1. Workflow Overview

This workflow transforms real-life portraits into 3D cartoon-style images with:

• Facial Feature Preservation: Via PulidFlux technology.

• Style Transfer: Uses LoRA models (e.g., miluo_cute_3D_cartoon).

• Upscaling: Optional SUPIR super-resolution (4K).

Core Models:

2. Key Nodes & Installation

Critical Nodes:

1. PulidFluxModelLoader

   • Purpose: Loads facial feature model (pulid_flux).

   • Install: Install PulidFlux plugin via ComfyUI Manager; place models in models/pulid_flux.

2. ApplyPulidFlux

   • Purpose: Merges input face features into generation.

   • Inputs: Source image + FACEANALYSIS data.

3. SUPIR_Upscale

   • Purpose: 4K upscaling (disabled by default).

   • Install: Clone ComfyUI-SUPIR from GitHub.

3. Workflow Structure

Group 1: Base Model Zone

• Inputs:

  • Base model: miluo-flux-pro.

  • LoRA: miluo_cute_3D_cartoon (strength 0.6).

• Output: Cartoon-style model.

Group 2: IPA (Facial Processing)

• Input: Upload portrait via LoadImage (clear face required).

• Process:

  • PulidFluxInsightFaceLoader: Detects facial landmarks.

  • ImageResize+: Resizes to 1024x1024.

Group 3: Sampling Zone

• Parameters:

  • Sampler: Euler (20 steps).

  • Prompt: pixar style, cute Chinese girl (customizable).

• Output: Cartoon latent result.

Group 4: Output & Upscaling

• Default Output: VAE-decoded image (768x1248).

• Optional Upscale: Enable SUPIR_Upscale for 4K.

4. Inputs & Outputs

Input Parameters:（json）

{
  "image": "Portrait (clear face required)",  
  "prompt": "pixar style, Asian little girl, flower background",  
  "negative_prompt": "low quality, blurry"  
}

Output:

• Cartoon image: /ComfyUI/output/xxxx.png.

• Upscaled image (optional): /ComfyUI/output/xxxx_upscaled.png.

5. Notes

1. Model Dependencies:

   • Download miluo-flux-pro (~7GB) and pulid_flux plugin.

   • Missing files trigger Model not found.

2. Optimization:

   • VRAM ≥12GB for base workflow; ≥16GB for upscaling.

   • Disable unused Groups (e.g., upscale) to save resources.

3. Troubleshooting:

   • Distorted faces: Ensure input image has a clear, front-facing face.

   • Style mismatch: Adjust LoRA strength (0.3~0.7).