Qwen Image Edit Docker

A Dockerized environment for Qwen-Image-Edit, served via a Gradio interface and API. This project allows you to easily edit images using text prompts, leveraging the power of local NVIDIA GPUs with FlowMatchEulerDiscreteScheduler for high-quality results.

Warning

Experimental Status: Improved generation quality has not yet been achieved on consumer hardware (tested on RTX 4080 16GB). Results may be unsatisfactory. Proceed with deployment primarily for testing or development purposes.

Features

• Easy Deployment: One-command startup using Docker Compose.
• GPU Acceleration: Optimized for NVIDIA GPUs with CUDA support, utilizing 4-bit quantization (NF4) and BFloat16 for efficiency.
• Advanced Scheduling: Uses FlowMatchEulerDiscreteScheduler for superior generation quality.
• Web Interface: User-friendly Gradio web UI with resolution controls.
• API Endpoint: REST-like API for integration with other applications.
• Persistent Cache: Hugging Face models are cached on the host.

Prerequisites

Before running this project, ensure you have the following installed:

• Docker Engine: Install Docker
• NVIDIA GPU: A generic NVIDIA GPU (e.g., RTX 4080) with at least 16GB VRAM recommended.
• NVIDIA Container Toolkit: Required for Docker to access your GPU. Installation Guide

Installation

1. Clone the repository:

   git clone https://github.com/artryazanov/qwen-image-edit-docker.git
   cd qwen-image-edit-docker

Usage

Option 1: Docker Compose (Recommended)

This is the simplest way to run the service.

1. Start the container:

   docker-compose up --build

   Add -d to run in detached mode (background).

2. Access the Web UI: Open your browser/client and go to: http://localhost:7860

3. Stop the container:

   docker-compose down

Option 2: Docker CLI

If you prefer using the Docker CLI directly:

1. Build the image:

   docker build -t my-qwen-edit .

2. Run the container:

   docker run --gpus all -p 7860:7860 \
       --ipc=host \
       -v ~/.cache/huggingface:/root/.cache/huggingface \
       my-qwen-edit

API Documentation

This service exposes an API endpoint via Gradio that you can use to integrate with your PHP or other applications.

Endpoint: /run/predict

Depending on the Gradio version, the endpoint might be available at /api/predict or /run/predict. Check the "Use via API" link in the footer of the web UI for the exact details for your version.

Method: POST

Content-Type: application/json

Payload Structure:

{
  "data": [
    "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAAAAAA6fptVAAAACklEQVR4nGNiAAAABgADNjd8qAAAAABJRU5ErkJggg==", 
    "make it look like a painting",
    "Square (1024x1024)"
  ]
}

• Argument 1: Input image (base64 string or URL).
• Argument 2: Text prompt.
• Argument 3: Resolution choice. Options are:
  • "Square (1024x1024)"
  • "HD (1280x720)"
  • "Full HD (1920x1080)"

Response Structure:

{
  "data": [
    "data:image/png;base64,..."
  ],
  "is_generating": false,
  "duration": 0.123,
  "average_duration": 0.123
}

Configuration

You can configure the deployment by modifying docker-compose.yml:

• Ports: Change "7860:7860" if you need to expose a different port on the host.
• Volumes: The command maps ~/.cache/huggingface to the container to persist model files. You can change the source path if your cache is located elsewhere.

Project Structure

• Dockerfile: Defines the environment and dependencies.
• app.py: Main entry point for the Gradio application.
• docker-compose.yml: Configuration for orchestration.
• .dockerignore: Files to exclude from the build context.

License

This project's code (Dockerfile, Python scripts) is licensed under the MIT License.

The underlying model Qwen-Image-Edit is subject to its own license from Alibaba Cloud (typically Apache 2.0 or Qwen License). By using this container, you also agree to the terms of the model's license.