Dragging with Geometry: From Pixels to Geometry-Guided Image Editing

Xinyu Pu    Hongsong Wang    Jie Gui    Pan Zhou
Southeast University   |   Singapore Management University

⭐ Like our project? Don’t forget to star it on GitHub and stay tuned for future updates!

Disclaimer

This is a research project, NOT a commercial product. Users are granted the freedom to create images using this tool, but they are expected to comply with local laws and utilize it in a responsible manner. The developers do NOT assume any responsibility for potential misuse by users.

News and Updates

• 🚨 Released the implementation code of GeoDrag.
• 🚨 Paper Portal for Top Conferences in the Field of Artificial intelligence: CV_Paper_Portal
• 🚨 Abstract Paper Portal of ICLR 2026

Installation

Recommended environment: Linux system with an NVIDIA GPU.
The project has not been fully tested on other operating systems or hardware configurations.
Running our method currently needs around 6 GB of GPU memory. We will continue to optimize memory efficiency.

To install the required libraries, simply run the following command:

conda env create -f environment.yaml
conda activate geodrag

Next, download the pretrained Depth Anything V2 weights by running:

bash download.sh model

Run GeoDrag

Gradio demo

From the command line, run the following to launch the Gradio user interface:

python3 app.py

You can refer to GIF above for a step-by-step demonstration of how to use the UI.

CLI demo

To run GeoDrag directly from the command line, use:

python3 inference_drag.py \
    -c path/to/config \
    -i path/to/image \
    -m path/to/meta-data \
    -o path/to/output-dir
  
# example
# python3 inference_drag.py \
#     -c configs/base_configs.yaml \
#     -i datasets/DragBench/animals/JH_2023-09-14-1820-16/original_image.png \
#     -m datasets/DragBench/animals/JH_2023-09-14-1820-16/meta_data.pkl \
#     -o output

# -c: config file path
# -i: input image
# -m: metadata (drag points, mask, and prompt)
# -o: output directory

DRAGBENCH demo

Download DragBench into the folder "datasets" by running:

bash download.sh dragbench

To evaluate GeoDrag on the DRAGBENCH benchmark, run the following command:

python3 inference_dragbench.py \
    -c path/to/config \
    -b path/to/benchmark \
    -o path/to/output

# example
# python3 inference_dragbench.py \
#     -c configs/base_configs.yaml \
#     -b datasets/DragBench
#     -o outputs

Quantitative evaluation

Single-result evaluation

To compute quantitative evaluation metrics on a single result, run:

python3 -m evaluation single \
    -s '["MD", "IF", "DAI"]' \
    -i path/to/image \
    -m path/to/meta-data \
    -e path/to/edited-image 
  
# The `-s` argument specifies which scores to compute.
# You can choose any combination of the following:
#   MD  – Mean Distance
#   IF  – Image Fidelity
#   DAI – Dragging Accuracy Index
# For example:
#   -s '["MD"]'
#   -s '["MD", "IF"]'
#   -s '["MD", "IF", "DAI"]'

DRAGBENCH evaluation

To directly compute quantitative results on the DRAGBENCH benchmark, run:

python3 -m evaluation benchmark \
    -b path/to/benchmark-results

Results will be summarized and printed in the terminal.

Explanation for parameters

See parameters.md for additional information.

BibTeX

If you find our repo helpful, please consider leaving a star or cite our paper :)

@article{pu2025geodrag,
  title={Dragging with Geometry: From Pixels to Geometry-Guided Image Editing}, 
  author={Xinyu Pu and Hongsong Wang and Jie Gui and Pan Zhou},
  journal={arXiv preprint arXiv:2509.25740},
  year={2025}
}

Contact

For any questions on this project, please contact Xinyu (xinyupu@seu.edu.cn)

Acknowledgement

The code is built builds upon DragDiffusion, FastDrag and diffusers, thanks for their outstanding work!
We’d also like to express our appreciation to the amazing open-source community behind diffusion models, libraries, and research that inspired this work.

Related Links

• Drag Your GAN: Interactive Point-based Manipulation on the Generative Image Manifold
• DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing
• Emergent Correspondence from Image Diffusion
• DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models
• FreeDrag: Feature Dragging for Reliable Point-based Image Editing
• FastDrag: Manipulate Anything in One Step
• Drag Your Noise: Interactive Point-based Editing via Diffusion Semantic Propagation
• Lightning fast text-guided image editing via one-step diffusion