🧠 Kiko9 ComfyUI WAN 2.1 Native Workflow

ComfyUI image-to-video (I2V) pipeline built around WAN 2.1 using native ComfyUI and Torch compilation (torch.compile) for performance gains. The design includes 2-pass generation, frame interpolation, upscaling, and slow motion — tailored for high-fidelity AI-enhanced video generation.

Link to workflow I use for start image:

📦 Workflow Overview

🛠️ Project Breakdown

🔧 Project Settings

• Project File Path Generator: Allows saving outputs with a defined base path. Set this to your local output folder.

  • ✅ User Action: Update root_path to your preferred save location.

🧮 Aspect Ratio Logic (Don't Touch)

• Calculates width and height from image size using a float-to-int conversion for maintaining aspect ratio.

  • ⚠️ Do not modify unless you understand aspect ratio propagation.

📸 Image Generation for Video (Optimized Resolution)

• When creating video frames using image generation tools like FLUX / SDXL, it's important to generate at the right resolution to maintain sharpness and consistency.

🎯 Target Video Resolution

• Target Size: 480x832

• Aspect Ratio: 480 ÷ 832 ≈ 0.577

✅ Ideal Generation Resolution

To preserve details and allow for high-quality downscaling, generate at 2x or higher resolution. A perfect match in aspect ratio ensures you avoid cropping or distortion.

Gen ResolutionAspect RatioNotes960x1664960 ÷ 1664 ≈ 0.577✅ Perfect aspect ratio match1024x15361024 ÷ 1536 ≈ 0.6667🔶 Slight crop or padding needed

🔄 Workflow

1. Generate High-Res Images Use 960x1664 or larger with the same aspect ratio. Using FLUX, SDXL, etc.

🧮 Why This Works

• High-res generation reduces artifacts and increases fidelity.

• Downscaling averages pixels, smoothing jagged edges and noise.

• Maintaining the same aspect ratio avoids warping or unnecessary padding.

📥 Loaders

• Load Checkpoint (WAN2.1): Load the WAN 2.1 native (ComfyUI) model checkpoint.

• VAE & CLIP Loader: Loads required VAE and CLIP encoders.

• Power LoRA Loader (optional): For Power LoRa.

• Tile Cache, Enhance, and CLIP Vision: Load auxiliary models.

  • ✅ User Action:

    • Set ckpt_name, vae_name, and clip_name according to local model files.

    • Ensure files are in your configured ComfyUI model folders.

🖼️ Image / Resize

• Load Image / Resize: Loads the input image or first frame from a video clip, resizes it to model-appropriate dimensions.

🌍 Global Settings

• CLIP Text Encode (Prompt & Negative): Prompts for conditioning the model.

  • ✅ User Action: Customize these prompts per your subject/style.

• Seed Generator / Upscale Factor: Controls random seed and image scale-up.

  • ✅ User Action: Set seed for reproducibility or leave -1 for random.

🔁 1st Pass (Initial Generation)

• KSampler: Runs the initial inference.

• VAE Decode & Video Combine: Decodes latent space to image, combines with source.

• Slow Motion / PlaySound: Optional audio sync and slow-mo settings.

• Select last frame for 2nd pass start frame. (Pop Up window)

🔁 2nd Pass (Refine & Extend)

• Similar to 1st Pass but optimized for longer inference or higher quality.

• Take last frame from 1st pass as 2nd pass starting image.

• Get Mask Range From Clip: Extracts mask regions for attention.

• Image Batch Multi: Processes multiple frames simultaneously.

📈 Upscaling & Frame Interpolation

• Image Sharpen / Restore Faces: Post-processing enhancements.

• Upscale Image (Real-ESRGAN or similar).

• Frame Interpolation (RIFE / FILM): Smooth transitions for higher FPS.

• Slow Motion: Optional, adds frames and blends for cinematic slow-mo.

🧪 Experimental (Optional, Long Runtime)

• Advanced enhancement or second-stage denoising/refinement.

• Useful for batch rendering with very high quality needs.

  • ⏱️ Warning: These steps significantly increase processing time.

⚡ Torch Compile Setup (VERY IMPORTANT)

To unlock native acceleration via torch.compile, ensure you meet these requirements:

✅ Requirements

• PyTorch 2.1+ with CUDA

• NVIDIA GPU with Ampere or later architecture (RTX 30XX, 40XX)

• Use latest nightly ComfyUI or manually apply torch.compile() patching.

💾 Saving Outputs

• Controlled via Project Path Generator and Video Combine nodes.

• Output format (e.g. .mp4, .png, .webm) should be explicitly set in Video Combine.

📋 Notes

• ⚠️ First run of torch.compile will be slow due to graph tracing.

• 🧠 Prompt tuning is crucial for WAN 2.1 — try detailed descriptions.

• ⚠️ Not optimized for older machines.

🙋 FAQ

Q: My output is laggy or missing frames.

• Check interpolation settings and slow motion settings — disable one if not needed.

Q: Workflow crashes during torch compile.

• Ensure you're using PyTorch 2.1+, and your GPU is Ampere or newer.

Q: Can I use this with other models like SDXL?

• You can, but WAN 2.1 is optimized for this specific setup. Results may vary.

📎 Credits

• Workflow design by Kiko9

• WAN 2.1

• ComfyUI team for the powerful modular engine

📂 Folder Structure Example

ComfyUI/
├── models/
│ ├── checkpoints/
│ ├── vae/
│ ├── clip/
├── output/
│ └── generated/
├── custom_nodes/ │

📊 End-to-End WAN 2.1 Generation Summary

StepDescriptionTime / Count. Resolution

Prompt StartInitial prompt execution begins 92.95 sec

Model LoadLoaded WAN21 model weights ~15,952 ms

First Comfy-VFI PassGenerated frames with TeaCache initialized ~6 min 13sec 480x832

Frames Generated (1st pass)Comfy-VFI output 231 frames 480x832

Second Comfy-VFI PassRepeats generation with same steps ~6 min 28 sec 480x832

Frames Generated (2nd pass)Comfy-VFI output(Implied 480x832

WanVAE Load (1st)Loaded latent space model ~1220 ms —

WanVAE Load (2nd)Loaded again for reuse ~1304 ms —

Face Restoration (GFPGAN)GFPGANv1.4 restored images 152 frames 512x512

Comfy-VFI Run (3rd)Generated additional frames ~unknown 960x1664 Frames Generated

(3rd pass)Comfy-VFI output 456 frames 960x1664

Comfy-VFI Run (4th)Final batch of generation~unknown 960x1664 Frames Generated

(4th pass)Comfy-VFI output304 frames960x1664Prompt EndFinal step of pipeline 1050.60 sec—

ℹ️ Notes:

• "TeaCache skipped" 12 conditional + 12 unconditional steps per 30 = ~20% optimization.

• Face restoration step was applied to a subset (152 frames).

• The 960x1664 resolution used in the last two passes matches the 480x832 aspect ratio perfectly, ideal for downscaling or 2x video output.

🗨️ Feedback & Contributions

Feel free to submit issues if you encounter bugs or want to contribute improvements.

🔥 Happy rendering!