🎨 Qwen Image for macOS

Fast AI image generation for Apple Silicon. Native CLI with MPS acceleration or containerized with Cog. Lightning LoRA for 4–8 step generation.

🚀 Quick Start

Choose your approach:

Option A: Native CLI (⚡ 2 minutes - Recommended for Apple Silicon)

git clone https://github.com/zsxkib/qwen-image-macos.git
cd qwen-image-macos
pip install -r requirements.txt
python qwen.py generate "cyberpunk cityscape" --ultra-fast

Option B: Containerized with Cog (🐳 Reproducible - Works everywhere)

git clone https://github.com/zsxkib/qwen-image-macos.git
cd qwen-image-macos
python3 precache.py  # downloads model once (~10 min)
cog predict -i prompt="cyberpunk cityscape" --output city.png

Note: Cog runs ~30x slower on Apple Silicon (60+ min) due to x86_64 emulation. Use native CLI for speed, cog for reproducibility/deployment.

⚡ Performance Comparison

🎯 Native CLI Usage

# Ultra-fast (4 steps with Lightning LoRA)
python qwen.py generate "cyberpunk cityscape" --ultra-fast

# Fast mode (8 steps)
python qwen.py generate "mountain landscape" --fast

# Custom settings
python qwen.py generate "robot on mars" --steps 20 --seed 42

# Test your setup
python qwen.py test

🧰 Requirements

• Apple Silicon Mac (M1/M2/M3/M4)
• Python 3.8+
• 32GB+ RAM recommended (64GB+ ideal)

✅ Features

• Native MPS acceleration on macOS
• Simple, single-file CLI (qwen.py)
• Auto-opens generated image in Preview on macOS
• Reproducible seeds and custom sizes

🐳 Cog Usage (Containerized)

Prerequisites:

• Docker Desktop (macOS: increase memory to 64GB+ in Settings → Resources)
• Cog CLI: brew install replicate/cog/cog

Quick workflow:

# 1. Pre-download model (recommended, ~10 min)
python3 precache.py

# 2. Generate images
cog predict -i prompt="robot on mars" --output robot.png
cog predict -i prompt="cyberpunk city" -i steps=10 --output city.png

When to use cog:

• ✅ Linux/NVIDIA: Fast with CUDA acceleration
• ✅ Reproducible deployments: Exact same environment everywhere
• ✅ Replicate cloud: Deploy to replicate.com
• ❌ Apple Silicon: Use native CLI instead (30x faster)

🔧 Technical Details

• Model: Qwen-Image (57GB)
• Acceleration: Lightning LoRA for 4-8 step generation
• Memory: Attention slicing + VAE tiling for efficiency
• Cache: Models stored in model_cache/ (~63GB)

🛠️ Troubleshooting

Check Apple Silicon GPU:

python -c "import torch; print('MPS available:', torch.backends.mps.is_available())"

Common fixes:

• MPS not available → Update macOS/PyTorch
• Images look unfinished → Increase --steps to 20-30
• Cog OOM errors → Increase Docker memory to 64GB+

🎆 Example Output

Native CLI generates high-quality images in ~2 minutes:

---

Built for Apple Silicon. Optimized for speed. Ready for deployment.