Qwen-Image-Edit-2509-LoRAs-Fast-Lazy-Load

Qwen-Image-Edit-2509-LoRAs-Fast-Lazy-Load is a high-performance image editing and style-transfer platform built on top of the Qwen/Qwen-Image-Edit-2509 base model and an optimized transformer architecture (prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V4). The application integrates Flash Attention 3 (QwenDoubleStreamAttnProcessorFA3) to achieve low VRAM footprints and accelerated 4-step image manipulation.

Using a Lazy Loading design for LoRA adapters, the system dynamically downloads and fuses task-specific adapters on demand—including Photo-to-Anime, Multiple Angles, Light Restoration, Relight, Multi-Angle Lighting, Edit Skin, Next Scene, Flat Log, and Upscaling. The web workspace is served via a custom, single-page web app built with a FastAPI backend server (gradio.Server) and a dark-mode frontend interface featuring a dual-view canvas, A/B comparison slider, history filmstrip, and interactive prompt suggestions.

Key Features

• Lazy-Loaded Adapter Registry: On-demand downloading and weight-fusing for 11+ specialized LoRA adapters (e.g., Photo-to-Anime, Multiple-Angles, Light-Restoration, Relight, Multi-Angle-Lighting, Edit-Skin, Next-Scene, Flat-Log, Upscale-Image, Upscale2K, and Dotted-Illustration).
• Flash Attention 3 (FA3) Acceleration: Hooks natively into the QwenDoubleStreamAttnProcessorFA3 processor layer to accelerate cross-attention inference phases while reducing active GPU memory consumption.
• Text-Guided Image Editing: Offers camera angle rotations, shadow removal, uniform studio relighting, skin detail refinement, scene propagation, and 4K upscaling.
• Studio SPA Interface: An interactive single-page application built with modern vanilla web components—featuring an A/B image comparison slider, history filmstrip, quick prompt chips, and drag-and-drop file support.
• Smart Aspect Ratio Snapping: Automatically resizes uploaded images to stay within 1024px while snapping width and height to multiples of 8 to prevent shape mismatch errors during inference.

Repository Structure

├── examples/
│   ├── 1.jpg
│   ├── 10.jpeg
│   ├── 11.jpg
│   ├── 12.jpg
│   ├── 13.jpg
│   ├── 14.jpg
│   ├── 2.jpeg
│   ├── 4.jpg
│   ├── 5.jpg
│   ├── 6.jpg
│   ├── 7.jpg
│   ├── 8.jpg
│   ├── 9.jpg
│   ├── DI.jpg
│   └── ELS.jpg
├── qwenimage/
│   ├── __init__.py
│   ├── pipeline_qwenimage_edit_plus.py
│   ├── qwen_fa3_processor.py
│   └── transformer_qwenimage.py
├── app.py
├── index.html
├── LICENSE
├── pre-requirements.txt
├── pyproject.toml
├── README.md
├── requirements.txt
└── uv.lock

Installation and Requirements

To set up the Qwen-Image-Edit-2509-LoRAs-Fast-Lazy-Load environment locally, configure your system according to the specifications below. A modern CUDA-enabled GPU is required.

• Python Version: Minimum Python 3.12 is required; Python 3.12 or 3.14 is recommended.
• PyTorch Version: torch==2.11.0 or above is required for best compatibility.
• CUDA Version: CUDA 13.0 is recommended (--extra-index-url https://download.pytorch.org/whl/cu130), matching the environment used on the live Hugging Face demo.

Running with uv (Recommended)

uv is an ultra-fast Python package and project manager written in Rust. It ensures rapid virtual environment setup and exact dependency synchronization based on the uv.lock file.

Step 1 — Install uv

• macOS / Linux: curl -LsSf https://astral.sh/uv/install.sh | sh
• Windows: powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Step 2 — Clone the repository

git clone https://github.com/PRITHIVSAKTHIUR/Qwen-Image-Edit-2509-LoRAs-Fast-Lazy-Load.git
cd Qwen-Image-Edit-2509-LoRAs-Fast-Lazy-Load

Step 3 — Initialize the project and install dependencies

uv sync

Step 4 — Run the script

uv run app.py

---

Standard PIP Implementation

1. Update Package Manager Upgrade your local package manager:

pip install pip>=26.1.2

2. Install Core Dependencies Install the primary deep learning stack, transformer libraries, and core computing utilities listed in requirements.txt:

pip install -r requirements.txt

Core Requirements List (requirements.txt)

--extra-index-url https://download.pytorch.org/whl/cu130
torch==2.11.0
torchvision==0.26.0
transformers==5.14.1
accelerate==1.14.0
diffusers==0.39.0
peft==0.19.1
gradio==6.22.0
av==17.1.0
spaces==0.51.1
huggingface-hub==1.24.0
kernels==0.16.0

Usage

Once the web server initializes, open your browser to the local address output in your terminal (typically http://127.0.0.1:7860/).

1. Upload Asset: Drag and drop an image into the main canvas workspace, paste an image from your clipboard, or click the upload icon in the left rail.
2. Select Style / LoRA: Choose your target editing task from the Style / LoRA dropdown menu in the right inspector panel. The adapter weights will download lazily on first use.
3. Refine Instructions: Type your instructions inside the prompt field, or click one of the Quick Prompts chips to instantly fill it. Press ⌘/Ctrl + Enter or click Edit Image.
4. Compare & Chain: Use the Compare tool on the left rail to view an A/B slider of the before and after states. Click Use as Input to chain multiple edits sequentially.

Links and Source

• GitHub Repository: https://github.com/PRITHIVSAKTHIUR/Qwen-Image-Edit-2509-LoRAs-Fast-Lazy-Load.git
• Hugging Face Live Space: https://huggingface.co/spaces/prithivMLmods/Qwen-Image-Edit-2509-LoRAs-Fast
• License: Apache License 2.0
• Thanks to AK (aka akhaliq) for the initial PR #27, customization, and motivation.