1. Workflow Overview

This workflow uses Alibaba’s WAN2.1 model for Image-to-Video (I2V) generation, transforming a 2D anime-style character image (e.g., a girl cycling) into a dynamic video with camera motion.

Core Models:

• Wan2.1-I2V-14B: Video generation model (480P output).

• UMT5-XXL: Multilingual text encoder for Chinese prompts.

• OpenCLIP-ViT-H: Image encoder for input analysis.

2. Key Nodes

Dependencies:

• Manual model downloads:

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

  • umt5-xxl-enc-bf16.safetensors

  • Save to: ComfyUI/models/wan_video/

3. Workflow Groups

1. Load Image: Input character image (e.g., 00028-2257557710.png).

2. Model Load: Load WAN2.1, VAE, and text encoders.

3. Prompt Processing: Positive/Negative prompts (e.g., "speed effect").

4. Video Generation: WanVideoSampler generates latent frames.

5. Post-Processing: Upscaling (CR Upscale Image).

6. Video Export: MP4 output (VHS_VideoCombine).

7. Frame Interpolation: Boosts FPS to 32 via RIFE VFI.

4. Inputs & Outputs

• Inputs:

  • Image: 480x768 PNG (anime character).

  • Prompts: Chinese/English descriptions.

  • Params: Frame rate (16→32FPS), CFG=8.0, Steps=30.

• Output:

  • MP4 video (default: xiao.mp4).

5. Notes

• VRAM: ≥16GB GPU recommended (use bf16 for lower usage).

• Debugging:

  • Missing model → Check wan_video folder path.

  • RIFE error → Ensure rife47.pth is in ComfyUI/models/rife/.

• Optimization:

  • Reduce resolution (e.g., 360P) for lower VRAM.

  • Disable experimental features to speed up.