Thermodynamic Equilibrium of UI Layouts: A Non-Autoregressive Graph Diffusion Approach to Frontend Code Generation

LiHan, Rui Zhang, fen Huang, Yanan.cai, Bin Wang, Ya Xiong

Abstract: The automatic derivation of runnable applications from UI screenshots is a persistent challenge in engineering. Although MLLMs have advanced static web generation, their application to complex frameworks such as Vue, React, and Angular remains unstable. Standard prompting strategies often yield substantial performance deficits due to strict framework-specific syntax requirements. In response, we propose the \textbf{Constraint-Guided Orchestrator (CGO)}, an architecture that encapsulates multi-framework nuances into strict rule sets managed by cooperative agents. Instead of unconstrained generation, CGO employs a structured protocol: a perception module first extracts structured data tables from visual elements, which informs a strategic planner that outlines the hierarchical tree and file contracts. Actual coding is executed via strictly bounded write operations, while a real-time audit layer enforces logical consistency through pre-computation checks. This design allows for rapid error correction without expensive build cycles. Experiments demonstrate that CGO delivers superior multi-framework portability and drastically reduces syntax violations.