ComfyUI FluxMod 🚀

A modulation layer addon for Flux that reduces model size to 8.8B parameters without significant quality loss.

Overview

ComfyUI_FluxMod acts as a plugin for Flux, enabling you to run Flux Dev and Flux Schnell on more consumer-friendly hardware. This is achieved by utilizing a modulation layer that significantly reduces the parameter count while maintaining quality.

Note

: You still need the original Flux Dev or Flux Schnell model to use this addon.

Table of Contents

• Requirements
• Installation
• Usage
• Node Information
• Quantization Guide
• FAQ
• Support
• Contributing
• Issues
• Sample Workflow

Requirements

• ComfyUI installation
• Original Flux model (Dev or Schnell)
• universal_modulator.safetensors

Installation

Via Comfy Registry (Recommended)

comfy node registry-install comfyui_fluxmod

Manual Installation

1. Navigate to your ComfyUI's custom_nodes folder
2. Clone the repository:

git clone https://github.com/lodestone-rock/ComfyUI_FluxMod.git

3. Restart ComfyUI
4. Refresh your browser if ComfyUI is already running

Node Information

Node	Description	Options
FluxModCheckpointLoader	Primary checkpoint loading node	• ckpt_name: Original Flux model path • guidance_name: Modulation addon path • quant_mode: Quantization selection

Usage

1. Double click workspace → search "FluxModCheckpointLoader"
2. Select your Flux model in ckpt_name
3. Select modulation addon in guidance_name
4. Choose quantization mode
5. Configure remaining nodes as per standard Flux workflow

💡 Tip: Check the examples folder for sample workflows. Drag and drop the workflow image into ComfyUI to get started quickly.

Quantization Guide

Mode	Recommended GPU	VRAM Usage	Recommended
bf16	24GB+	~20GB	✅
float8_e4m3fn	12-16GB	~10GB	✅
float8_e5m2	12-16GB	~10GB	❌

FAQ

Will my outputs be different?

Yes, outputs will likely differ as we're reducing parameters. However, the difference is often minimal for most use cases.

How much does this degrade image quality?

Testing shows minimal quality degradation in most cases. The most notable exception is long text generation, which shows a moderate degradation.

Why use this over regular quantization?

You don't have to use this over regular quantisation! You can combine them, or if you don't want to use quantisation at all and you have enough VRAM, you can also just stick with bf16. If you combine quantisation, you can make the model even smaller and allow it to run on consumer hardware.

Support

Need help? Join our Discord community for support and discussions.

Contributing

Pull requests are welcome! Feel free to contribute to the project by:

• Fixing bugs
• Adding new features
• Improving documentation
• Suggesting enhancements

Issues

Found a bug? Have a suggestion? Please create a GitHub issue with:

• A clear description of the problem
• Steps to reproduce
• Expected vs actual behavior
• Your environment details (OS, GPU, etc.)

Sample Workflow