qwen-image-mps 0.1.0

Generate images with Qwen-Image on Apple Silicon (MPS) and other devices

Qwen Image (MPS/CUDA/CPU)

Generate and edit images from text prompts using the Hugging Face Diffusers pipeline for Qwen/Qwen-Image, with automatic device selection for Apple Silicon (MPS), NVIDIA CUDA, or CPU fallback.

Features

• Auto device selection: prefers MPS (Apple Silicon), then CUDA, else CPU
• Simple CLI: provide a prompt and number of steps
• Image generation: create new images from text prompts
• Image editing: modify existing images using text instructions
• Timestamped outputs: avoids overwriting previous generations
• Fast mode: 8-step generation using Lightning LoRA (auto-downloads if needed)
• Ultra-fast mode: 4-step generation using Lightning LoRA (auto-downloads if needed)
• Multi-image generation: generate multiple images in one run with --num-images

Example

Example result you can create with this project:

Installation

Option 1: Install from PyPI (Recommended)

Install the package using pip:

pip install qwen-image-mps

Then run it directly from the command line:

qwen-image-mps --help
qwen-image-mps generate --help  # For image generation
qwen-image-mps edit --help      # For image editing

Option 2: Direct script execution with uv

You can run this script directly using uv run without installation - it will install all dependencies automatically in an isolated environment:

uv run https://raw.githubusercontent.com/ivanfioravanti/qwen-image-mps/refs/heads/main/qwen-image-mps.py --help

Or download the file first:

curl -O https://raw.githubusercontent.com/ivanfioravanti/qwen-image-mps/refs/heads/main/qwen-image-mps.py
uv run qwen-image-mps.py --help

Option 3: Install from source

Clone the repository and install in development mode:

git clone https://github.com/ivanfioravanti/qwen-image-mps.git
cd qwen-image-mps
pip install -e .

Note: The first time you run the tool, it will download the 57.7GB model from Hugging Face and store it in your ~/.cache/huggingface/hub/models--Qwen--Qwen-Image directory.

Usage

After installation, use the qwen-image-mps command with either generate or edit subcommands:

qwen-image-mps --help
qwen-image-mps generate --help  # For image generation
qwen-image-mps edit --help      # For image editing

Image Generation Examples:

# Default prompt and steps
qwen-image-mps generate

# Custom prompt and fewer steps
qwen-image-mps generate -p "A serene alpine lake at sunrise, ultra detailed, cinematic" -s 30

# Fast mode with Lightning LoRA (8 steps)
qwen-image-mps generate -f -p "A magical forest with glowing mushrooms"

# Ultra-fast mode with Lightning LoRA (4 steps)
qwen-image-mps generate --ultra-fast -p "A magical forest with glowing mushrooms"


# Custom seed for reproducible generation
qwen-image-mps generate --seed 42 -p "A vintage coffee shop"

# Generate multiple images (incrementing seed per image when seed is provided)
qwen-image-mps generate -p "Retro sci-fi city skyline at night" --num-images 3 --seed 100

# Generate multiple images with a fresh random seed for each image (omit --seed)
qwen-image-mps generate -p "Retro sci-fi city skyline at night" --num-images 3

Image Editing Examples:

# Basic image editing
qwen-image-mps edit -i input.jpg -p "Change the sky to sunset colors"

# Edit with custom steps
qwen-image-mps edit -i photo.png -p "Add snow to the mountains" -s 30

# Edit with custom output filename
qwen-image-mps edit -i landscape.jpg -p "Make it autumn colors" -o autumn_landscape.png

# Edit with custom seed and steps
qwen-image-mps edit -i portrait.jpg -p "Change hair color to blonde" --seed 123 -s 30

If using the direct script with uv, replace qwen-image-mps with uv run qwen-image-mps.py in the examples above.

Command Arguments

Generate Command Arguments

• -p, --prompt (str): Prompt text for image generation.
• -s, --steps (int): Number of inference steps (default: 50).
• -f, --fast: Enable fast mode using Lightning LoRA for 8-step generation.
• --ultra-fast: Enable ultra-fast mode using Lightning LoRA v1.0 for 4-step generation.
• --seed (int): Random seed for reproducible generation (default: 42). If not explicitly provided and generating multiple images, a new random seed is used for each image.
• --num-images (int): Number of images to generate (default: 1).

Edit Command Arguments

• -i, --input (str): Path to the input image to edit (required).
• -p, --prompt (str): Editing instructions (required).
• -s, --steps (int): Number of inference steps (default: 50).
• --seed (int): Random seed for reproducible generation (default: 42).
• -o, --output (str): Output filename (default: edited-.png).

What the script does

Image Generation

• Loads Qwen/Qwen-Image via diffusers.DiffusionPipeline
• Selects device and dtype:
  • MPS: bfloat16
  • CUDA: bfloat16
  • CPU: float32
• Uses a light positive conditioning suffix for quality
• Generates at a 16:9 resolution (default 1664x928)
• Saves the output as image-YYYYMMDD-HHMMSS.png for a single image, or image-YYYYMMDD-HHMMSS-1.png, image-YYYYMMDD-HHMMSS-2.png, ... when using --num-images
• Prints the full path of the saved image

Image Editing

• Loads Qwen/Qwen-Image-Edit via QwenImageEditPipeline for image editing
• Takes an existing image and editing instructions as input
• Applies transformations while preserving the original structure
• Saves the edited image as edited-YYYYMMDD-HHMMSS.png or custom filename
• Prints the full path of the edited image

Fast Mode & Ultra-Fast Mode (Lightning LoRA)

Fast Mode (-f/--fast)

When using the -f/--fast flag, the tool:

• Automatically downloads the Lightning LoRA v1.1 from Hugging Face (cached in ~/.cache/huggingface/hub/)
• Merges the LoRA weights into the model for accelerated generation
• Uses fixed 8 inference steps with CFG scale 1.0
• Provides ~6x speedup compared to the default 50 steps

Ultra-Fast Mode (--ultra-fast)

When using the --ultra-fast flag, the tool:

• Automatically downloads the Lightning LoRA v1.0 from Hugging Face (cached in ~/.cache/huggingface/hub/)
• Merges the LoRA weights into the model for maximum speed generation
• Uses fixed 4 inference steps with CFG scale 1.0
• Provides ~12x speedup compared to the default 50 steps
• Ideal for rapid prototyping and iteration

The fast implementation is based on Qwen-Image-Lightning. The Lightning LoRA models are available on HuggingFace at lightx2v/Qwen-Image-Lightning.

Note: Fast modes are currently only available for image generation, not editing.

Notes and tweaks

• Aspect ratio / resolution: The script currently uses the 16:9 entry from an aspect_ratios map. You can change the selection in the code where width, height is set.
• Determinism: Use the --seed parameter to control the random generator for reproducible results. On MPS, the random generator runs on CPU for improved stability.
• Performance: If you hit memory or speed issues, try reducing --steps.

Troubleshooting

• If you see "Using CPU" in the console on Apple Silicon, ensure your PyTorch build includes MPS and you are running on Apple Silicon Python (not under Rosetta).
• If model download fails or is unauthorized, log in with huggingface-cli login or accept the model terms on the Hugging Face model page.

Development

To contribute or modify the tool:

1. Clone the repository:

git clone https://github.com/ivanfioravanti/qwen-image-mps.git
cd qwen-image-mps

2. Install in development mode with dev dependencies:

pip install -e ".[dev]"

3. Install pre-commit hooks:

pre-commit install

The project uses:

• black for code formatting
• isort for import sorting
• ruff for linting
• Pre-commit hooks for code quality

Repository contents

• src/qwen_image_mps/: Main package source code
• qwen-image-mps.py: Script wrapper for direct URL execution
• pyproject.toml: Package configuration and dependencies
• uv.lock: Locked dependencies for reproducible builds
• .github/workflows/: CI/CD pipelines for testing and publishing
• example.png: Sample generated image