Jingyi Lu, Kai Han
Visual AI Lab, The University of Hong Kong

TLDR: Inpaint4Drag decomposes drag-based image editing into bidirectional warping (0.01s) and image inpainting (0.3s) (measured at 512×512). The method warps selected regions using bidirectional mapping, then fills revealed areas with inpainting models. Key advantages:

• Physics-inspired: Treats image regions as elastic materials for intuitive deformation
• Real-time interaction: Instant warping preview unlike existing methods requiring minutes per edit
• Universal adapter: Works with any inpainting model without modifications, inheriting future improvements

After cloning this repository, you can install the dependencies through the following steps:

# Create and activate environment
conda create -n inpaint4drag python=3.10 -y
conda activate inpaint4drag

# Install requirements 
pip install torch torchvision
pip install -r requirements.txt

# Install EfficientViT-SAM (Optional)
pip install git+https://github.com/mit-han-lab/efficientvit.git

After installing the requirements, you can simply launch the user inferface through:

python app.py

You can evaluate our method by (1) downloading the DragBench dataset, (2) running evaluation.py on DragBench-DR and DragBench-SR. You may use utils/evaluator.py to evaluate other drag editing methods.

# Install gdown and download drag_data.zip from Google Drive
pip install gdown
gdown 1rdi4Rqka8zqHTbPyhQYtFC2UdWvAeAGV

# Extract and clean up
unzip -q drag_data.zip
rm -rf drag_data.zip

Evaluate on DragBench-DR and DragBench-SR:

python evaluation.py --data_dir drag_data/dragbench-dr --output_dir output/dragbench-dr

python evaluation.py --data_dir drag_data/dragbench-sr --output_dir output/dragbench-sr

Found this repo useful? We'd be grateful if you could give it a star ⭐ or cite our paper!

@inproceedings{lu2025inpaint4drag,
 author    = {Jingyi Lu and Kai Han},
 title     = {Inpaint4Drag: Repurposing Inpainting Models for Drag-Based Image Editing via Bidirectional Warping},
 booktitle = {International Conference on Computer Vision (ICCV)},
 year      = {2025},
}

• Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold
• DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing
• The Blessing of Randomness: SDE Beats ODE in General Diffusion-based Image Editing
• InstantDrag: Improving Interactivity in Drag-based Image Editing
• FastDrag: Manipulate Anything in One Step