Wan 2.2 14B Low-VRAM Workflow for ComfyUI

Generate high-quality text-to-video and image-to-video content using Wan 2.2 14B on budget hardware!

🚀 Performance Highlight: Generate a 1-second video in under 5 minutes on an RTX 3050 6GB!

✨ Features

Low VRAM Optimization - Works on 6GB VRAM cards (RTX 3050, 3060, AMD equivalents)
User-Friendly Interface - Simple FPS & Seconds input (automatic frame calculation)
Dual Mode Support - Text-to-Video (t2v) and Image-to-Video (i2v) with easy switching
Last Frame Extraction - Optional last frame saving for video chaining/extension
GGUF Quantization - Leverages quantized models for memory efficiency
Smart Memory Management - Aggressive CPU offloading for low-VRAM systems
Fast Generation - 4-step sampling for rapid results

🖼️ Screenshot

📋 System Requirements

Minimum

VRAM: 6GB (RTX 3050, 3060 6GB, AMD equivalents)
System RAM: 16GB
Storage: ~15GB for models

🔧 Required Custom Nodes

Install these ComfyUI custom nodes (via ComfyUI Manager or manual installation):

rgthree-comfy - Context system, workflow organization, group muting
cg-use-everywhere - Global parameter distribution
ComfyUI-Easy-Use - LoRA loading and model management
ComfyUI-VideoHelperSuite - Video processing and frame selection
ComfyUI-KJNodes - Utility nodes
ComfyUI-GGUF - GGUF model loading (CRITICAL for this workflow)
comfyui_memory_cleanup - Memory management for low-VRAM
ComfyUI-Custom-Scripts - Math expression nodes
efficiency-nodes-comfyui - Workflow efficiency nodes

Quick Installation via ComfyUI Manager

Open ComfyUI Manager
Click "Install Custom Nodes"
Search for each node pack above
Click "Install" on each
Restart ComfyUI

Manual Installation

cd ComfyUI/custom_nodes

git clone https://github.com/rgthree/rgthree-comfy
git clone https://github.com/chrisgoringe/cg-use-everywhere
git clone https://github.com/yolain/ComfyUI-Easy-Use
git clone https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite
git clone https://github.com/kijai/ComfyUI-KJNodes
git clone https://github.com/city96/ComfyUI-GGUF
git clone https://github.com/LAOGOU-666/Comfyui-Memory_Cleanup
git clone https://github.com/pythongosssss/ComfyUI-Custom-Scripts
git clone https://github.com/jags111/efficiency-nodes-comfyui

# Restart ComfyUI

📥 Required Model Files

This workflow uses the wan2.2-t2v-rapid-aio-v10-nsfw GGUF model from Phr00t/befox rapid merge.

Important Naming Note

The "D" prefix you may see in some screenshots (e.g., Dwan2.2-t2v-rapid-aio-v10-nsfw-Q3_K.gguf) is just my local folder organization. Download files with their original names as shown below - ComfyUI will load them correctly from the standard model folders.

1. Main Model (GGUF) - Choose based on your VRAM

Repository: https://huggingface.co/befox/WAN2.2-14B-Rapid-AllInOne-GGUF/tree/main/v10

Recommended for 6-8GB VRAM: Q3_K or Q4_K

Q3_K (~8 GB total memory, very low VRAM friendly)

Download: wan2.2-t2v-rapid-aio-v10-nsfw-Q3_K.gguf
Placement: ComfyUI/models/unet or models/checkpoints

Q4_K (~9-10 GB total memory, good balance) ⭐ Recommended

Download: wan2.2-t2v-rapid-aio-v10-nsfw-Q4_K.gguf
Placement: ComfyUI/models/unet or models/checkpoints

Q5_K (~11-12 GB total memory, higher quality)

Download: wan2.2-t2v-rapid-aio-v10-nsfw-Q5_K.gguf
Placement: ComfyUI/models/unet or models/checkpoints

2. Text Encoder / CLIP (umt5-xxl GGUF)

Repository: https://huggingface.co/city96/umt5-xxl-encoder-gguf/tree/main

Recommended: Q3_K_M (lowest VRAM usage, good performance)

Q3_K_M (~3 GB) ⭐ Recommended

Download: umt5-xxl-encoder-Q3_K_M.gguf
Placement: ComfyUI/models/clip or models/text_encoders

Q4_K_M (~3.7 GB, slightly higher quality)

Download: umt5-xxl-encoder-Q4_K_M.gguf
Placement: ComfyUI/models/clip or models/text_encoders

3. VAE

Repository: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/vae

wan_2.1_vae.safetensors (~254 MB)

Download: wan_2.1_vae.safetensors
Placement: ComfyUI/models/vae

Quick Tips

Verify file hashes after download (Hugging Face shows them on each file page)
If links change, search Hugging Face for "WAN2.2-14B-Rapid-AllInOne-GGUF" (befox), "umt5-xxl-encoder-gguf" (city96), or "Wan_2.2_ComfyUI_Repackaged" (Comfy-Org)
For best low-VRAM performance: Use Q3_K or Q4_K model + Q3_K_M CLIP + standard Wan VAE

Recommended Configurations by VRAM

VRAM	Main Model	Text Encoder	Total Memory	Performance
6GB	Q3_K	Q3_K_M	~11GB system	Fast, good quality
8GB	Q4_K	Q3_K_M	~13GB system	Balanced ⭐
10GB+	Q5_K	Q4_K_M	~15GB system	High quality

🎬 How to Use

Basic Workflow

Load the workflow in ComfyUI
Set your parameters:
- Width: 512 (default)
- Height: 512 (default)
- Seconds: 10 (duration of video)
- FPS: 15 (frames per second)
- Frames are calculated automatically!
Choose mode:
- Text-to-Video (t2v): Generate from text prompt
- Image-to-Video (i2v): Animate from input image
Enter your prompt (positive and negative)
Queue prompt and generate!

Text-to-Video (t2v) Mode

Make sure "Load Image" node is disabled/muted
Enter your text prompt describing the video
Queue prompt

Image-to-Video (i2v) Mode

Enable the "Load Image" node
Load your input image
Enter your prompt (can enhance/modify the image animation)
Queue prompt

Last Frame Extraction (Optional)

The workflow includes automatic last frame saving for video chaining:

Enabled by default - Last frame is automatically saved
To disable: Mute the "Save the Last Frame" group using rgthree Group Muter
Use case: Chain multiple generations together for longer videos

⚙️ Workflow Features Explained

Automatic Frame Calculation

Input: FPS × Seconds
Calculation: (FPS × Seconds) + 1
The +1 frame is required for Wan's t2v initialization
Works for both t2v and i2v modes automatically

Memory Optimization

GGUF quantization reduces model size
Aggressive CPU offloading (~9GB offloaded to system RAM)
Dynamic memory management between pipeline stages
Memory cleanup nodes for low-VRAM systems

Mode Switching

Clean interface with POS/NEG/Switch nodes
One control point switches both two paths (Text2Image and Image2Video) simultaneously
No manual rewiring needed between modes

🐛 Troubleshooting

"Out of Memory" Errors

Drop to Q3_K model instead of Q4_K
Close other applications to free system RAM
Enable Windows page file (set to system managed)
Reduce video length (fewer seconds)

"Missing Nodes" Error

Install all required custom nodes (see list above)
Use ComfyUI Manager's "Install Missing Nodes" feature
Restart ComfyUI after installing nodes

GGUF Models Not Loading

Ensure ComfyUI-GGUF is installed correctly
Verify model files are in correct folders with exact filenames
Check that files downloaded completely (verify file sizes)

Poor Quality Output

Try Q4_K or Q5_K model if you have more VRAM
Increase steps (though 4 steps works well with this model)
Adjust prompt strength and detail
Use better quality input images for i2v mode

Slow Generation

This is normal for low-VRAM setups (CPU offloading takes time)
Expected: ~5 minutes for 1 second video on RTX 3050 6GB
Close background applications
Ensure models are on SSD for faster loading

🎯 Performance Benchmarks

RTX 3050 6GB (Tested Configuration)

Model: Q4_K + Q3_K_M CLIP
Video length: 1 second (16 frames @ 15 FPS)
Total time: ~4.6 minutes
- Model loading: ~1 min
- Sampling (4 steps): ~1.8 min
- VAE decode: ~30s
Quality: Good, suitable for most use cases

Expected Performance on Other Cards

RTX 3060 12GB: ~3 minutes (less offloading overhead)
RTX 4060 8GB: ~2.5 minutes (faster architecture)
RTX 4070 12GB: ~2 minutes (minimal offloading)

Times are approximate and depend on system configuration

Optional LoRAs

The workflow supports standard LoRA loading through the EasyUse nodes. You can add any compatible Wan 2.2 LoRAs to enhance or modify output.

📚 Additional Resources

🤝 Contributing

This workflow is released into the public domain with no restrictions. Use it however you want - personal, commercial, modified, redistributed.

Contributions, improvements, and forks are welcome! No attribution required, but appreciated! ❤️

🙏 Credits

befox - GGUF quantized Wan models
city96 - ComfyUI-GGUF loader and umt5-xxl encoder
Comfy-Org - Wan 2.2 repackaged models
ComfyUI community - All the amazing custom node developers

📜 License

This workflow is released into the public domain with no restrictions.

You are free to:

Use commercially
Modify and redistribute
Use in closed-source projects
Not provide attribution (though it's appreciated!)

No warranty provided. Use at your own risk.

💬 Support & Feedback

If you find this workflow helpful:

⭐ Star this repository
🐛 Report issues or bugs
💡 Suggest improvements
📢 Share with the community!

Name		Name	Last commit message	Last commit date
Latest commit History 5 Commits
LICENSE		LICENSE
README.md		README.md
WAN2.2.json		WAN2.2.json

Folders and files

Latest commit

History

Repository files navigation