ScreenParser

Model Summary

ScreenParser is a YOLO11-Large model fine-tuned at 1280px resolution on ScreenParse v2. The v2 detector uses the filtered leaf-element annotations released with the current dataset main branch.

Supported Classes (55)

Table, Column/Browser, Button, Utility Button, App Icon, Navigation Bar, Status Bar, Search Field, Toolbar, Tooltip, Video, Tab Bar, Side Bar, Slider, Picker, ContextMenu, DockMenu, EditMenu, Image, Scroll, Switch, File Icon, Chart, Window, Screen, List, List Item, PopUp Menu, Steppers, Toggles, Text Input, Rating Indicator, Checkbox, Radiobox, Select, Avatar, Badge, Alert, Progress bar, Bottom navigation, Breadcrumb, Page control, Link, Menu, Pagination, Tab, Search Bar, Date-Time picker, Calendar, Text, Heading, Code snippet, Carousel, Notification, Logo

Usage

Single Image Inference

from ultralytics import YOLO

model = YOLO("docling-project/ScreenParser")
results = model.predict("screenshot.png", imgsz=1280, conf=0.10, iou=0.10)

for r in results:
    for box, cls_id, conf in zip(r.boxes.xyxy, r.boxes.cls, r.boxes.conf):
        x1, y1, x2, y2 = box.tolist()
        label = model.names[int(cls_id)]
        print(f"{label:20s}  conf={conf:.2f}  bbox=({int(x1)}, {int(y1)}, {int(x2-x1)}, {int(y2-y1)})")

Batch Inference

import os
from ultralytics import YOLO

model = YOLO("docling-project/ScreenParser")
image_dir = "screenshots/"
images = sorted(
    os.path.join(image_dir, f)
    for f in os.listdir(image_dir)
    if f.lower().endswith((".png", ".jpg", ".jpeg"))
)

results = model.predict(images, imgsz=1280, conf=0.10, iou=0.10, batch=16)

Save Visualizations

from ultralytics import YOLO

model = YOLO("docling-project/ScreenParser")
results = model.predict("screenshot.png", imgsz=1280, conf=0.10, iou=0.10, save=True)
# Annotated image saved under runs/detect/predict/

Training Data

The current main checkpoint was trained on ScreenParse v2, which provides 1,447,100 high-quality training screenshots and 25,575,213 UI element annotations. The dataset uses filtered leaf-element annotations to reduce noisy nested boxes and includes multiple viewport resolutions.

The original ScreenParser checkpoint trained on ScreenParse v1 remains available with revision="v1".

Limitations

• Produces bounding boxes and element labels only; it does not produce text content for detected elements. Pair it with OCR or ScreenVLM when text extraction is needed.
• The model is trained on rendered web screenshots, so performance may vary on native desktop, mobile, or application screenshots outside the training distribution.

Citation

@misc{gurbuz2026movingsparsegroundingcomplete,
      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},
}