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.

• 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
• Fast editing: 8-step and 4-step editing using Lightning LoRA
• Timestamped outputs: avoids overwriting previous generations
• Clean outputs: saves images to output/ by default (configurable with --outdir)
• 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
• Multi-image editing: blend multiple input images in a single edit command
• Batman mode: Add a LEGO Batman minifigure photobombing your images with --batman 🦇

Example generation result:

Example edit results showing winter transformation:

Anime photo-to-anime example using --anime:

Input (einstein.jpg) vs. anime output (anime-einstein.png):

Photo-to-anime mode is powered by Choomba’s (X: @begina​non) LoRA, available on Hugging Face as autoweeb/Qwen-Image-Edit-2509-Photo-to-Anime.

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

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

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.

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

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

Prefer a graphical interface? Launch the bundled Gradio app:

qwen-image-mps-gradio

This starts a local “Qwen-Image Studio” with two tabs:

• Generate – Enter prompts, toggle fast/ultra-fast Lightning LoRAs, add LEGO Batman, pick aspect ratios, and queue up to four images per run.
• Edit – Upload one or more images, enable anime (Photo-to-Anime) mode, apply Rapid-AIO fast edits, combine custom LoRAs, or add Batman photobombs.

Dark mode is enabled by default; use the “Toggle light / dark theme” button near the top of the UI to switch to the light palette on demand. (Behind the scenes, this updates Gradio’s __theme URL parameter the same way documented in Gradio’s theme guide.)

Outputs are still saved under output/ (or whatever directory you specify), mirroring the CLI defaults.

# 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"
# Or use the short form
qwen-image-mps generate -uf -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

# Generate with a custom LoRA for anime style
qwen-image-mps generate -p "A magical forest" --lora flymy-ai/qwen-image-anime-irl-lora

# Generate with custom LoRA and fast mode combined
qwen-image-mps generate -p "A futuristic city" --lora your-username/your-lora-model --fast

# Use a negative prompt to discourage artifacts
qwen-image-mps generate -p "Portrait photo" --negative-prompt "blurry, watermark, text, low quality"
qwen-image-mps generate -p "Portrait photo" -np "blurry, watermark, text, low quality"

# Batman mode: LEGO Batman photobombs your image!
qwen-image-mps generate -p "A magical forest with elves" --batman

# Combine Batman mode with ultra-fast generation
qwen-image-mps generate -p "A serene mountain lake" --batman --ultra-fast

# Specify aspect ratio (default is 16:9)
qwen-image-mps generate -f -p "Cozy reading nook, soft morning light" --aspect 1:1
qwen-image-mps generate -f -p "Tall cyberpunk city street, neon rain" --aspect 9:16

# Save images into a custom directory
qwen-image-mps generate -p "A cozy cabin in the woods" --outdir my-outputs

# Override CFG scale (default 4.0 normal, 1.0 fast/ultra-fast)
qwen-image-mps generate -p "Portrait photo" --cfg-scale 2.5

# Basic image editing (uses Rapid-AIO transformer, 4 steps by default)
qwen-image-mps edit -i input.jpg -p "Change the sky to sunset colors"

# Fast mode with Rapid-AIO transformer (8 steps)
qwen-image-mps edit -i photo.png -p "Add snow to the mountains" --fast

# Ultra-fast mode with Rapid-AIO transformer (4 steps)
qwen-image-mps edit -i landscape.jpg -p "Make it autumn colors" --ultra-fast
# Or use the short form
qwen-image-mps edit -i landscape.jpg -p "Make it autumn colors" -uf

# Edit with custom output filename
qwen-image-mps edit -i portrait.jpg -p "Change hair color to blonde" -o blonde_portrait.png

# Edit with custom seed and steps
qwen-image-mps edit -i scene.jpg -p "Add dramatic lighting" --seed 123 -s 30

# Edit with a custom LoRA for specific style
qwen-image-mps edit -i photo.jpg -p "Make it anime style" --lora flymy-ai/qwen-image-anime-irl-lora

# Edit with custom LoRA and ultra-fast mode combined
qwen-image-mps edit -i landscape.jpg -p "Add cyberpunk elements" --lora your-username/your-lora-model --ultra-fast

# Combine multiple input images in a single edit
qwen-image-mps edit -i input1.png input2.png -p "Merge the two scenes into a city skyline"

# Use a negative prompt during editing
qwen-image-mps edit -i photo.jpg -p "Studio portrait" --negative-prompt "blurry, watermark, text, low quality"

# Batman mode for editing: LEGO Batman photobombs your edited image!
qwen-image-mps edit -i photo.jpg -p "Change to sunset lighting" --batman

# Combine Batman mode with fast editing
qwen-image-mps edit -i portrait.jpg -p "Add dramatic shadows" --batman --fast

# Transform photo to anime style (prompt is optional with --anime) fast mode with Rapid-AIO transformer recommended 
qwen-image-mps edit -i photo.jpg --anime --fast

# Anime transformation with additional instructions
qwen-image-mps edit -i photo.jpg -p "make it colorful" --anime --fast

# Save edited image into a custom directory
qwen-image-mps edit -i photo.jpg -p "Add autumn colors" --outdir edits

# Override CFG scale for editing (default 4.0 normal, 1.0 fast/ultra-fast)
qwen-image-mps edit -i input.jpg -p "Studio portrait" --cfg-scale 2.0

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

• -p, --prompt (str): Prompt text for image generation.
• --negative-prompt (str): Text to discourage (negative prompt), e.g. "blurry, watermark, text, low quality".
• -s, --steps (int): Number of inference steps (default: 50).
• -f, --fast: Enable fast mode using Lightning LoRA for 8-step generation.
• -uf, --ultra-fast: Enable ultra-fast mode using Lightning LoRA v2.0 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).
• --lora (str): Hugging Face model URL or repo ID for additional LoRA to load (e.g., 'flymy-ai/qwen-image-anime-irl-lora' or full HF URL).
• --batman: Add a LEGO Batman minifigure photobombing your image in unexpected ways!
• --outdir (str): Directory to save generated images (default: ./output).
• --cfg-scale (float): Classifier-free guidance scale (overrides mode defaults).

• -i, --input (str): Path(s) to the input image(s) to edit (required). Provide multiple paths to blend results.
• -p, --prompt (str): Editing instructions. Optional when --anime is used (defaults to anime transformation prompt).
• --negative-prompt (str): Text to discourage in the edit (negative prompt).
• -s, --steps (int): Number of inference steps for normal editing (default: 50).
• -f, --fast: Enable fast mode using Rapid-AIO transformer for 4-step editing.
• -uf, --ultra-fast: Enable ultra-fast mode using Rapid-AIO transformer for 4-step editing.
• --seed (int): Random seed for reproducible generation (default: 42).
• -o, --output (str): Output filename (default: edited-.png).
• --outdir (str): Directory to save edited images (default: ./output). If --output is a basename, it is saved under this directory.
• --cfg-scale (float): Classifier-free guidance scale (overrides mode defaults).
• --lora (str): Hugging Face model URL or repo ID for additional LoRA to load (e.g., 'flymy-ai/qwen-image-anime-irl-lora' or full HF URL).
• --batman: Add a LEGO Batman minifigure photobombing your edited image!
• --anime: Transform photo to anime style using Photo-to-Anime LoRA. Can be combined with --fast or --ultra-fast for faster processing.

• 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 images under output/ by default. Filenames are image-YYYYMMDD-HHMMSS.png for a single image, or image-YYYYMMDD-HHMMSS-1.png, image-YYYYMMDD-HHMMSS-2.png, ... when using --num-images. Use --outdir to change the directory.
• Prints the full path of the saved image

• Loads Qwen/Qwen-Image-Edit-2509 via QwenImageEditPlusPipeline (falling back to QwenImageEditPipeline when needed)
• Uses Rapid-AIO transformer (linoyts/Qwen-Image-Edit-Rapid-AIO) for optimized fast inference
• Takes an existing image and editing instructions as input
• Applies transformations while preserving the original structure
• Uses 4 inference steps by default (optimized for Rapid-AIO transformer)
• Saves the edited image under output/ by default as edited-YYYYMMDD-HHMMSS.png, or to a custom filename.
• Prints the full path of the edited image

All image editing operations use the Rapid-AIO transformer (linoyts/Qwen-Image-Edit-Rapid-AIO) for optimized fast inference:

• Automatically loads the Rapid-AIO transformer (cached in ~/.cache/huggingface/hub/)
• Uses 4 inference steps by default with CFG scale 1.0
• Provides fast, high-quality image editing results

When using the -f/--fast flag with edit command:

• Uses Rapid-AIO transformer (already loaded by default)
• Uses 4 inference steps with CFG scale 1.0 (Rapid-AIO optimized)

When using the -uf/--ultra-fast flag with edit command:

• Uses Rapid-AIO transformer (already loaded by default)
• Uses 4 inference steps with CFG scale 1.0 (Rapid-AIO optimized)
• Both fast and ultra-fast modes use the same optimized 4-step inference

For image generation, the tool uses Lightning LoRA models:

• Fast mode (-f/--fast): Uses Lightning LoRA v2.0 for 8-step generation
• Ultra-fast mode (-uf/--ultra-fast): Uses Lightning LoRA v2.0 for 4-step generation
• The fast implementation is based on Qwen-Image-Lightning
• Lightning LoRA models are available on HuggingFace at lightx2v/Qwen-Image-Lightning

The --batman flag adds a fun twist to your image generation and editing by having a LEGO Batman minifigure photobomb your images! This feature works with both generate and edit commands.

When enabled, the tool randomly selects from various photobombing styles:

• LEGO Batman doing dramatic cape poses
• Sneaking into frame from the sides
• Peeking from behind objects
• Hanging upside down from the top
• Doing the Batusi dance
• Striking heroic poses
• Shouting his famous catchphrases

This feature adds a playful element to your images while keeping the main subject intact. The LEGO Batman appears small but noticeable, creating unexpected and humorous compositions.

The --lora argument allows you to load custom LoRA models from Hugging Face Hub:

# Using a repo ID
qwen-image-mps generate -p "Your prompt" --lora flymy-ai/qwen-image-anime-irl-lora

# Using a full Hugging Face URL
qwen-image-mps generate -p "Your prompt" --lora https://huggingface.co/flymy-ai/qwen-image-anime-irl-lora

# Combine with Lightning LoRA for both speed and style
qwen-image-mps generate -p "Your prompt" --lora your-username/style-lora --fast

The tool will automatically:

• Download the LoRA from Hugging Face Hub (cached locally)
• Find the appropriate safetensors file in the repository
• Merge the LoRA weights into the model
• Apply any Lightning LoRA if --fast or --ultra-fast is also specified

• Aspect ratio / resolution: Use --aspect to select output size. Available choices: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3. Default is 16:9.
• 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.

• 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.

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

The project includes integration tests that verify the image generation functionality:

# Run all tests
pytest tests/

# Run only fast tests (skip integration tests)
pytest -m "not slow"

# Run integration tests (generates real images with minimal steps)
pytest -m slow -v

# Run a specific test with verbose output and print statements
pytest tests/integration/test_generate_function.py::TestGenerateImageIntegration::test_generator_yields_expected_steps -v -s

# Run a specific test class
pytest tests/integration/test_generate_function.py::TestGenerateImageIntegration -v

Integration tests generate actual images (by default under output/) using ultra-fast mode (4 steps) to minimize execution time while ensuring the pipeline works correctly. Use -v for verbose output and -s to see print statements during test execution.

• 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