README.md

---
title: GroundGUI GUI Grounding
emoji: 🖱️
colorFrom: pink
colorTo: blue
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Ground click coordinates on any GUI screenshot.
python_version: "3.12"
startup_duration_timeout: 1h
license: apache-2.0
models:
  - PrentisAI/GroundGUI-8B
---

# GroundGUI · GUI Grounding

This demo runs [PrentisAI/GroundGUI-8B](https://huggingface.co/PrentisAI/GroundGUI-8B),
an 8B GUI-grounding vision-language model fine-tuned from Qwen3-VL-8B-Instruct via
high-quality data curation and distillation-to-reinforcement learning.

Give it a GUI screenshot plus a short instruction, and it emits the click action
as a tool call with coordinates in a **normalized 0–1000 space** —
`{"name": "computer_use", "arguments": {"action": "left_click", "coordinate": [x, y]}}`.
The demo maps the coordinates back onto your original image and marks the
predicted click point with a crosshair.

The system prompt is the exact frozen per-platform grounding prompt the model was
trained with (`scalecua_toolcall`): pick **desktop**, **web**, or **mobile** to
serve the matching tool schema (`computer_use` / `browser_use` / `mobile_use`).
Inference is greedy (`do_sample=False`), matching the paper's evaluation.

- Paper: [GroundGUI: GUI Grounding by High-Quality Data and Distillation-to-Reinforcement Learning](https://huggingface.co/papers/2610.19100)
- Code & data: [github.com/PrentisAI/GroundGUI](https://github.com/PrentisAI/GroundGUI)
- Model: [PrentisAI/GroundGUI-8B](https://huggingface.co/PrentisAI/GroundGUI-8B)

**Benchmark results** (from the paper): 95.2 ScreenSpot-v2 · 63.9 ScreenSpot-Pro ·
86.8 MMBench-GUI L2 · 38.3 UI-Vision · 65.4 OSWorld-G.

## Example screenshots

The bundled example screenshots come from
[likaixin/ScreenSpot-Pro](https://huggingface.co/datasets/likaixin/ScreenSpot-Pro)
(MIT license), a professional high-resolution GUI grounding benchmark.