A Nix flake for Fooocus - an image generating software focused on simplicity.

• Cross-platform: Linux (x86_64, aarch64) and macOS (Intel, Apple Silicon)
• Reproducible builds with Nix
• Automatic GPU detection (NVIDIA CUDA, Apple Silicon MPS, CPU fallback)
• Persistent data storage in ~/.config/fooocus
• Docker images (CPU and CUDA variants)
• Multiple preset support (default, anime, realistic)

Run Fooocus directly without installation:

nix run github:utensils/fooocus-nix

Or with a specific preset:

nix run github:utensils/fooocus-nix#anime
nix run github:utensils/fooocus-nix#realistic

Add to your flake.nix:

{
  inputs.fooocus-nix.url = "github:utensils/fooocus-nix";
}

Then use the overlay or package:

# Using the overlay
nixpkgs.overlays = [ fooocus-nix.overlays.default ];

# Or directly reference the package
environment.systemPackages = [ fooocus-nix.packages.${system}.default ];

git clone https://github.com/utensils/fooocus-nix
cd fooocus-nix
nix develop  # Enter development shell
nix run      # Run Fooocus

# Run with default settings
nix run .

# Open browser automatically when ready
nix run . -- --open

# Use a specific port
nix run . -- --port=7866

# Listen on all interfaces (for network access)
nix run . -- --listen 0.0.0.0

# Use a preset
nix run . -- --preset=anime

# Enable debug logging
nix run . -- --debug

Build and run the Docker image:

# CPU version
nix run .#buildDocker
docker run -p 7865:7865 -v $PWD/data:/data fooocus:latest

# CUDA version (requires nvidia-container-toolkit)
nix run .#buildDockerCuda
docker run --gpus all -p 7865:7865 -v $PWD/data:/data fooocus:cuda

Fooocus stores persistent data in ~/.config/fooocus:

~/.config/fooocus/
├── app/              # Fooocus application code
├── venv/             # Python virtual environment
├── models/           # Downloaded models
│   ├── checkpoints/  # Main model files
│   ├── loras/        # LoRA models
│   ├── embeddings/   # Text embeddings
│   ├── controlnet/   # ControlNet models
│   └── ...
└── outputs/          # Generated images

The flake automatically detects your GPU:

• NVIDIA: Uses CUDA (cu124 by default, configurable via CUDA_VERSION env var)
• Apple Silicon: Uses Metal Performance Shaders (MPS) acceleration
• Intel Mac: Falls back to CPU-only mode
• CPU: Falls back to CPU-only mode on any platform

Override CUDA version:

CUDA_VERSION=cu121 nix run .

Supported CUDA versions: cu118, cu121, cu124, cpu

On M1/M2/M3/M4 Macs, PyTorch automatically uses MPS (Metal Performance Shaders) for GPU acceleration. No additional configuration is required.

Note: First run will download macOS-specific PyTorch wheels which may take a few minutes.

nix develop

This provides:

• Python 3.12 environment
• Development tools (git, shellcheck, shfmt)
• Linting tools (ruff, pyright)
• Nix formatting (nixfmt-rfc-style)

# Run all checks
nix flake check

# Individual checks
nix run .#lint        # Python linting
nix run .#format      # Python formatting
nix run .#type-check  # Type checking
nix run .#check-all   # All Python checks

Fooocus is tracked as a flake input, so updating to HEAD is simple:

nix flake update fooocus-src

If port 7865 is already in use, the launcher will offer options to:

1. Open browser to existing instance
2. Use a different port
3. Kill the process using the port

If CUDA isn't detected despite having an NVIDIA GPU:

1. Ensure NVIDIA drivers are installed
2. Check that nvidia-smi works
3. Try setting CUDA_VERSION explicitly

The first run downloads:

• Python dependencies
• PyTorch with appropriate GPU support
• Base models (automatically on first generation)

Subsequent runs use cached dependencies.

This issue is now handled automatically by the flake. The launcher sets the path_outputs environment variable to bypass Gradio's symlink security restrictions.

If you still encounter {"detail":"File not allowed: .../log.html"}, manually edit ~/.config/fooocus/app/config.txt:

"path_outputs": "/home/YOUR_USER/.config/fooocus/outputs"

Change from .../app/outputs to the real path (without the app/ symlink), then restart Fooocus.

This project contains two components with different licenses:

• Packaging code (Nix expressions, shell scripts): MIT License - see LICENSE
• Fooocus application: GPL-3.0 - see the Fooocus repository

The Nix package metadata reflects GPL-3.0 as that is the license of the packaged software.

• Fooocus by lllyasviel