1. Workflow Overview

This workflow leverages WAN2.1 I2V model to generate videos of traditional Chinese costume transformation. Key features:

• Image Input: Processes portrait images (e.g., 79-花瓣网.jpg) at 1280x720 resolution.

• Text Guidance: Uses prompts like "gufengbianshen" to control costume effects.

• Video Generation: Integrates LoRA (WAN2.1 I2V 国风戏服变身_V1) for dynamic transitions.

• Post-Processing: Optional frame interpolation (GIMMVFI) and upscaling (2xNomosUni).

Use Cases: Historical role animation, short video effects, cultural content creation.

2. Core Models

3. Key Nodes

4. Workflow Structure

1. Input Preprocessing

   • Nodes: LoadImage → ImageScale

   • Input: Source image (auto-resized to 1280x720).

   • Output: Normalized image.

2. Feature Encoding

   • Nodes: WanVideoTextEncode + WanVideoImageClipEncode

   • Input: Text prompts + resized image.

   • Output: Joint text-image embeddings.

3. Video Generation

   • Nodes: WanVideoSampler → WanVideoDecode

   • Params: Seed 797198568875963, CFG=5, 16FPS.

   • Output: Raw video (costume transformation).

4. Post-Processing (Optional)

   • Interpolation: GIMMVFI boosts to 30FPS.

   • Upscaling: 2xNomosUni enhances resolution.

5. Inputs & Outputs

• Required Inputs:

  • Image: Full-body portrait (plain background recommended).

  • Prompts: Must include keywords like "traditional costume".

• Output:

  • Video: MP4 (H.264, 16FPS default, 30FPS interpolated).

  • Save Path: ComfyUI/output/WanVideoWrapper_I2V_xxxxx.mp4.

6. Notes

⚠️ Common Issues:

• VRAM Overflow: Requires ≥10GB GPU. Enable WanVideoVRAMManagement for optimization.

• LoRA Failure: Verify filename includes _V1.safetensors.

• Interpolation Errors: Ensure gimmvfi_r_arb_lpips_fp32.safetensors is in models/gimmvfi.

🔧 Optimization Tips:

• Disable redundant options in ExperimentalArgs for faster generation.

• Use ImageUpscaleWithModel for HD output (requires extra VRAM).