Detecting UI components from pixels is often hampered by inconsistent DOM structures and noisy heuristic label extraction. AtomBlock-WebUI sidesteps those issues by generating semantic HTML with explicit yolo-* markers, rendering pages with Playwright, and extracting getBoundingClientRect coordinates to produce tightly aligned visual labels that are ready for object-detection training.
What Sets It Apart
- Pixel-perfect DOM extraction: labels come from rendered DOM queries instead of heuristic DOM parsing or human box-drawing, so bounding boxes align precisely with the visual output — this reduces label noise that typically hurts small, dense UI element detection.
- Real-image injection for realism: CC3M images are retrieved and injected into generated HTML via FAISS to narrow the visual domain gap between purely synthetic layouts and real webpages — this helps models generalize better to natural content while keeping synthetic control.
- Task-oriented splits and YOLO-ready layout: the release includes 9,683 images, 1.32M+ bboxes, and train/val/test splits plus a ready data.yaml for YOLO workflows, letting practitioners plug straight into training pipelines without heavy pre-processing.
Who It's For & Trade-offs
Great fit if you train or benchmark UI element detectors, conduct domain-adaptation experiments for web UI vision tasks, or need reproducible synthetic data with precise geometry. Look elsewhere if you need commercial-redistributable assets (dataset is CC BY-NC-SA 4.0), deeply nested/edge-case DOM complexity present in some production sites, or pixel-perfect photorealism — the HTML is LLM-generated and may lack certain real-world DOM clutter.
Where It Fits
AtomBlock-WebUI sits between handcrafted DOM-extracted datasets and wholly human-annotated web UI screenshots: it provides controlled layout variability and high-quality geometric labels while injecting real images to improve visual fidelity. Use it as a synthetic augmentation source, a pretraining corpus for small-element detectors, or as a benchmark for web UI detection models.
