Researchers have introduced UI2Code^N, a novel approach to generating front-end code from UI screenshots by treating it as an interactive visual optimization problem. This method embeds code generation within a feedback loop of execution, visual inspection, and iterative refinement, moving beyond single-pass generation. To handle non-differentiable visual objectives, they developed Relative Visual Policy Optimization (RVPO), a reinforcement learning technique that optimizes relative rankings of rendered candidates. The open-source 9B model trained with this paradigm achieves state-of-the-art results on UI drafting, polishing, and editing tasks, with performance improving through the iterative process. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Introduces a new iterative approach to UI-to-code generation that improves performance through visual feedback loops.
RANK_REASON This is a research paper detailing a new method and model for UI-to-code generation. [lever_c_demoted from research: ic=1 ai=1.0]