Qwen-Image-Edit-2511-LoRAs-Fast-Lazy-Load

A Gradio-based demonstration for the Qwen/Qwen-Image-Edit-2511 model with lazy-loaded LoRA adapters for advanced single- and multi-image editing. Supports 7+ specialized LoRAs including photo-to-anime, multi-angle camera control, pose transfer (Any-Pose), upscaling, style transfer, light migration, and manga tone. Features fast inference (4 steps default) with Flash Attention 3 and dynamic adapter loading to optimize memory.

Features

• Multi-Image Support: Upload one or more images via gallery (e.g., subject + reference for pose/style transfer).
• Lazy LoRA Loading: 7 adapters load on-demand only when selected, minimizing VRAM usage.
• Advanced Editing Tasks:
  • Photo-to-Anime: Realistic to anime style
  • Multiple-Angles: Camera rotation/view changes
  • Any-Pose: Precise pose transfer from reference
  • Upscaler: High-resolution enhancement
  • Style-Transfer: Apply artistic style from reference
  • Light-Migration: Match lighting/color tone
  • Manga-Tone: Black-and-white manga aesthetic & more.
• Rapid Inference: Flash Attention 3 enabled; 4-step default with bfloat16.
• Auto-Resizing: Preserves aspect ratio up to 1024px max edge (multiples of 8).
• Custom Theme: OrangeRedTheme with clean, responsive layout.
• Examples: 7 curated multi/single-image scenarios.
• Queueing: Up to 30 concurrent jobs.

Note: This is an experimental Space for the newer Qwen-Image-Edit-2511 model. For stable performance, consider the 2509 version.

Prerequisites

• Python 3.10 or higher.
• CUDA-compatible GPU (required for bfloat16 and Flash Attention 3).
• Stable internet for initial model/LoRA downloads.

Installation

1. Clone the repository:

   git clone https://github.com/PRITHIVSAKTHIUR/Qwen-Image-Edit-2511-LoRAs-Fast-Lazy-Load.git
   cd Qwen-Image-Edit-2511-LoRAs-Fast-Lazy-Load
   

2. Install dependencies: Create a requirements.txt file with the following content, then run:

   pip install -r requirements.txt
   

   requirements.txt content:

    git+https://github.com/huggingface/accelerate.git
    git+https://github.com/huggingface/diffusers.git
    git+https://github.com/huggingface/peft.git
    transformers==4.57.3
    huggingface_hub
    sentencepiece
    torchvision
    kernels
    spaces
    hf_xet
    gradio==6.17.3
    torch==2.11.0
    numpy
    av
   

3. Start the application:

   python app.py
   

   The demo launches at http://localhost:7860.

Running with uv (Recommended)

uv is an ultra-fast Python package installer and dependency resolver. It isolates execution contexts instantly and securely.

1. Install uv

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

2. Clone and Synchronize the Workspace

git clone https://github.com/PRITHIVSAKTHIUR/Qwen-Image-Edit-2511-LoRAs-Fast-Lazy-Load.git
cd Qwen-Image-Edit-2511-LoRAs-Fast-Lazy-Load
uv sync

3. Launch the Web Interface

uv run app.py

Usage

1. Upload Images: Use gallery to add one or more images (e.g., person + pose reference).

2. Select Adapter: Choose from 7 styles (default: Photo-to-Anime).

3. Enter Prompt: Describe the edit (e.g., "Make the person do the exact same pose").

4. Configure (Optional): Expand "Advanced Settings" for seed, guidance, steps.

5. Edit Image: Click "Edit Image" to generate output.

Supported Adapters

Examples

Troubleshooting

• Adapter Loading: First selection downloads LoRA; monitor console.
• OOM: Reduce steps/resolution; clear cache with torch.cuda.empty_cache().
• Flash Attention Fails: Fallback to default; requires compatible CUDA.
• Gallery Input: Supports filepaths, tuples, or PIL objects.
• No Output: Ensure at least one valid image and descriptive prompt.

Contributing

Contributions welcome! Add new adapters to ADAPTER_SPECS, improve multi-image handling, or enhance prompts.

Repository: https://github.com/PRITHIVSAKTHIUR/Qwen-Image-Edit-2511-LoRAs-Fast-Lazy-Load.git

License

Apache License 2.0. See LICENSE for details.

Built by Prithiv Sakthi. Report issues via the repository.