osunlp/GUI-Drag-dataset

GUI-Drag Dataset

Project Page | Paper | GitHub

GUI-Drag is a diverse dataset of 161K text dragging examples synthesized through a scalable pipeline. It is designed to advance Graphical User Interface (GUI) grounding beyond simple clicking, focusing on the essential interaction of dragging the mouse to select and manipulate textual content.

Our dataset is built upon the Uground, Jedi, and additional public paper-style academic document screenshots.

NOTE: Before you use this dataset, make sure you understand the logic of absolute coordinates and image processor for Qwen2.5-VL. This dataset is set with the image processor max_pixels=2116800, which is aligned with the Jedi model series. Make sure you also follow it in your training procedure; the performance will not be as expected.

Note

If you'd like to check the annotated coordinates on the screenshots, please refer to images_resized.zip. In this zip file, all images are preprocessed by the same pre-procession pipeline from Qwen2.5-VL

from transformers.models.qwen2_vl.image_processing_qwen2_vl_fast import (
    smart_resize as qwen_smart_resize,
)
import torchvision.transforms.functional as tvF

resized_height, resized_width = qwen_smart_resize(
    image.height,
    image.width,
    max_pixels=2116800,
    min_pixels=12544,
)
resized_image = tvF.resize(
    image_tensor,
    [resized_height, resized_width],
    interpolation=tvF.InterpolationMode.BILINEAR,
    antialias=True,
)

Citation

If you find our data, model, benchmark or the general resources useful, please consider citing:

@article{wu2025beyond,
  title={Beyond Clicking: A Step Towards Generalist GUI Grounding via Text Dragging},
  author={Wu, et al.},
  journal={arXiv preprint arXiv:2601.06031},
  year={2025}
}