1. Workflow Overview

This workflow leverages Wan2.1 model for Image-to-Video (I2V) generation, with:

1. Input: Single image + text prompt → Low-res video generation

2. RIFE interpolation + CR upscaling for higher FPS (32fps) and resolution

3. KJ acceleration (TeaCache/BlockSwap) to optimize VRAM usage

4. Outputs: Original (16fps) and enhanced videos (32fps)

2. Core Models

3. Key Nodes

4. Workflow Groups

• Group 1: Wan2.1 Model Loading

  • Input: Model files, VAE, LoRA

  • Output: Initialized video generation model

• Group 2: Text & Image Encoding

  • Input: Prompt (e.g., "1girl, golden hair"), negative prompt, uploaded image

  • Output: CLIP image embeds + T5 text embeds

• Group 3: Acceleration Nodes

  • Critical params: BlockSwap=20 (VRAM safety), TeaCache=0.04 (speed boost)

• Group 4: Initial Video Synthesis

  • Output: 480P raw video (16fps)

• Group 5: Upscale + Interpolation

  • Pipeline: CR upscale → RIFE interpolation → 1080P output (32fps)

5. Inputs & Outputs

• Input Parameters:

  • Required: Image path, Prompt, Seed (41387343190862)

  • Optional: LoRA model, RIFE multiplier (10)

• Output:

  • teacache_00002.mp4 (raw video)

  • xiao_00001.mp4 (HD interpolated video)

6. Notes

1. VRAM: Recommended ≥12GB GPU (e.g., RTX 3060+). Enable BlockSwap to reduce usage.

2. Compatibility: Wan models must be bf16/fp8 format.

3. Troubleshooting: If umt5-xxl-enc-bf16.safetensors is missing, download from HuggingFace to ComfyUI/models/wan_video.

4. Optimization: Adjust TeaCache (0.01~0.05) for speed/quality trade-off.