1. Workflow Overview

This workflow is designed for first/last frame video generation and enhancement, featuring:

• Dynamic video synthesis from start/end images

• 4x super-resolution upscaling (using 4xRealWebPhoto_v4)

• Frame interpolation (10x via RIFE VFI)

• Auto frame labeling (e.g., "start_frame"/"end_frame")

• Dual-format output (MP4 + GIF)

2. Core Models

3. Key Nodes

4. Workflow Groups

• Group 1: Start Frame Processing

  • Input: 1288x1024 PNG

  • Nodes: LoadImage → ImageResizeKJ (480x640) → AddLabel ("start_frame")

• Group 2: End Frame Processing

  • Input: 1288x1024 PNG

  • Nodes: LoadImage → ImageResizeKJ → AddLabel ("end_frame")

• Group 3: Video Generation

  • Core: WanVideoSampler (30 steps, UniPC) → WanVideoDecode (to image sequence)

• Group 4: Post-Processing

  • Upscale: ImageUpscaleWithModel (4x)

  • Interpolation: RIFE VFI (10x)

  • Output: VHS_VideoCombine (MP4 16fps + GIF 32fps)

5. Inputs & Outputs

• Inputs:

  • Required: Start/End frame images (drag-and-drop)

  • Optional: Seed (default: 999), Prompt (e.g., "change clothes"), Negative prompt (pre-configured)

• Outputs:

  • output.mp4 (H.264, 16fps)

  • AnimateDiff.gif (32fps loop)

6. Notes

• VRAM: Recommended ≥16GB GPU (FP8 optimization reduces usage to ~12GB)

• Troubleshooting:

  • Check ImageResizeKJ parameters if image dimensions mismatch

  • Verify rife47.pth path if RIFE fails

• Optimization:

  • Enable torch.compile (configured with inductor mode)

  • Use easy cleanGpuUsed for manual VRAM release