DeepMind publishes 'LayoutFormer' checkpoints for hierarchical UI parsing

AI · 6 min read

LayoutFormer uses a transformer architecture that encodes positional and semantic signals to output structured representations of UI screenshots.

Researchers and tool builders can use the checkpoints to extract component trees, approximate CSS, and accessibility labels as a starting point for design automation.

DeepMind shared evaluation scripts and baseline metrics, encouraging the community to build more robust reverse-engineering and prototyping tools on top of the model.