The layout lives in XML. What users see lives in pixels.
Gcat builds two commits and explores each one, saving a screenshot and GUI metadata at every step.
Redundant screens are filtered out. The rest are paired by bipartite matching on color and layout.
Perceptual image differencing finds what changed. Gcat classifies it, then says what and where.
Average recall: share of real GUI changes each one found.
“Seeing how components moved with the highlighted red boxes.”
— Study participant