1. Workflow Overview

This is a professional-grade character-scene fusion pipeline featuring:

• Dual-model architecture: SDXL for scenes + FLUX for characters

• Precise composition control via ControlNet

• Auto-segmentation with SAM+GroundingDINO

• Advanced color matching and pixel-level compositing

• HD output with comparison tools

2. Core Models

3. Key Components

Required Custom Nodes:

1. ComfyUI-Impact-Pack

   • Provides GroundingDinoSAMSegment node

   • Install via ComfyUI Manager

2. ImageColorMatch+

   • Advanced color matching (manual install)

   • Critical params: RGB mode/strength=0.35/GPU acceleration

4. Pipeline Stages

Stage 1: Scene Processing

• Input: ControlNet reference image + auto-tags

• Key params: SDXL/20 steps/Euler ancestral/CFG=0.7

Stage 2: Character Generation

• Input: Prompt "woman nude in room"

• Key params: FLUX model/low CFG(0.4) for soft details

Stage 3: Smart Compositing

• Core techniques:

  1. Text-guided segmentation ("human body")

  2. LAB color space matching

  3. Multi-layer blending (feathered edges)

5. Input/Output

Input Requirements:

• Scene reference image (1024x1536 recommended)

• Character image (clear outline required)

Outputs:

• Final composite (PNG)

• Comparison view (slider/split-screen)

6. Critical Notes

1. Hardware:

   • Min VRAM: 8GB (FP16)

   • Recommended: 12GB (FP8)

2. Troubleshooting:

   • Download GroundingDINO model to models/grounding-dino if missing

   • Adjust gamma in ImageColorMatch+ for color correction

3. Optimization:

   • Use LCM sampler for faster scene generation

   • Reduce character KSampler to 15 steps