ComfyUI-UNO-Wrapper

This extension integrates ByteDance’s UNO-FLUX model into ComfyUI, allowing you to use UNO’s powerful text-to-image generation with reference capabilities.

Features

• Generate images with text prompts and up to 4 reference images
• Full control over guidance scale, steps, and dimensions
• Batch generation support
• Leverages your existing ComfyUI text encoders

Installation

1. Install ComfyUI-UNO-Wrapper

cd ComfyUI/custom_nodes
git clone https://github.com/ShmuelRonen/ComfyUI-UNO-Wrapper

2. Download the LoRA Model (REQUIRED)

Most models will be installed automatically, but you must manually download the ByteDance LoRA model:

1. Create a folder named uno_lora in your ComfyUI/models/loras directory
2. Download dit_lora.safetensors and place it in this folder

Download Link:

• bytedance-research/UNO

3. Hugging Face Authentication Setup

The UNO-FLUX implementation automatically attempts to download model files from Hugging Face repositories. For private or gated models, you’ll need to set up authentication:

Option 1: Using config.json (Recommended)

1. Create a file named config.json in the ComfyUI/custom_nodes/ComfyUI-UNO-Wrapper directory
2. Add the following content, replacing with your actual Hugging Face token:

{
    "hf_token": "your_huggingface_token_here"
}

3. Restart ComfyUI for the changes to take effect

Getting a Hugging Face Token

1. Visit Hugging Face and create an account or log in
2. Go to your profile settings
3. Navigate to “Access Tokens”
4. Create a new token with “read” permissions
5. Copy the token and paste it into your config.json file

Usage

Basic Workflow

1. Add a “UNO Model Loader” node

   • Configure device settings if needed (most settings work with defaults)

2. Connect it to a “UNO Image Generator” node

   • Enter your text prompt
   • Set dimensions, guidance scale, and steps as desired
   • Connect up to 4 reference images
   • Set batch size and seed

3. Run the workflow to generate images

Working with Multiple Reference Images

UNO-FLUX’s standout feature is its ability to incorporate elements from multiple reference images:

1. Connect different reference images to image_ref1, image_ref2, etc.
2. Each reference image will influence different aspects of the generated output
3. Use a descriptive prompt to guide how the references are combined

Troubleshooting

• Generation Fails: Make sure you’ve downloaded the LoRA model to models/loras/uno_lora/dit_lora.safetensors
• CUDA Errors: Try enabling offload in the UNO Model Loader
• Poor Results: Try adjusting the guidance scale (higher for more prompt adherence, lower for more creativity)
• Download Errors: Check your Hugging Face token in the config.json file is correct and has proper permissions
• LoRA Not Working: Ensure your LoRA files are in the correct format and located in the models/loras directory

License

Based on UNO ByteDance project: https://github.com/bytedance/UNO

This project follows the same license as UNO-FLUX (Apache License 2.0)