Siwei Xia    Li Sun    Tiantian Sun    Qingli Li   
East China Normal University

• Release benchmark evaluation code
• Release Gradio user interface
• Update readme for detailed usage guide
• Release paper on arXiv

To set up the environment, run:

conda env create -f environment.yaml
conda activate draglora

To use DragLoRA with your own images:

1. Launch the interface:

python drag_ui.py

2. Follow these steps:

   • Upload your image to the left-most box
   • (Optional) Add a descriptive prompt below the image
   • Click "Train RecLoRA" to optimize for identity preservation
   • (Optional) Draw a mask to specify editable regions
   • Set handle-target points in the middle box:
     1. Click to place a handle point
     2. Click to place its target point
     3. Repeat for additional point pairs as needed
   • Click "Run" to process the image

3. Output and Storage:

   • Results appear in the right-most box
   • Temporary files are saved in "lora_tmp" (overwritten for each new image)
   • Input images, prompts, masks, points, and outputs are saved with unique names (e.g. 2025-05-17-0105-05) in "lora_tmp"

To evaluate the algorithm using benchmark data:

1. Navigate to the evaluation directory:

cd drag_bench_evaluation

2. Download and extract DragBench to "drag_bench_data"

3. Train reconstruction LoRA:

python run_lora_training.py

Results will be saved in "drag_bench_lora"

4. Run DragLoRA:

python run_dragbench_draglora.py

Results will be saved in "drag_results"

5. Evaluate performance:
   • Point matching accuracy (MD and m-MD):

   python run_eval_point_matching.py --eval_root drag_results

   • Image consistency (LPIPS, CLIP, MSE):

   python run_eval_similarity.py --eval_root drag_results

To simultaneously measure editability and consistency through two symmetric editing operations, first drag once following above instructions. After that, drag twice with:

1. Train reconstruction LoRA:

python run_lora_training.py --img_path drag_results

Results will be saved in "drag_bench_lora_for_drag_results"

2. Run DragLoRA:

python run_dragbench_draglora.py \
--img_dir drag_results \
--lora_dir drag_bench_lora_for_drag_results \
--save_dir drag_back_results

Results will be saved in "drag_back_results"

3. Evaluate performance by compare the similarity between the drag-back images and original images:

   python run_eval_similarity.py --eval_root drag_back_results

If you find our work useful, please cite our paper:

@inproceedings{xia2025draglora,
  title={DragLoRA: Online Optimization of LoRA Adapters for Drag-based Image Editing in Diffusion Model},
  author={Xia, Siwei and Sun, Li and Sun, Tiantian and Li, Qingli},
  booktitle={The International Conference on Machine Learning (ICML)},
  year={2025}
}

This work builds upon DragDiffusion. We thank the authors and all contributors to the open-source diffusion models and libraries.