Qwen-Image Generator

A cross-platform desktop application providing a user-friendly GUI for Alibaba's Qwen-Image diffusion model. Built with Python/Gradio backend and designed for future Tauri desktop distribution.

🚀 Features

• 🎨 Easy Image Generation: Simple web interface for text-to-image generation
• 🧠 Smart Memory Management: CPU offloading for large models with GPU acceleration
• ⚡ Auto-Configuration: Automatic GPU/CPU detection and optimization
• 📦 One-Click Setup: Automatic model download and caching
• 🔧 Advanced Controls: Multiple aspect ratios, CFG scale, steps, and seed controls
• 📊 Real-Time Monitoring: Comprehensive logging and progress tracking
• 🌐 Cross-Platform: Works on Windows, macOS, and Linux
• 🔄 True Async Generation: ThreadPoolExecutor ensures UI interactions don't interrupt generation

🖼️ Generated Image Quality

The Qwen-Image model produces high-quality images with:

• High Resolutions: Up to 2048×2048 (1:1) and 2048×1152 (16:9) maximum quality
• Intuitive UI: Separate aspect ratio and resolution selection with live preview
• Multiple Aspect Ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3 with 4 quality levels each
• Advanced Controls: CFG scale, negative prompts, custom seeds
• Professional Quality: 4K, ultra HD, cinematic composition enhancement

🛠️ Quick Start

Prerequisites

• Python 3.10+ with pip
• PowerShell (Windows) for best compatibility
• 16GB+ RAM recommended
• NVIDIA GPU with 8GB+ VRAM (optional, but recommended)

Installation

1. Clone the repository:

git clone https://github.com/DexterLagan/qwen-image-generator.git
cd qwen-image-generator

2. Install dependencies:

pip install -r requirements.txt

3. Launch the application:

python qwen_gui.py

4. Open your browser to http://localhost:7860

5. First run: Click "Download & Load Model" (one-time ~4GB download)

📦 Model Variants

Two versions of the Qwen-Image model are available:

• FP8 (~20GB) – https://huggingface.co/Qwen/Qwen-Image
• FP16 (~40GB) – https://huggingface.co/Qwen/Qwen-Image-FP16

Select the desired variant in the application before downloading.

🧠 Advanced Memory Management

Automatic Optimization

The application automatically detects your hardware and optimizes accordingly:

• GPU Available (16GB+ VRAM): Full GPU acceleration
• GPU Available (<16GB VRAM): CPU offloading with GPU compute
• CPU Only: Sequential CPU offloading with memory optimization

Features

• CPU Offloading: Stores large models in RAM, uses GPU for computation
• Attention Slicing: Reduces memory usage during generation
• Smart Caching: Reuses downloaded model files
• Memory Monitoring: Real-time RAM/VRAM usage reporting

📁 Project Structure

qwen-image-generator/
├── qwen_gui.py                           # 🚀 Main Gradio application
├── qwen_sample_hf.py                     # 📖 Official HuggingFace reference
├── qwen_gradio_proto_chatgpt_wrong.py    # 📁 Legacy prototype (archived)
├── test_launch.py                        # 🧪 Basic functionality test
├── requirements.txt                      # 📦 Python dependencies
├── console.log                           # 📋 Application logs
├── model/cache/                          # 🤖 Downloaded model files
├── output/                               # 🖼️ Generated images
└── screenshots/                          # 📸 Documentation images

⚙️ Technical Details

Model Integration

• API: Uses diffusers.DiffusionPipeline (not transformers)
• Model: Qwen/Qwen-Image from Hugging Face
• Parameters: Official true_cfg_scale, aspect ratios, and resolutions
• Caching: Intelligent model file verification and reuse

Compatibility

• PyTorch: CUDA-enabled version for GPU acceleration
• Gradio: Latest version with compatibility fixes
• Cross-Platform: Windows tested, Linux/macOS compatible

Performance Optimizations

• Environment Variables: CUDA memory optimization settings
• Progressive Loading: Components loaded sequentially
• Error Recovery: Multiple fallback strategies for reliability

🔧 Configuration

Environment Variables

The application sets optimal configurations automatically:

# Gradio analytics disabled for privacy
os.environ['GRADIO_ANALYTICS_ENABLED'] = 'False'
# Note: CUDA expandable_segments removed for platform compatibility

Model Storage

• Windows: model/cache/models--Qwen--Qwen-Image/
• Output: output/qwen_image_YYYYMMDD_HHMMSS_seedXXXXXX.png

🚀 Future Development

Planned Tauri Integration

• Native Desktop App: Rust-based wrapper for system integration
• Embedded Runtime: Self-contained Python environment
• Cross-Platform Distribution: .exe, .app, and .deb packages
• Enhanced UX: Native file dialogs, system tray, auto-updater

Roadmap

• Tauri desktop wrapper implementation
• Batch image generation
• Image gallery and history
• Advanced prompt templates
• Model fine-tuning support

📦 Package Versions

Exact Tested Versions

The following versions have been tested and work together:

torch==2.7.1
torchvision==0.22.1
transformers==4.55.0
accelerate==1.10.0
safetensors==0.5.3
gradio==5.41.1
psutil==7.0.0

CUDA Installation (Windows/Linux)

For GPU support (though CPU-only mode is recommended for this model):

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

🐛 Troubleshooting

Critical Issues & Solutions

Model Loading Crashes at "Checkpoint Shards 6/9" (MOST COMMON)

Symptoms: Process exits/crashes during model loading without error message Cause: Insufficient Windows virtual memory for large model loading Solution:

1. Press Win + R, type sysdm.cpl, press Enter
2. Advanced tab → Performance Settings → Advanced → Virtual Memory Change
3. Uncheck "Automatically manage", select "Custom size"
4. Set Initial: 50000 MB, Maximum: 50000 MB
5. Click Set → OK → Restart computer Impact: CRITICAL - Model will not load without this fix

"CUDA Out of Memory" Errors

Symptoms: CUDA OOM errors despite CPU-only configuration Cause: Qwen-Image model (~20GB) too large for most GPUs Solution: Application automatically uses CPU-only mode User Action: Restart application if error persists

"Cannot copy out of meta tensor" Error

Symptoms: NotImplementedError during image generation Cause: Mixed GPU/CPU operations creating invalid tensor states Solution: Application uses pure CPU-only mode to prevent this User Action: Restart application, ensure CPU-only mode is active

Model File Corruption

Symptoms: Loading fails partway through, safetensors errors Cause: Interrupted downloads, disk issues, or network problems Solution: Click "🗑️ Clear Cache & Re-download" button Prevention: Ensure stable internet connection during initial download

Gradio Server Won't Start

Symptoms: "Unable to configure formatter" errors Cause: Logging conflicts between custom code and Gradio Solution: Fixed in current version with proper logging configuration User Action: Use PowerShell instead of CMD on Windows

System Requirements

Minimum Requirements

• RAM: 24GB available RAM (32GB total recommended)
• Storage: 10GB free space (4GB model + 6GB temporary)
• CPU: Multi-core processor (Intel i5/AMD Ryzen 5 or better)
• OS: Windows 10/11 (PowerShell required)

Critical Configuration

• Virtual Memory: 50GB Windows page file (ESSENTIAL)
• Python: 3.10+ with pip
• Network: Stable connection for initial 4GB download

Performance Notes

• Generation Time: 5-15 minutes per image on CPU
• Loading Time: 5-10 minutes first run (one-time)
• GPU Usage: Not recommended due to 20GB model size

Debug Information

• Logs: Check console.log for detailed operation information
• Status: UI shows real-time model status and error details
• Memory: Application monitors and reports RAM usage
• Recovery: Automatic cache clearing on corruption detection

Platform-Specific Notes

Windows

• Use PowerShell (not CMD) for better Unicode support
• Page file configuration is essential for model loading
• Antivirus: May quarantine model files, add exclusion for project folder

Linux

• Install packages may require sudo permissions
• Virtual memory configuration varies by distribution
• Monitor /tmp space during model loading

Common Error Messages

Performance Optimization

• Close other applications to free RAM during loading
• Use SSD storage if available for faster model loading
• Disable antivirus real-time scanning for project folder temporarily
• Increase page file on fastest drive (usually C:) for better performance

📄 License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.

Model License

The Qwen-Image model is licensed under the Qianwen License 1.0:

• ✅ Free for personal, research, and evaluation use
• ❌ Commercial use requires separate license from Alibaba

👏 Credits

• Model: Alibaba Qwen Team (Qwen-Image)
• Framework: Hugging Face (Diffusers, Gradio)
• Future Desktop: Tauri
• Author: Dexter Santucci

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

📞 Support

If you encounter any issues:

1. Check the console.log file for error details
2. Ensure you're using PowerShell on Windows
3. Verify your Python and dependency versions
4. Open an issue on GitHub with your log file