ComfyUI Flux

ComfyUI Flux is a Docker-based setup for running ComfyUI with FLUX.1 models and additional features.

Features

• Dockerized ComfyUI environment
• Automatic installation of ComfyUI and ComfyUI-Manager
• Low VRAM Mode: Download and use FP8 models for reduced VRAM usage
• Pre-configured with FLUX models and VAEs
• Easy model management and updates
• GPU support with CUDA 12.1 or 12.4

Prerequisites

• Docker and Docker Compose
• NVIDIA GPU with CUDA support (for GPU acceleration)
• (Optional) Huggingface account and token (for downloading official FLUX.1[dev] and FLUX.1[schnell] models)

Quick Start

1. (Optional) Create a .env file in the project root and add your Huggingface token:

   # For downloading FLUX.1[dev] and FLUX.1[schnell] model files.
   # Get your token from https://huggingface.co/settings/tokens
   HF_TOKEN=your_huggingface_token

2. Download the docker-compose.yml file:

   wget https://raw.githubusercontent.com/frefrik/comfyui-flux/main/docker-compose.yml

   Alternatively, you can create a docker-compose.yml file and copy/paste the following contents:

   services:
     comfyui:
       container_name: comfyui
       image: frefrik/comfyui-flux:latest
       restart: unless-stopped
       ports:
         - "8188:8188"
       volumes:
         - "./data:/app"
       environment:
         - CLI_ARGS=
         - HF_TOKEN=${HF_TOKEN}
         - LOW_VRAM=${LOW_VRAM:-false}
       deploy:
         resources:
           reservations:
             devices:
               - driver: nvidia
                 device_ids: ['0']
                 capabilities: [gpu]

   Note: The default Docker image uses CUDA 12.1 (cu121). If you require CUDA 12.4, you can specify the cu124 tag in the image name. For example:

   image: frefrik/comfyui-flux:cu124

3. Run the container using Docker Compose:

   docker-compose up -d

   Note: The first time you run the container, it will download all the included models before starting up. This process may take some time depending on your internet connection.

4. Access ComfyUI in your browser at http://localhost:8188

Environment Variables

• HF_TOKEN - Your Hugging Face access token. Required to download the official flux1-dev, flux1-schnell models.
• LOW_VRAM - Set to true to download and configure ComfyUI for FP8 models, reducing VRAM usage. See Low VRAM Mode section.
• MODELS_DOWNLOAD - Comma-separated list specifying which FLUX base models to download (schnell, dev). If not set, both models will be downloaded.
• CLI_ARGS - Additional command-line arguments to pass directly to the ComfyUI.

Low VRAM Mode

By setting the LOW_VRAM environment variable to true, the container will download and use the FP8 models, which are optimized for lower VRAM usage. The FP8 versions have CLIP and VAE merged, so only the checkpoint files are needed.

Enable Low VRAM Mode:

LOW_VRAM=true

Model Files

Overview of the model files that will be downloaded when using this container.
Both official FLUX.1[dev] and FLUX.1[schnell] model files now require a HF_TOKEN for download.

When LOW_VRAM=false (default)

The following model files will be downloaded by default, unless specified otherwise in MODELS_DOWNLOAD:

Type	Model File Name	Size	Notes
UNet	flux1-schnell.safetensors	23 GiB	requires HF_TOKEN for download
UNet	flux1-dev.safetensors	23 GiB	requires HF_TOKEN for download
CLIP	clip_l.safetensors	235 MiB
CLIP	t5xxl_fp16.safetensors	9.2 GiB
CLIP	t5xxl_fp8_e4m3fn.safetensors	4.6 GiB
LoRA	flux_realism_lora.safetensors	22 MiB
VAE	ae.safetensors	320 MiB	requires HF_TOKEN for download

When LOW_VRAM=true

Type	Model File Name	Size	Notes
Checkpoint	flux1-dev-fp8.safetensors	17 GiB
Checkpoint	flux1-schnell-fp8.safetensors	17 GiB

Workflows

Download the images below and drag them into ComfyUI to load the corresponding workflows.

Official versions

FLUX.1[schnell]	FLUX.1[dev]
Download	Download

FP8 versions (LOW_VRAM)

FLUX.1[schnell] FP8	FLUX.1[dev] FP8
Download	Download

Updating

The ComfyUI and ComfyUI-Manager are automatically updated when the container starts. To update the base image and other dependencies, pull the latest version of the Docker image using:

docker-compose pull

Additional Notes

• Switching Between Modes: If you change the LOW_VRAM setting after the initial run, the container will automatically download the required models for the new setting upon restart.
• Model Selection: Use the optional MODELS_DOWNLOAD environment variable to specify which FLUX models to download:
  • MODELS_DOWNLOAD="schnell": Download only FLUX.1[schnell] model
  • MODELS_DOWNLOAD="dev": Download only FLUX.1[dev] model
  • MODELS_DOWNLOAD="schnell,dev" or not set: Download both models (default)
  • When LOW_VRAM=false, necessary dependencies (VAE, text encoders, LoRA) are always included
• Model Downloading: The scripts are designed to skip downloading models that already exist, so you won't waste bandwidth re-downloading models you already have.
• Huggingface Token: The HF_TOKEN is necessary for downloading the flux1-dev.safetensors, flux1-schnell.safetensors, and ae.safetensors models when LOW_VRAM=false.