1. Workflow Overview

This workflow generates interpolated frames between start/end images and upscales the video. Key steps:

1. Input start/end frames, generate intermediate frames via VACE model.

2. Enhance video sequence using WanVideo sampler.

3. Apply 4x super-resolution with NMKD-Siax model.
   Use Case: Smooth animation transitions (e.g., character movements, scene morphing).

2. Core Models

3. Key Nodes

Dependencies:

• Plugin ComfyUI-WanVideoWrapper (GitHub)

• Models vace_preview_fp32.safetensors & Wan2_1_VAE_bf16.safetensors (manual download)

4. Workflow Groups

• Group 1: Frame Input

  • Nodes: LoadImage → ImageResizeKJ

  • Output: Resized images (832x480).

• Group 2: VACE Interpolation

  • Nodes: WanVideoVACEStartToEndFrame → GetImageSizeAndCount

  • Output: 81 interpolated frames (5s at 16fps).

• Group 3: Video Sampling

  • Nodes: WanVideoSampler → WanVideoDecode

  • Parameters: 20 steps, CFG=4.0, DPM++ scheduler.

• Group 4: Upscaling

  • Nodes: UpscaleModelLoader → VHS_VideoCombine

  • Output: 4K MP4 video (CRF=19).

5. Inputs & Outputs

• Inputs:

  • Start/end images (JPG/PNG).

  • Prompt (e.g., "A woman practicing yoga").

  • Frame count (default: 81), resolution (832x480).

• Output:

  • Upscaled MP4 video (filename prefix xiao).

6. Notes

⚠️ VRAM: Minimum 8GB (RTX 3060+ recommended).
⚠️ Errors:

• OOM error no attribute 'is_directml_enabled' → Reduce frames/resolution.

• Upscale OOM → Disable ImageUpscaleWithModelBatched or switch to 2x model.
  ⚠️ Tips:

• Start/end frames must match dimensions (e.g., 512x512).

• Enable WanVideoTeaCache for >100 frames to save VRAM.

Demo Steps

1. Replace LoadImage paths with your start/end frames.

2. Modify prompts in WanVideoTextEncode.

3. Click Queue Prompt. Output saves to ComfyUI/output.