1. Workflow Overview

This "Wan Image-to-Video" workflow transforms static images into dynamic videos (e.g., making portraits blink/smile). Powered by Wan 2.1 specialized model, key features include:

• Preserving original image details

• Generating natural motions

• Outputting 720P HD videos (24FPS)

2. Core Models

• Main Model: Quantized wan2.1-i2v-14b (4bit GGUF)

• CLIP Vision: clip_vision_h.safetensors

• Video VAE: Custom wan_2.1_vae

3. Key Nodes

Dependencies:

• Input image 1.png

• umt5-xxl text encoder (GGUF format)

4. Pipeline Stages

Stage 1: Initialization

• Load Wan model trio (UNet/CLIP/VAE)

• CLIP vision encode reference image

Stage 2: Motion Control

• Positive prompt drives actions (e.g., "slowly turns head")

• Negative prompt filters artifacts

Stage 3: Rendering

• uni_pc sampler (20 steps)

• MP4 output via FFmpeg

5. I/O Specification

• Inputs:

  • Source image (e.g., portrait)

  • Action description in English

• Output:

  • 720P video (wan_i2v_xxxx.mp4)

6. Critical Notes

⚠️ Requirements:

• 8GB+ VRAM (optimized with GGUF)

• NVIDIA 30/40 series recommended

🔧 Pro Tips:

• Adjust tile_size in VAEDecodeTiled for performance

• Keep "static" in negative prompts for smoother motion