1. Workflow Overview

This workflow leverages the Wan2.1-I2V-14B model to generate dynamic videos with "Avatar Summoning" effects (e.g., semi-transparent phantom synchronized with character movements). It combines text prompts + input images and custom LoRAs (e.g., spell effects).

2. Core Models

• Wan2.1-I2V-14B-480P_fp8_e4m3fn.safetensors

  • Main model for video generation (image-to-video). Requires BF16 precision.

• umt5-xxl-enc-bf16.safetensors

  • T5 text encoder for processing complex prompts (supports Chinese).

• Wan2.1_VAE_bf16.safetensors

  • Decodes latent frames to images.

3. Key Nodes

• WanVideoModelLoader

  • Loads the main model. Manual download required (place in ComfyUI/models/wan_video).

• WanVideoTextEncode

  • Processes text prompts (positive/negative) using T5.

• WanVideoSampler

  • Uses DPM++ SDE sampler (25 steps default).

• WanVideoLoraSelect

  • Applies custom LoRAs (e.g., Avatar Summoning_beta).

• VHS_VideoCombine

  • Renders frames into MP4 (16 FPS).

4. Workflow Structure

1. Input Group

   • Text prompts (e.g., "A woman swings a sword, summoning a purple phantom").

   • Reference image (e.g., "修仙女子.png").

2. Generation Group

   • Model initialization via WanVideoModelLoader and WanVideoVAELoader.

   • Frame generation via WanVideoSampler.

3. Output Group

   • Video synthesis with VHS_VideoCombine (480x832 resolution).

5. Inputs & Outputs

• Inputs: Text prompts, image, seed (e.g., 1057359483639287).

• Outputs: MP4 video (H.264, with metadata).

6. Notes

• Dependencies: Manually download Wan2.1 models and LoRAs.

• VRAM: 16GB+ GPU recommended. Use BF16 to reduce usage.

• Compatibility: Requires ComfyUI-WanVideoWrapper (install via ComfyUI Manager).

• Troubleshooting:

  • FileNotFoundError if models are missing.

  • Reduce resolution in WanVideoBlockSwap for CUDA OOM errors.