Simple Qwen Image Layered Inference Server

A FastAPI server for decomposing images into semantic layers using the Qwen-Image-Layered diffusion model.

Features

• Decompose input images into multiple semantic layers
• RESTful API with automatic documentation
• GPU acceleration support
• Returns layered outputs as a ZIP file containing PNG images

Requirements

• Python 3.11 or higher
• CUDA-capable GPU (recommended for performance)
• uv package manager

Setup

1. Clone the repository:

git clone https://github.com/creative-graphic-design/simple-qwen-image-layered-inference
cd simple-qwen-image-layered-inference

2. Install dependencies:

uv sync

Usage

Starting the Server

uv run fastapi run

The server will start at http://localhost:8000

API Documentation

Once the server is running, visit:

• Interactive API docs: http://localhost:8000/docs
• Alternative docs: http://localhost:8000/redoc

API Endpoint

POST /decompose

Decomposes an uploaded image into semantic layers.

Example Request

curl -X POST "http://localhost:8000/decompose?layers=4&num_inference_steps=50&seed=42" \
  -H "accept: application/zip" \
  -H "Content-Type: multipart/form-data" \
  -F "image_file=@your_image.jpg" \
  --output decomposed_layers.zip

Key Parameters

• image_file: The image to decompose (required)
• layers: Number of layers to generate (default: 4)
• num_inference_steps: Denoising steps (default: 50, higher = better quality but slower)
• seed: Random seed for reproducibility (default: 0)
• prompt: Optional text prompt to guide layer generation
• true_cfg_scale: Guidance scale for classifier-free guidance (default: 4.0)
• resolution: Output resolution, choose from 640 or 1024 (default: 640)

Response

The endpoint returns a ZIP file containing the decomposed layers as PNG images (layer_1.png, layer_2.png, etc.).

License

MIT