1. Workflow Overview

This workflow generates dynamic videos with angry facial expressions from a single image, leveraging Image-to-Video (I2V) technology. Key features:

1. Transforms input images (e.g., portrait photos) into videos with emotion transitions.

2. Uses WanVideo 2.1 model + LoRA (Angry Expression｜i2V_v1) for enhanced details.

3. Applies 2x upscaling for higher resolution.
   Use Case: Film VFX, game character animation, short video content.

2. Core Models

3. Key Nodes

Dependencies:

• Plugin ComfyUI-WanVideoWrapper (GitHub)

• Model Wan2_1-I2V-14B-480P_fp8_e4m3fn.safetensors

• Llama model unsloth/Meta-Llama-3.1-8B-bnb-4bit

4. Workflow Groups

• Group 1: Model Loading

  • Nodes: WanVideoModelLoader + WanVideoVAELoader

  • Loads main model, VAE, and LoRA.

• Group 2: Image Encoding

  • Nodes: LoadImage → Joy_caption → WanVideoTextEncode

  • Input: Portrait image (e.g., "Vol.2野蛮学姐").

  • Dynamic prompt: "acial expression abruptly changes to angry face".

• Group 3: Video Generation

  • Nodes: WanVideoSampler → VHS_VideoCombine

  • Output: 480P → 2x upscaled MP4 (CRF=19).

5. Inputs & Outputs

• Inputs:

  • Character image (1024x1024 recommended).

  • Negative prompt: "static, blurry, bad quality".

  • Frame rate: 16fps (81 frames ≈5s).

• Output:

  • Upscaled MP4 (filename prefix wangyi AI Studio-wan).

6. Notes

⚠️ VRAM: Minimum 12GB (Llama model required).
⚠️ Troubleshooting:

• Unnatural expressions → Adjust CFG (current: 6.0, try 4.0-8.0).

• Upscale OOM → Disable PurgeVRAM or switch to 1x model.
  ⚠️ Tips:

• Use WanVideoTeaCache (enabled with offload_device).

• Input images: Close-up portraits work best.

Demo Steps

1. Replace LoadImage with your character design.

2. Tweak negative prompts in WanVideoTextEncode.

3. Run Queue Prompt. Output saves to ComfyUI/output.