Lens: Rethinking Training Efficiency for Foundational Text-to-Image Models Paper โข 2605.21573 โข Published May 20 โข 111