🎞️ ComfyUI Image-to-Video Workflow - WAN 2.1 Wrapper (Kiko WAN v3)
This is a high-performance, multi-pass Image-to-Video workflow for ComfyUI, powered by the WAN 2.1 Wrapper, with advanced optimizations like `torch.compile` and Sage Attention for faster and smarter frame generation. I tried to expose all the settings that Kijai exposes that I can understand 😉, This is not the fastest workflow you will find on here, but it is one I use to make 20 secons videos.
Crafted with ❤️ on Arch Linux BTW, using an RTX 4090 and 128 GB of RAM—this setup is tuned for heavy-duty inference and silky-smooth video generation.
---
🚀 Features
- 🧠 WAN 2.1 Wrapper for cinematic image-to-video transformations
- 🔂 Two-pass generation: initial + refinement/extension
- 🐌 Optional Slow Motion + Frame Interpolation (RIFE, FILM, etc.)
- 🧽 Sharpening and Upscaling (e.g., RealESRGAN, SwinIR)
- 🛠️ Includes torch.compile for faster inference
- 🌀 Integrates Sage Attention for improved attention efficiency
- 📏 Customizable prompts, seed, duration, and aspect ratio logic
- 🌀 Final loop polish with "Extend Last Frame"
---
⚙️ System Specs
- OS: Arch Linux (rolling release)
- GPU: NVIDIA RTX 4090 (24GB VRAM)
- RAM: 128 GB DDR5
- Python: 3.12.9 via `pyenv`
- ComfyUI: Latest build from GitHub
- torch: 2.x with `torch.compile` enabled
- Sage Attention: Enabled via patched attention mechanism
---
🛠️ Workflow Overview
🔹 Input & Resize
- Drop an image and optionally resize to fit WAN 2.1's expected input.
🔹 WAN 2.1 Wrapper Core
- Uses `torch.compile` for speed boost
- Enhanced with Sage Attention (set via the custom node or environment)
🔹 Pass 1: Generate + Optional Slow Motion
- Frame-by-frame synthesis
- Add slow motion via interpolation node (RIFE or FILM)
🔹 Pass 2: Extend + Merge
- Extends the motion, ensures smoother transitions
- Combines motion with refined prompt guidance
🔹 Final Polish
- Sharpening and Upscaling
- Final interpolation if needed
- Loop-ready output by extending the last frame
---
🧪 Performance Tips
- Tune torch compile for you system, they are all different, my setting might not work for you.
- For Sage Attention:
- Use the node
- Running on lower-end GPUs? Disable upscaling and reduce frame count.
---
🧰 Requirements
- ComfyUI
- WAN 2.1 Wrapper Node
- Optional:
- `RIFE`, `FILM`, or `DAIN` for interpolation
- `RealESRGAN` / `SwinIR` for upscaling
- `Sage Attention` patch or node
---
▶️ How to Use
1. Load the `kiko-wan-v3.json` file into ComfyUI.
2. Drop your image into the input node.
3. Customize prompts, duration, and frame count.
4. Click `Queue Prompt` to generate.
5. Your video will be rendered in the output folder.
---
📁 Files
- `kiko-wan-v3.json` — Exported workflow (coming soon)
- `kiko-wan-v3.png` — Workflow diagram
---
🧠 Inspirations & Credits
- [ComfyUI]()
- [WAN 2.1 Wrapper]()
- Real-ESRGAN, RIFE, FILM, Sage Attention contributors
- Arch Linux + NVIDIA ecosystem for elite workstation performance 😉
---
💡 Future Plans
- [ ] Add batch image-to-video mode
- [ ] Audio?
⚙️ Custom Nodes Used in `kiko-wan-wrapper-v3.json`
- Anything Everywhere:
- Display Any (rgthree):
- Fast Bypasser (rgthree):
- Fast Groups Bypasser (rgthree):
- GetImageRangeFromBatch:
- GetImageSize+:
- GetNode:
- SetNode:
- Image Filter:
- ImageBatchMulti:
- ImageFromBatch+:
- ImageListToImageBatch:
- ImageResizeKJ:
- Int:
- LoadWanVideoClipTextEncoder:
- LoadWanVideoT5TextEncoder:
- MarkdownNote: NOT FOUND
- PlaySound|pysssss:
- ProjectFilePathNode:
- RIFE VFI:
- ReActorRestoreFace:
- Seed Generator:
- SimpleMath+:
- Text Input [Dream]:
- VHS_VideoCombine:
- WanVideoBlockSwap:
- WanVideoDecode:
- WanVideoEnhanceAVideo:
- WanVideoFlowEdit:
- WanVideoImageClipEncode:
- WanVideoLoopArgs:
- WanVideoLoraBlockEdit:
- WanVideoLoraSelect:
- WanVideoModelLoader:
- WanVideoSLG:
- WanVideoSampler:
- WanVideoTeaCache:
- WanVideoTextEncode:
- WanVideoTorchCompileSettings:
- WanVideoVAELoader:
- WanVideoVRAMManagement:
