[ICML 2026] ScreenParse

Official code release for ScreenParse: Moving Beyond Sparse Grounding with Complete Screen Parsing Supervision, accepted to ICML 2026.

ScreenParse studies complete screen parsing for computer-use agents: recovering visible UI elements, their locations, semantic types, text, and hierarchy from a screenshot. The project introduces:

• ScreenParse, a large-scale dataset with dense UI annotations over web screenshots.
• Webshot, the automated data generation and filtering pipeline used to build ScreenParse.
• ScreenVLM, a compact vision-language model trained for structured screen parsing.

Repository Layout

.
|-- webshot/      # Dataset generation, refinement, export, and evaluation toolkit
|-- docs/         # Project website published with GitHub Pages
|-- assets/       # Figures used by this repository README
`-- LICENSE

The runnable code currently lives in webshot/. Its README contains installation, dataset generation, VLM refinement, YOLO export, and evaluation instructions.

Quick Start

cd webshot
uv sync
uv run playwright install chromium
uv run wsd --help

To run a small Webshot pipeline example:

cd webshot
uv run wsd pipeline --urls examples/urls_sample.csv --workers 4

See webshot/README.md and webshot/USAGE.md for detailed usage.

Links

• Project page: https://saidgurbuz.github.io/screenparse/
• Paper: https://openreview.net/forum?id=H8KE7sudq5
• Dataset: https://huggingface.co/datasets/docling-project/screenparse
• ScreenVLM: https://huggingface.co/docling-project/ScreenVLM
• ScreenParser: https://huggingface.co/docling-project/ScreenParser
• Webshot toolkit: webshot/

Citation

The official ICML proceedings citation will be added when available. For now, please cite:

@misc{gurbuz2026screenparse,
      title={ScreenParse: Moving Beyond Sparse Grounding with Complete Screen Parsing Supervision},
      author={A. Said Gurbuz and Sunghwan Hong and Ahmed Nassar and Marc Pollefeys and Peter Staar},
      year={2026},
      eprint={2602.14276},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.14276},
      note={Accepted to ICML 2026}
}

License

This repository is released under the MIT License. See LICENSE for details.