ComfyUI-FlashFace

ComfyUI implementation of FlashFace: Human Image Personalization with High-fidelity Identity Preservation

Officially implemented here: https://github.com/ali-vilab/FlashFace

Installation

1. Navigate to the custom_nodes directory in your ComfyUI installation.
2. Install the custom node by running git clone https://github.com/cold-hand/ComfyUI-FlashFace.git in the custom_nodes directory.
3. If you are using a virtual environment, activate it.
4. cd into the ComfyUI-FlashFace directory and run setup.sh or setup.bat depending on your OS.
   • This will install all required dependencies including pydash
   • Alternatively, you can run pip install pydash==7.0.7 manually
5. Restart ComfyUI and refresh your browser and you should see the FlashFace node in the node list.
6. Load the provided example-workflow.json in file in the examples/comfyui folder of this repo to see how the nodes are used.

Dependencies

This extension requires the following key dependencies:

• Python 3.8+
• PyTorch (included in ComfyUI)
• pydash 7.0.7 or newer
• See requirements.txt for the full list

If you encounter any import errors, run python install_dependencies.py from the extension directory.

Documentation

• Installation Guide - Detailed installation instructions
• Technical Documentation - Architecture and developer information
• Improvements - Summary of recent refactoring and improvements

Results

The FlashFace node takes in an image and outputs a personalized image with the desired attributes. In comparison with InstantId I feel that it yields more realistic results. The original repo had a limit of 4 reference images, that limitation has been removed in this implementation. I've found that generally the more reference images you provide the better the results. as seen below:

Example Reference Image:

Nicki Minaj

Output with 1 Reference Image	Output with 4 Reference Images	Output with 8 Reference Images	Output with 16 Reference Images

Change Log

• [5/22/2024]: Initial release
• [Current Date]: Major refactoring and improvements:
  • Added robust model loading with SafeTensors support
  • Implemented automatic model downloading from HuggingFace
  • Improved model path resolution for different ComfyUI installation configurations
  • Enhanced error handling and reporting
  • Added pydash dependency to requirements
  • Added model registry system for easy management of model files