Qwen-Image-Edit-2511-LoRAs-Fast-Multi-Image-Rerun

Qwen-Image-Edit-2511-LoRAs-Fast-Multi-Image-Rerun

A Gradio-based experimental demonstration for the Qwen/Qwen-Image-Edit-2511 model with lazy-loaded LoRA adapters supporting multi-image input editing. Users can upload one or more images (gallery format) and apply advanced edits such as pose transfer, anime conversion, or camera angle changes via natural language prompts. Features integrated Rerun SDK visualization for interactive side-by-side comparison of inputs and outputs, with downloadable results and persistent .rrd recordings.

Features

• Multi-Image Input: Upload multiple images (e.g., reference pose + subject) for complex tasks like precise pose transfer.
• Lazy LoRA Loading: 3 specialized adapters (Photo-to-Anime, Multiple-Angles, Any-Pose) load only when selected to minimize memory usage.
• Rerun Visualization: Interactive viewer via gradio-rerun and rerun-sdk; logs all input images and final result; saves .rrd recordings in tmp_rerun/.
• Download Support: Direct download of generated output image.
• Rapid Inference: Flash Attention 3 enabled; 4-step default generations with bfloat16.
• Auto-Resizing: Preserves aspect ratio up to 1024px max edge (multiples of 8).
• Custom Theme: OrangeRedTheme with responsive layout and clean styling.
• Examples: 3 curated multi/single-image scenarios (anime style, angle change, pose transfer).
• Queueing: Up to 30 concurrent jobs.

Note: This is an experimental Space for the Qwen-Image-Edit-2511 model. For more 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-Multi-Image-Rerun.git
   cd Qwen-Image-Edit-2511-LoRAs-Fast-Multi-Image-Rerun
   

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
   gradio-rerun
   torchvision
   supervision
   rerun-sdk
   kernels
   spaces
   hf_xet
   torch
   numpy
   av
   

3. Start the application:

   python app.py
   

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

Usage

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

2. Select Adapter: Choose from “Photo-to-Anime”, “Multiple-Angles”, or “Any-Pose”.

3. Enter Prompt: Describe the desired edit (e.g., “Make the person in image 1 do the exact same pose of the person in image 2”).

4. Edit Image: Click “Edit Image”.

5. Output:

   • Interactive Rerun viewer showing all inputs and final result.
   • Download button for the generated PNG.
   • Recording saved as .rrd in tmp_rerun/.

Supported Adapters

Adapter	Use Case
Photo-to-Anime	Convert realistic photos to anime style
Multiple-Angles	Change camera viewpoint/angle
Any-Pose	Transfer precise pose from reference image(s)

Examples

Input Images	Prompt Example	Adapter
examples/B.jpg	“Transform into anime.”	Photo-to-Anime
examples/A.jpeg	“Rotate the camera 45 degrees to the right.”	Multiple-Angles
examples/P1.jpg + examples/P2.jpg	“Make the person in image 1 do the exact same pose of the person in image 2. Changing style/background undesirable. Match head tilt, eye gaze, arms/legs position exactly.”	Any-Pose

Rerun Viewer

• Logs paths: images/input_0, input_1, …, images/edited_result.
• Supports zoom, pan, and comparison.
• Recordings persist in tmp_rerun/ until manually cleared.

Troubleshooting

• Rerun Issues: Ensure gradio-rerun and rerun-sdk installed; handles SDK version differences.
• Adapter Loading: First use 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 uploaded.

Contributing

Contributions welcome! Fork the repo, add new adapters to ADAPTER_SPECS, enhance Rerun logging, or improve multi-image handling, and submit PRs with tests.

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

License

Apache License 2.0. See LICENSE for details.

Built by Prithiv Sakthi. Report issues via the repository.