Vision2Web: A Hierarchical Benchmark for Visual Website Development with Agent Verification

Vision2Web is a hierarchical benchmark for visual website development spanning static UI, interactive frontend, and full-stack tasks.

The benchmark comprises 193 tasks from 16 categories with 918 prototype images and 1,255 test cases drawn from real websites.

The authors propose a workflow-based agent verification combining a GUI agent verifier and a VLM-based judge to evaluate coding agents.

Evaluations show significant performance gaps, with state-of-the-art models struggling on full-stack development.