WebUI-95 Turns Screenshots Into Lean Code

WebUI-95 makes a strong case for AI-generated code as a data-cleaning layer, but its impressive visual metrics should not be mistaken for production-ready frontend reconstruction.

• –The dataset converts messy real-world HTML into compact semantic structures while retaining substantial visual diversity.
• –Cleaner DOMs could improve screenshot-to-code models, layout prediction, accessibility tooling, and web-agent perception.
• –The generated pages are static, so buttons, forms, navigation, and other interactive behaviors are lost.
• –Fidelity drops on overlapping layouts, embedded media, and custom-font-heavy pages, limiting use in pixel-perfect workflows.
• –Its biggest value may be normalization: preserving real design variety without forcing models to learn framework boilerplate.