1. Workflow Overview

This workflow utilizes Alibaba's Wan2.1 model to generate videos from static images (I2V). Key features:

• Extracts image features via CLIP vision encoder

• Processes multilingual prompts with T5 text encoder

• Generates video latent using 14B-parameter Wan2.1-I2V model

• Outputs animated WEBP/MP4 files

2. Core Models

3. Key Nodes

4. Workflow Structure

• Group 1: Input Processing

  • LoadImage: Loads input image (e.g., 576x1024)

  • WanVideoTextEncode: Processes prompts (e.g., "A smiling ancient beauty")

• Group 2: Model Loading

  • LoadWanVideoT5TextEncoder: Loads T5 encoder

  • WanVideoModelLoader: Loads 14B video model

• Group 3: Video Generation

  • WanVideoSampler: Generates latent (30 frames, CFG=6)

  • WanVideoDecode: Decodes to image sequence via VAE

5. Inputs & Outputs

• Required Inputs:

  • Image file (PNG/JPG)

  • Positive prompt (e.g., style description)

  • Negative prompt (e.g., "low quality, static")

• Outputs:

  • Animated WEBP (default) or MP4

  • Resolution: 272x272 (adjustable)

6. Notes

⚠️ Troubleshooting:

1. VRAM: 14B model requires ≥16GB GPU, enable bf16 precision

2. Plugin: Manual install required:

   git clone https://github.com/AI-ModelScope/comfyui-wanvideo-plugin

3. Models: Place all .safetensors in models/wanvideo/