1. Workflow Overview

This is a dance video stylization workflow based on WanVideo-UniAnimate, featuring:

• Real dance video to anime conversion

• Pose control (DWPose) + Face swapping (ReActor)

• 3x frame interpolation (RIFE VFI)

• 720P/1080P output

Key Models:

• Wan2.1-I2V-14B: 14B-parameter base model (FP8 quantized)

• UniAnimate Pose: Precise dance motion capture

• RIFE 4.7: 3x frame interpolation

2. Critical Nodes

Dependencies:

1. Model Files:

   • Wan2_1-I2V-14B_fp8_e4m3fn.safetensors → Save to models/wanvideo/

   • umt5-xxl-enc-bf16.safetensors → Save to models/clip/

3. Workflow Structure

Phase 1: Input Prep

• VHS_LoadVideo: Loads dance video (downsampled to 8FPS)

Phase 2: Generation Core

1. DWPose extracts skeleton data

2. UniAnimate generates frames (720x1280)

3. Style control via LoRA

Phase 3: Post-Processing

• RIFE VFI: 3x frame interpolation (output 24FPS)

• Optional ReActor face swap

4. Inputs & Outputs

Required Inputs:

1. Source dance video (ae3135303eee...mp4)

2. Target face image (07b81177eb57...png)

Outputs:

• Final video as MP4 (H.264, CRF=19)

• Example filename: WanVideoWrapper_I2V_00011.mp4

5. Important Notes

⚠️ Hardware:

• Minimum 16GB VRAM (14B model requirement)

• Launch with --medvram flag

🔧 Troubleshooting:

1. VRAM overflow solutions:

   • Reduce frame_load_cap (e.g., set to 40)

   • Decrease resolution (set width to 640 in ImageResizeKJv2)

2. Motion glitches:

   • Increase motion_scale in WanVideoExperimentalArgs (default=1.25)