MFlux Commander

Use flux models on your Apple Silcon machine.

filipstrand/mflux is a A MLX port of FLUX based on the Huggingface Diffusers implementation, and mflux-commander is tool that lets you manage all of the parameters, so you can do things like this:

./mflux-wrapper.py --prompt "a magical forest" --iterations 3

And you can change one parameter and see a couple more variations

mflux-wrapper.py --style ghibli

And if you like what you see, you can see interations of a specific seed variant:

./mflux-wrapper.py --seed 185769 --vary-steps 1,3,5,9

./mflux-wrapper.py --brainstorm "homegrown software"

Which returns

Generated Prompts:
----------------------------------------
1. A seedling sprouting from lines of code, with soil made of circuit board patterns, captured in macro photography with morning dew on the tender green leaves representing fresh, organic software development.
2. Weathered hands typing on a keyboard where the keys are tiny garden plots, each growing miniature digital elements like pixels and icons among the soil, with warm sunlight filtering across the workspace.
3. A close-up of intertwined roots and ethernet cables beneath a small flowering plant, where binary code is visible as texture on the roots, creating a seamless blend between technology and organic growth.
4. A terrarium-like environment inside a transparent computer case, where tiny plants grow among microchips and memory modules, with thin fiber-optic filaments glowing softly like irrigation lines.
5. A macro shot of a wooden desk workspace where handwritten code notes on paper are transforming into digital elements, with coffee stains merging into circuit patterns and a small potted plant casting pixel-shaped shadows.
----------------------------------------

And then you can try them out with out copy and paste.

./mflux-wrapper.py --run-prompt 2

Overview

MFlux Commander is a CLI tool for working directly with image generation models, designed for both local development and direct model exploration. Unlike polished products like Midjourney, it provides raw access to the underlying model capabilities, making it invaluable for understanding model behavior and prototyping new features.

A key feature is the ability to control the number of inference steps - from quick 1-step previews that generate in seconds to detailed 9-step renders that can take minutes. This granular control lets you rapidly prototype ideas with fast previews, then refine promising directions with higher-quality renders. The tool's session management and live preview features make it feel like a proper development environment, while its command-line interface and detailed logging support automation and systematic experimentation.

A powerful command-line tool for managing and exploring image generation with mflux-generate. Features include:

• Running multiple iterations with random or fixed seeds
• Automatic session management and history tracking
• Real-time progress monitoring with live preview
• Detailed metadata tracking and command reproduction
• Style management and reuse
• Interactive HTML galleries for reviewing results
• Seed and step variation modes for exploring different parameters

Example Sessions

Basic Image Generation

Generate 4 iterations of an image using the default settings:

./mflux-wrapper.py --prompt "a serene lake at sunset"

This will:

1. Create a new session directory or use a recent one
2. Generate 4 iterations with random seeds
3. Open a live preview in your browser
4. Save all results with metadata
5. Create an interactive HTML gallery

Style Exploration

Save and apply styles to maintain consistency:

# Save a style
./mflux-wrapper.py --save-style ghibli "in the style of Studio Ghibli, hand-drawn animation"

# List available styles
./mflux-wrapper.py --list-styles

# Apply a style
./mflux-wrapper.py --prompt "a magical forest" --style ghibli

Iterative Refinement

Explore variations of seeds and steps:

# Generate iterations with different step counts but same seed
./mflux-wrapper.py --prompt "a cyberpunk city" --seed 12345 --vary-steps 1,3,5,9

# Generate iterations with different random seeds
./mflux-wrapper.py --prompt "a cyberpunk city" --vary-seed --iterations 4

# Refine a specific result
./mflux-wrapper.py --prompt "a cyberpunk city" --seed 12345 --steps 5

Resolution Experiments

Try different aspect ratios and sizes:

# Quick preview with small format
./mflux-wrapper.py --prompt "mountain range" --landscape-sm

# Landscape format
./mflux-wrapper.py --prompt "mountain range" --landscape-xl

# Portrait format
./mflux-wrapper.py --prompt "tall tree in forest" --portrait-lg

# Custom resolution
./mflux-wrapper.py --prompt "city skyline" --resolution 1920x1080

Command Reference

Core Options

• --prompt TEXT - The generation prompt (reuses last prompt if not specified)
• --model [schnell|dev] - Model to use (default: schnell)
• --new - Force creation of new output directory
• --no-watch - Disable live preview
• --iterations N - Number of iterations to generate (default: 4)
• --vary-seed - Generate variations using different random seeds
• --seed N - Starting seed (random if not provided, cannot be used with --vary-seed)
• --steps N - Number of steps (defaults: 1 for schnell, 5 for dev)
• --metadata - Include generation metadata in output

Brainstorming Features

Generate and explore multiple prompt variations using Claude:

# Generate 5 prompt variations for a concept
./mflux-wrapper.py --brainstorm "home grown software"

# Run specific prompts from the last brainstorm
./mflux-wrapper.py --run-prompts "2,4,5"

The brainstorm feature:

• Uses Claude to generate 5 creative prompt variations
• Saves results in the current session directory
• Allows running specific prompts by their index numbers
• Creates separate run directories for each selected prompt
• Maintains all results in the same session for easy comparison

Resolution Options

Standard Formats:

• --resolution WxH - Custom resolution (default: 1024x1024)
• --landscape - 16:9 format (1024x576)
• --portrait - 3:4 format (768x1024)

Small Formats:

• --landscape-sm - Small 16:9 (512x288)
• --portrait-sm - Small 3:4 (384x512)
• --square-sm - Small square (512x512)

Large Formats:

• --landscape-lg - Large 16:9 (1536x864)
• --portrait-lg - Large 3:4 (1152x1536)

Extra Large Formats:

• --landscape-xl - XL 16:9 (2048x1152)
• --portrait-xl - XL 3:4 (1536x2048)
• --square-xl - XL square (2048x2048)

Style Management

• --save-style NAME DESC - Save a new style
• --list-styles - Show all saved styles
• --style NAME - Apply a saved style
• --style none - Clear any existing style

Advanced Features

• --vary-steps N,N,N - Comma-separated list of step counts to iterate through (requires --seed)
• --output-dir DIR - Custom output directory
• --metadata - Include metadata in output
• --brainstorm CONCEPT - Generate 5 prompt variations for a concept
• --run-prompt INDEX - Run specific prompts from last brainstorm (e.g., "2")

Output Structure

Each session creates a directory with format mflux_output_YYYYMMDD_HHMMSS containing:

• index.html - Main gallery of all runs
• brainstorm_results.json - Results from brainstorm sessions
• run_N/ - Directories for each generation run
  • index.html - Detailed view of the run
  • image_N.png - Generated images
  • image_N.png.json - Metadata (if enabled)
  • run_info.json - Run configuration and results

The tool automatically manages sessions, keeping related generations together and making it easy to track and reproduce results. Sessions are maintained for 4 hours before creating a new one, or you can force a new session with the --new flag.

Future actions

• Image to image
• Lora models
• Training your own model

Installation

Clone this repository and ensure the script has executable permissions:

git clone https://github.com/yourusername/mflux-wrapper.git
cd mflux-wrapper
chmod +x mflux-wrapper.py

Requirements

• Python 3.6+
• mflux - uv tool install -p 3.12 --upgrade mflux
• mflux-generate command installed and in your PATH
• Optional: live-server (install with npm install -g live-server) for live preview

License

MIT