New benchmark tackles design-to-code translation for e-commerce UI workflows

Converting visual designs into functional code remains a challenging frontier for AI systems. A new research benchmark aims to establish clearer standards for measuring progress in this space by providing a consistent dataset, evaluation framework, and toolchain for the design-to-code task.

The benchmark framework#

Researchers have unveiled 1D-Bench, a testing suite grounded in actual e-commerce environments where design conversion happens under real-world constraints. The name reflects the goal of completing such tasks efficiently—within a single day. Each benchmark instance includes both a reference visual rendering and an exported intermediate representation of the design structure, which may contain imperfections from the extraction process.

The system requires models to generate executable React code with an explicit component hierarchy, all within a fixed technical environment. This setup tests whether AI can work with imperfect structural data while still producing visually accurate outputs.

Iterative refinement improves results#

Rather than expecting perfect code generation on the first try, 1D-Bench adopts a multi-round approach where models iteratively refine their output. During each cycle, the system applies component-level modifications informed by execution feedback—essentially allowing the AI to see what went wrong and adjust accordingly.

Testing on both commercial and open-source multimodal models revealed that this iterative strategy generally enhances final performance:

• Rendering success rates increase across iterations
• Visual similarity to target designs often improves with refinement
• The approach demonstrates robustness to structural data defects

Emerging challenges with learning strategies#

The research team also explored advanced training techniques, including synthetic repair trajectories and reinforcement learning approaches for guided editing. However, these methods produced only modest and inconsistent gains. The researchers suggest this limitation may stem from sparse reward signals at task completion and the inherent unpredictability of file-level modifications—areas that warrant further investigation.

The benchmark and evaluation scripts are available in an anonymized repository for the research community.