1. Workflow Overview

This workflow specializes in e-commerce glasses model generation with:

• Multi-angle portrait synthesis from reference glasses image

• Dual-ControlNet architecture for precise glasses positioning

• Expression fine-tuning and HD refinement
  Core technologies:

  • AWPortrait FL_bf16: Professional portrait model

  • RealPortrait-Glasses_v1.0 LoRA: Glasses adaptation

  • DepthAnything+DWPreprocessor: Dual-ControlNet

2. Key Components

3. Critical Nodes

• Dual-ControlNet Setup:（Python）

  # ControlNet1 (Depth)
  "AIO_Preprocessor": "DepthAnythingPreprocessor"
  "ControlNetApplyAdvanced": [0.4, 0, 0.8]  # strength/start/end
  
  # ControlNet2 (Edge)
  "AIO_Preprocessor": "DWPreprocessor"
  "ControlNetApplyAdvanced": [0.8, 0, 0.8]

• LoRA Configuration:（Python）

  "LoraLoader": ["RealPortrait-Glasses_v1.0", 0.6, 1]  # model/strength/CLIP_strength

4. Workflow Architecture

graph TB
  A[Group1: Base Model Generation] --> B[Group2: Head Pose Adjustment]
  A --> C[Output Preview]

• Group1 (Green):

  • Input: Glasses reference (04.png)

  • Process:

    1. JanusProDescribeImage auto-prompting

    2. Dual-ControlNet processing

    3. AWPortrait+LoRA rendering

  • Output: 1536x1536 base image

• Group2 (Blue):

  • ExpressionEditor for head rotation:（Python）

    widgets_values = [0,-10.5,0,0,0,0,0,0,0,0,0,0,1,1,"OnlyExpression",1.7]
    # -10.5° Y-axis tilt

5. I/O Specifications

• Required Inputs:

  • LoadImage: Glasses reference (PNG with transparency)

  • Text: Base prompt (e.g. "professional 8K photo")

• Optional:

  • KSampler: seed=776058480007604, steps=30, cfg=1

  • ExpressionEditor: 23 facial parameters

• Outputs:

  • Multi-angle portrait sequence (1536x1536 PNG)

  • Final composited preview

6. Critical Notes

1. Hardware:

   • ≥12GB VRAM (1536x1536)

   • Launch with --medvram

2. Model Setup:

   # ControlNet path:
   ComfyUI/models/controlnet/FLUX.1-dev-ControlNet-Union-Pro-InstantX.safetensors

3. Troubleshooting:

   • Glasses misalignment: Adjust DWPreprocessor threshold

   • Face distortion: Reduce ControlNet strength (0.8→0.6)

   • VRAM overflow: Downscale to 1024x1024