Qwen Image Generator

This is a simple but enhanced image generator for Qwen-Image model, released on Huggingface at Qwen/Qwen-Image.

The original model contains the suggested Python code to startup, setup the pipeline and invoke the model. This project is just a slightly enhanced version of the very same code, with bare minimum of improvements to make its usage more comfortable. This is not indended to be all-encompassing Diffusion models launcher; more like, "make it more convenient to generate different images, with little clutter".

Features

• Client-server architecture:
  • The server loads the model and pipeline to keep them hot, receives requests, and generates the images in a common directory (auto-incrementing the number of the image), one by one.
  • The client allows you to quickly add a job to generate an image (or multiple images with the same prompt and different seeds).
• Job-based architecture:
  • Single (preloaded, prewarmed) worker for sequential jobs; multiple jobs don't slow down each other;
  • The CUDA (or CPU if feasible) based generation occurs in a separate thread, so the server API is still accessible.
• Better Python architecture:
  • Uv-based project, for easier maintenance of dependencies/Python version.
  • Tested with Python 3.12.
• Reasonable defaults:
  • The image directory is automatically generated for you, if missing.
  • The file names are auto-incremented, and don't block each other.
  • The communication IP port is has the same defaults both for client and server (so you don't need to mention it explicitly); you can still configure it if needed.
• CPU offloading support (enabled by default), to use the model on GPUs with insufficient VRAM, and still run reasonably efficiently.

Installation

Install Uv, as recommended by their official documentation.

Clone the repository.

(If you are going to use CUDA-based Torch libraries) Install the Torch libraries with CUDA support:

uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

Install/sync the rest of the packages:

uv sync

Usage

Run the server:

uv run python qwenimage server

Add a job to generate an image, with some prompt:

uv run python qwenimage generate "A cute fluffy white kitten looks at you with blue eyes, while holding a huge cleaver in its paws. Comicbook style, pastel colors."

... or a job generating multiple images with the same prompt:

uv run python qwenimage generate -n 10 "A cute fluffy white kitten looks at you with blue eyes, while holding a huge cleaver in its paws. Comicbook style, pastel colors."```

Note: you can always run the commands with --help argument, to see their options:

uv run python qwenimage server --help
uv run python qwenimage generate --help

Contacts

Author: Alex Myodov