Qwen-Image-Edit-2509-LoRAs-Fast

Overview

Qwen-Image-Edit-2509-LoRAs-Fast is a high-performance, user-friendly web application built with Gradio that leverages the advanced Qwen/Qwen-Image-Edit-2509 model from Hugging Face for seamless image editing tasks. This app specializes in rapid, specialized edits using lightweight LoRA (Low-Rank Adaptation) adapters, enabling users to transform photos into anime styles, adjust camera angles for multi-view generation, restore lighting by removing harsh shadows, or relight scenes with custom illumination—all in just a few inference steps (as low as 4) for near-instant results. Powered by a custom SteelBlueTheme for an intuitive interface, the app automatically resizes input images to optimal dimensions (multiples of 8 for efficient diffusion processing) while preserving aspect ratios, and supports seed randomization for creative variations.

Features

• Specialized LoRA Adapters: Choose from four pre-loaded adapters:

  • Photo-to-Anime: Converts real-world photos into vibrant anime artwork.
  • Multiple-Angles: Rotates or switches camera perspectives (e.g., 45° left, top-down, wide-angle) for dynamic multi-view edits.
  • Light-Restoration: Removes unwanted shadows and artifacts for cleaner, evenly lit images.
  • Relight: Applies custom lighting effects, such as soft golden-hour filters or diffused illumination.

• Fast Inference: Optimized for speed with FlowMatchEulerDiscreteScheduler, bfloat16 precision, and as few as 4 steps—ideal for real-time prototyping.

• User-Friendly Interface:

  • Drag-and-drop image upload with automatic resizing.
  • Text prompt for precise edits (e.g., "transform into anime" or "rotate camera 180° upside down").
  • Advanced accordion for sliders on seed, guidance scale (1.0–10.0), and steps (1–50).
  • Built-in examples showcasing diverse use cases.

• Technical Optimizations:

  • Double-stream attention processor (QwenDoubleStreamAttnProcessorFA3) for enhanced efficiency.
  • Negative prompting to avoid common artifacts (e.g., blurriness, extra digits).
  • CUDA device detection and multi-GPU support via device_map='cuda'.

• Deployment-Ready: Integrates with Hugging Face Spaces for GPU-accelerated hosting, with progress tracking and error handling.

Installation

To run this app locally or in a custom environment:

1. Clone the repository:

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

2. Create a virtual environment (recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate
   

3. Install dependencies:

   pip install -r requirements.txt
   

4. Download required models and LoRAs (handled automatically on first run via from_pretrained):

   • Base model: Qwen/Qwen-Image-Edit-2509
   • Transformer: linoyts/Qwen-Image-Edit-Rapid-AIO (subfolder: transformer)
   • LoRAs:
     • autoweeb/Qwen-Image-Edit-2509-Photo-to-Anime (Qwen-Image-Edit-2509-Photo-to-Anime_000001000.safetensors)
     • dx8152/Qwen-Edit-2509-Multiple-angles (镜头转换.safetensors)
     • dx8152/Qwen-Image-Edit-2509-Light_restoration (移除光影.safetensors)
     • dx8152/Qwen-Image-Edit-2509-Relight (Qwen-Edit-Relight.safetensors)

5. Launch the app:

   python app.py  # Assuming the main script is saved as app.py
   

   The Gradio interface will open at http://127.0.0.1:7860.

Requirements

Install the following packages via pip (full requirements.txt below):

• git+https://github.com/huggingface/accelerate.git
• git+https://github.com/huggingface/diffusers.git
• git+https://github.com/huggingface/peft.git
• huggingface_hub
• sentencepiece
• transformers
• torchvision
• kernels
• spaces
• torch
• numpy
• Additional implicit deps: gradio, PIL (Pillow), qwenimage (custom from repo)

requirements.txt:

git+https://github.com/huggingface/accelerate.git
git+https://github.com/huggingface/diffusers.git
git+https://github.com/huggingface/peft.git
huggingface_hub
sentencepiece
transformers
torchvision
kernels
spaces
torch
numpy
gradio
Pillow

Hardware Notes:

• GPU recommended (NVIDIA with CUDA 11.8+ for optimal performance).
• ~8GB VRAM minimum for bfloat16 mode.
• On CPU, inference will be slower; set device='cpu' in code.

Usage

1. Upload an Image: Drag a photo (JPG/PNG) into the input field. It auto-resizes to 1024px max dimension (aspect-preserved, 8px multiples).

2. Enter Prompt: Describe the edit, e.g.:

   • Anime: "Transform into anime style with vibrant colors."
   • Angles: "Switch to top-down view."
   • Lighting: "Apply soft morning light from the left."

3. Select Adapter: Dropdown for LoRA style (default: Photo-to-Anime).

4. Tune Advanced Settings (optional):

   • Randomize seed for variations.
   • Adjust guidance scale (higher = stricter prompt adherence).
   • Increase steps for higher quality (but slower).

5. Run: Click "Run" to generate. Output appears alongside.

Troubleshooting

• CUDA Errors: Ensure torch.cuda.is_available() returns True. Check nvidia-smi for GPU usage.
• Model Download Fails: Verify Hugging Face token if gated models are accessed.
• Slow Inference: Reduce steps or use fewer adapters; enable torch.backends.cudnn.benchmark = True.
• Out-of-Memory: Lower resolution or batch size (single-image mode here).

Repository

• GitHub: https://github.com/PRITHIVSAKTHIUR/Qwen-Image-Edit-2509-LoRAs-Fast.git
• Hugging Face Space🤗: Qwen-Image-Edit-2509-LoRAs-Fast