lazarevtill-Qwen-Image-Edit-2511-Local-Interface

Qwen-Image-Edit-2511 Local Interface

A local Gradio interface for running Qwen-Image-Edit-2511 image editing models using GGUF quantized weights.

Features

• Local inference - Run entirely on your machine, no cloud required
• Multiple quantization options - From Q2_K (7.2GB) to full precision BF16 (40.9GB)
• Multi-GPU support - NVIDIA CUDA, Intel XPU (Arc GPUs), AMD ROCm, and CPU
• Multi-GPU detection - Detects all available GPUs, allows forcing specific device
• Model management - Download, delete, and manage models from the UI
• Image saving - Automatically saves all input/output images
• Network mode - Share on local network for access from other devices

Quick Start

Windows

:: First run - installs everything and starts the app
setup.bat

:: Network mode (accessible from other devices)
setup.bat --share

:: Force specific GPU
setup.bat --cuda      :: Use NVIDIA GPU
setup.bat --xpu       :: Use Intel Arc GPU
setup.bat --cpu       :: Use CPU only

:: Clean reinstall if something breaks
setup.bat --reset

Linux / macOS

# Make executable
chmod +x setup.sh

# First run - installs everything and starts the app
./setup.sh

# Network mode
./setup.sh --share

# Force specific GPU
./setup.sh --cuda      # Use NVIDIA GPU
./setup.sh --xpu       # Use Intel Arc GPU
./setup.sh --rocm      # Use AMD GPU (Linux only)
./setup.sh --cpu       # Use CPU only

# Clean reinstall
./setup.sh --reset

Then open http://127.0.0.1:7860 in your browser.

Requirements

• Python 3.10 or higher
• RAM 16GB+ recommended
• VRAM 8-24GB depending on model (or system RAM for CPU mode)
• Disk 7-41GB per model

Supported Hardware

Device	Backend	Performance
NVIDIA GPU (RTX 20/30/40 series)	CUDA	Fast
Intel Arc GPU (A770, A750, etc.)	XPU	Moderate
Intel Core Ultra (integrated)	XPU	Moderate
AMD GPU (with ROCm)	ROCm	Moderate
CPU	CPU	Very Slow

Available Models

Model	Size	Quality	Recommended For
Q2_K	7.22 GB	Lower	Testing, low VRAM
Q3_K_M	9.74 GB	Fair	Budget GPUs (8GB)
Q4_K_M	13.1 GB	Good	General use (Recommended)
Q5_K_M	15 GB	Better	Quality focus
Q6_K	16.8 GB	High	High quality
Q8_0	21.8 GB	Very High	Near-original
BF16/F16	40.9 GB	Original	Maximum quality

Usage

Command Line Options

python app.py [options]

Options:
  --host HOST    Host to bind to (default: 127.0.0.1, use 0.0.0.0 for network)
  --port PORT    Port to bind to (default: 7860)
  --share        Create a public Gradio link

Setup Script Options

setup.bat [options]   # Windows
./setup.sh [options]  # Linux/macOS

Options:
  --share, --public    Enable network mode (bind to 0.0.0.0)
  --local              Local only mode (default)
  --port PORT          Custom port number
  --reset              Remove .venv and reinstall everything
  --cuda               Force NVIDIA CUDA device
  --xpu                Force Intel XPU device
  --rocm               Force AMD ROCm device (Linux only)
  --cpu                Force CPU mode

Multi-GPU Systems

On systems with multiple GPUs (e.g., NVIDIA + Intel Arc), the setup script will:

1. Detect ALL available GPUs
2. Display them in the console
3. Auto-select by priority: CUDA > XPU > ROCm > CPU

To override the automatic selection, use the force flags:

# System has both NVIDIA and Intel Arc, but you want to use Intel:
setup.bat --xpu
./setup.sh --xpu

Directory Structure

img/
├── app.py              # Main application
├── setup.bat           # Windows setup/run script
├── setup.sh            # Linux/macOS setup/run script
├── requirements.txt    # Python dependencies
├── README.md           # This file
├── AGENTS.md           # AI agent instructions
├── models/             # Downloaded GGUF models
└── images/
    ├── inputs/         # Saved input images
    └── outputs/        # Generated output images

Tips for Best Results

1. Be specific with prompts: “Add a red Santa hat to the person” works better than “add hat”
2. Guidance scale 4-6: Higher values follow prompts more strictly
3. 40-50 inference steps: Good balance of quality and speed
4. Start with Q4_K_M: Best quality/performance ratio for most GPUs

Troubleshooting

Out of memory

• Try a smaller quantization (Q3_K_M, Q2_K)
• Reduce output image dimensions
• Close other GPU-intensive applications

Slow generation

• Reduce inference steps (20-30 for previews)
• Use a smaller model
• Ensure GPU is being used (check Devices tab)

Intel XPU not detected

• Install PyTorch with XPU support:

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

• Install Intel Extension for PyTorch:

  pip install intel-extension-for-pytorch
  

Only NVIDIA detected on multi-GPU system

• The setup scripts now detect all GPUs automatically
• Use --xpu flag to force Intel Arc GPU
• Use --cpu flag to force CPU mode

Virtual environment issues

• Run with --reset flag to clean reinstall:

  setup.bat --reset   # Windows
  ./setup.sh --reset  # Linux/macOS
  

Links

• Qwen-Image-Edit-2511 (Original Model)
• GGUF Quantized Models
• Qwen-Image GitHub
• Diffusers GGUF Documentation
• Intel Extension for PyTorch

License

This project uses the Qwen-Image-Edit-2511 model which is licensed under Apache 2.0.