---

TL;DR: FocusUI teaches VLMs where to look in UI screenshots. 🔍

Mingyu Ouyang¹, Kevin Qinghong Lin², Mike Zheng Shou^1†, Hwee Tou Ng^1†
1National University of Singapore    ²University of Oxford
^†Corresponding authors

Overview ✨

Vision-Language Models (VLMs) have shown remarkable performance in UI grounding tasks, but high-resolution screenshots are tokenized into thousands of visual tokens (e.g., ~4700 for 2K resolution), causing significant computational overhead. In contrast, humans naturally focus on regions of interest when interacting with UI. FocusUI is an efficient UI grounding framework that selects patches most relevant to the instruction while preserving positional continuity for precise grounding.

Key Innovations

1. Query-Guided Visual Token Selection: Constructs patch-level supervision by fusing instruction-conditioned scores with rule-based UI-graph scores that down-weight large homogeneous regions.
2. POSPAD (Position-Preserving Padding): A novel strategy that compresses each contiguous sequence of dropped visual tokens into a single special marker placed at the sequence's last index, preserving positional continuity crucial for UI grounding.

Updates 📣

• [2026/02/08] 🤗 Models, dataset and benchmarks are available on HuggingFace.
• [2025/12/29] Project page and code base released.

Quick Start 🚀

Installation

# Clone the repository
git clone https://github.com/showlab/FocusUI.git
cd FocusUI

# Install dependencies
conda create -n focusui python=3.12 -y
conda activate focusui
pip install -r requirements.txt

To download checkpoints:

# download FocusUI-3B
hf download yyyang/focusui_3b_ft_final --repo-type model --local-dir ./checkpoints/focusui-3b

(Optional) To download benchmarks and training datasets:

# download benchmarks
hf download yyyang/UI-Grounding-Benchmarks --repo-type dataset --local-dir ./datasets/UI-Grounding-Benchmarks/

# download training datasets
hf download yyyang/FocusUI-Training-Data --repo-type dataset --local-dir ./datasets/FocusUI-Training-Data/

Quick Start: Inference with FocusUI

See inference_focusui.py for an example of how to use FocusUI for inference.

Training 🧠

FocusUI uses a two-stage training process:

Stage 1: Train Patch Scorer Only. This stage trains only the PatchScorer module while freezing the base VLM.

bash scripts/train/stage_1_ft_focusui_scorer.sh

Stage 2: Full Fine-tuning. This stage fine-tunes the entire model with the trained PatchScorer.

bash scripts/train/stage_2_ft_focusui.sh

Evaluation 📊

Run evaluation on grounding benchmarks:

# ScreenSpot-Pro
python -m evaluation.ss_pro_eval \
    --model_type focusui_3b \
    --model_name_or_path checkpoints/FocusUI-3B \
    --data_path ./datasets/UI-Grounding-Benchmarks/ScreenSpot-Pro \
    --save_path ./results/ss_pro \
    --visual_reduct_ratio 0.5

# ScreenSpot-V2
python -m evaluation.ss_v2_eval \
    --model_type focusui_3b \
    --model_name_or_path checkpoints/FocusUI-3B \
    --data_path ./datasets/UI-Grounding-Benchmarks/ScreenSpot-V2 \
    --save_path ./results/ss_v2 \
    --visual_reduct_ratio 0.5

# ScreenSpot-V2
python -m evaluation.ss_v2_eval \
    --model_type focusui_qwen3vl_2b \
    --model_name_or_path checkpoints/FocusUI-Qwen3-VL-2B \
    --data_path ./datasets/UI-Grounding-Benchmarks/ScreenSpot-V2 \
    --save_path ./results/ss_v2_2b \ 
    --visual_reduct_ratio 0.5

# UI-Vision
python -m evaluation.ui_vision_eval \
    --model_type focusui_3b \
    --model_name_or_path checkpoints/FocusUI-3B \
    --data_path ./datasets/UI-Grounding-Benchmarks/UI-Vision \
    --save_path ./results/ui_vision \
    --visual_reduct_ratio 0.5


# OSWorld-G
python -m evaluation.os_world_g_eval \
    --model_type focusui_3b \
    --model_name_or_path checkpoints/FocusUI-3B \
    --data_path ./datasets/UI-Grounding-Benchmarks/OSWorld-G \
    --save_path ./results/osworld_g \
    --visual_reduct_ratio 0.5

Key Evaluation Options

Argument	Description	Default
--apply_visual_token_select	Enable visual token selection	True
--visual_reduct_ratio	Token retention ratio (1.0 = keep all)	0.5

Model Zoo 🧩

Model	Backbone	Parameters	HuggingFace
FocusUI-3B	Qwen2.5-VL-3B	3B	FocusUI-3B
FocusUI-7B	Qwen2.5-VL-7B	7B	FocusUI-7B
FocusUI-2B	Qwen3-VL-2B	2B	FocusUI-Qwen3-VL-2B

Citation 📝

If you find FocusUI useful for your research, please cite:

@article{ouyang2025focusui,
  title   = {FocusUI: Efficient UI Grounding via Position-Preserving Visual Token Selection},
  author  = {Ouyang, Mingyu and Lin, Kevin Qinghong and Shou, Mike Zheng and Ng, Hwee Tou},
  year    = {2025},
  journal = {arXiv preprint},
}

Acknowledgements 🙏

FocusUI builds upon Qwen2.5/3-VL and GUI-Actor as backbone models. We thank the open-source community for their valuable contributions.