1. Workflow Overview

This workflow specializes in facial expression manipulation with:

• Multi-expression generation from a single input image

• Expression-to-video conversion

• Video-driven expression imitation
  Core technologies:

  • ExpressionEditor: Precise facial muscle control

  • AdvancedLivePortrait: Video-style expression transfer

  • VHS_VideoCombine: Frame sequence video rendering

2. Core Components

3. Key Nodes Deep Dive

• ExpressionEditor Parameters:（Python）

  widgets_values = [
    0,          # Brow vertical
    0,          # Brow tilt
    8.5,        # Left eye open (0-15)
    -23,        # Mouth corner (-25~25)
    "OnlyExpression"  # Mode selector
  ]

• Dependencies:

  • Required model: facial_landmark_68pts.dat in ComfyUI/models/facial_animation/

4. Workflow Architecture

graph LR
  A[Group1: Multi-Expression] --> B[Group2: Video Conversion]
  B --> C[Group3: Video Imitation]

• Group1 (Blue):

  • Input: Single portrait (1024x1024)

  • Process: 6 parallel ExpressionEditor nodes

  • Output: PNG sequence via SaveImage

• Group2 (Middle):

  • Converts frames to 15fps video (CRF=19)

• Group3 (Right):

  • Video-driven expression using driving_video.mp4

  • Output: H.264 MP4 with audio

5. I/O Specifications

• Required Inputs:

  • LoadImage: Front-facing portrait

  • VHS_LoadVideo: Reference MP4 video

• Optional:

  • 23 expression parameters per editor

  • select_every_nth frame sampling

• Outputs:

  • AdvancedLivePortrait_xxxx.mp4 (with metadata)

  • Intermediate PNG sequence

6. Critical Notes

1. Hardware:

   • ≥8GB VRAM (video processing)

   • Launch ComfyUI with --highvram

2. Troubleshooting:

   • Error Missing facial landmark model:

     wget https://github.com/kurnianggoro/GSOC2017/raw/master/data/lbfmodel.yaml -P models/facial_animation/

   • Unnatural expressions: Adjust OnlyExpression/OnlyEyes mode

3. Optimization:

   • Reduce frame_rate to 10-15fps for lower VRAM

   • Enable --fp16 for ExpressionEditor acceleration