Generate Breathtaking Asian-Style Female Portraits with AI

CN
ComfyUI.org
2025-04-22 14:09:38

1. Workflow Overview

m9skz0d58yvhthogk5m69534f73fd06ca753e26dbc4aa3042bcbd1cf6068bea209a9a1168ab1746031a.png
  • Purpose:
    Generates high-quality oriental-style female portraits from a reference image, combining image analysis, auto-prompting, multi-stage sampling, and post-processing.

  • Core Models:

    • Stable Diffusion (F.1_fp16): Base model for image generation.

    • AWPortrait CN_2.0 LoRA: Fine-tunes for Asian portrait aesthetics.

    • Florence-2-base: Analyzes reference images to generate detailed captions.

    • ControlNet (Implicit): Enhances details via tiled sampling (TTP toolkit).

2. Key Nodes

  • Critical Components:

    • Florence2Run: Analyzes reference images and outputs captions.

    • TTP_Image_Tile_Batch: Splits images for high-frequency detail refinement.

    • ImageColorMatch+: Matches color tones to the reference.

    • VAEDecodeTiled: Decodes latent images in tiles to save VRAM.

  • Dependencies:

    • Plugins:

      • comfyui-florence2 (for Florence-2 integration).

      • comfyui_ttp_toolset (tiled processing).

    • Models:

      • Download Florence-2-base from HuggingFace.

      • LoRAs: AWPortrait CN_2.0 and XLabs F.1 Realism LoRA_V1.

3. Workflow Structure

  • Groups:

    1. Reference Analysis:

      • Input: Reference image.

      • Output: Auto-generated prompts via Florence-2.

    2. Primary Sampling:

      • Input: Prompts, base model, LoRAs.

      • Output: Initial latent image.

    3. TTP Upscaling:

      • Input: Tiled image, high/low-frequency sigma splits.

      • Output: High-res image with enhanced details.

    4. Color Adjustment:

      • Input: Generated image + reference.

      • Output: Final color-matched portrait.

4. Inputs & Outputs

  • Inputs:

    • Reference image (e.g., output(1).png).

    • Optional manual prompts.

    • Resolution (default 1024x1024, supports 9:16 portrait).

    • Seed (e.g., 459202764244119).

  • Output:

    • Final portrait (via PreviewImage node).

5. Tips & Warnings

  • VRAM: ≥12GB GPU recommended; tiling reduces memory usage.

  • Installation: Use ComfyUI Manager for Florence-2 and TTP Toolset.

  • Debugging:

    • Missing nodes? Check comfyui-layerstyle installation.

    • Color mismatch? Adjust LAB in ImageColorMatch+.