metadata
dataset_info:
features:
- name: image
dtype: 'null'
- name: image_url
dtype: string
- name: instruction
sequence: string
- name: bbox
sequence:
sequence: float64
- name: point
sequence:
sequence: float64
- name: type
sequence: string
splits:
- name: train
num_bytes: 59376321
num_examples: 21988
download_size: 8450810
dataset_size: 59376321
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
Github | arXiv | HF Paper | Spaces | Datasets | Quick Start
ShowUI-web is a UI-grounding dataset focused on Web grounding, with screenshots and annotations originally sourced from OmniAct.
We developed a parser and collected 22K screenshots, retaining only visual-related elements such as those tagged with ‘Button’ or ‘Checkbox’ by removing static text.
If you find our work helpful, please consider citing our paper.
@misc{lin2024showui,
title={ShowUI: One Vision-Language-Action Model for GUI Visual Agent},
author={Kevin Qinghong Lin and Linjie Li and Difei Gao and Zhengyuan Yang and Shiwei Wu and Zechen Bai and Weixian Lei and Lijuan Wang and Mike Zheng Shou},
year={2024},
eprint={2411.17465},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2411.17465},
}