1. Workflow Overview

This workflow, named "AI Portrait Workflow", is designed for high-quality portrait generation and editing, integrating AI generation, face detection, inpainting, and style control.

• Core Models:

  • Stable Diffusion XL (SDXL): LEOSAM HelloWorld _ SDXL_v7.0 (base model).

  • ControlNet: xinsir_union_promax.safetensors (pose/structure control).

  • IPAdapter: Style transfer and detail enhancement.

  • Florence-2: Face detection and segmentation.

• Functions:

  • Automatic face detection and masking.

  • Multi-ControlNet integration (pose, depth, edges).

  • Background replacement and style fusion.

  • Low-strength inpainting for detail refinement.

2. Key Nodes

3. Workflow Structure

Key functional groups:

1. Face Detection & Masking:

   • Florence2Run detects faces → Sam2Segmentation refines masks.

2. ControlNet Control:

   • ControlNetApplySD3 applies pose/depth constraints.

3. IPAdapter Style Transfer:

   • Transfers style from reference images (e.g., "Cyberpunk Girl").

4. Inpainting & Refinement:

   • Light inpainting + PS masking for detail fixes.

4. Inputs & Outputs

• Inputs:

  • Source image (e.g., portrait).

  • Optional prompts (default uses Florence-2 captions).

  • Control parameters (e.g., ControlNet weight).

• Outputs:

  • Final portrait with stylized background.

  • Side-by-side comparisons via Image Comparer.

5. Notes

• Model Dependencies:

  • Download SDXL, ControlNet, and Florence-2 (total >20GB).

  • Model links:

    • SDXL: LEOSAM HelloWorld

    • ControlNet: xinsir_union_promax

• Hardware:

  • ≥12GB VRAM (RTX 3080+), 24GB recommended for full pipeline.

• Troubleshooting:

  • Poor masking: Adjust Sam2Segmentation bboxes or manual coordinates.

  • Over-stylization: Reduce weight in IPAdapterAdvanced (default: 0.7).