineArtPreprocessor: Line art generation

Stage 2: Feature Fusion

• Core Technology:

  • IPAdapterFaceIDKolors:

    • Strength=1.2 / Steps=2 / Blend mode="linear"

    • "K+mean(V) w/ C penalty" algorithm for color retention

  • ControlNetApplyAdvanced: Full-weight line art control

Stage 3: Generation

• Sampling:

  • 30 steps DPM++ 2M Karras

  • Resolution 512x1024

• Output: Material-transferred image

5. Input/Output

Input Requirements:

• Minimum 2 images:

  • Source (style/material reference)

  • Target (content reference)

• Prompt: Simple (e.g. "4k") + Negative prompt "Fuzzy, low quality"

Output:

• Generated image (auto-saved to ComfyUI/output)

6. Critical Notes

1. Hardware:

   • ≥8GB VRAM required (IPAdapterFaceIDKolors is resource-intensive)

   • InsightFace works better on GPU

2. Troubleshooting:

   • Adjust strength to 0.8-1.2 if face distortion occurs

   • Verify ControlNet model loading if line art fails

3. Extensions:

   • Add Detailer node for post-processing face refinement

   • Experiment with CLIP vision models (e.g. ViT-L/14)