Vision2Web: A Hierarchical Benchmark for Visual Website Development Based on Agentic Verification

Vision2Web is a hierarchical benchmark designed to evaluate how well multimodal coding agents perform the entire process of visual website development. It comprehensively measures UI understanding, requirement reasoning, interactive logic, and full-stack implementation capability.

The benchmark consists of three levels according to difficulty:

• Level 1 (Static Webpage): Generating responsive webpages based on UI prototypes for various devices (Visual Score measured)
• Level 2 (Interactive Frontend): Developing multi-page interactive frontends using text specifications and prototypes (Visual Score + Functional Score measured)
• Level 3 (Full-Stack Website): Building an entire system including state management and backend logic based on requirement documents and prototypes (Visual Score + Functional Score measured)

The evaluation uses a workflow-based approach that combines functional correctness verification through a GUI agent verifier with visual fidelity assessment through a VLM-based judge. This enables scalable evaluation regardless of implementation method.

The dataset includes 193 tasks across 4 major domains: e-commerce, SaaS, content, and public services. For this, 918 prototype images and 1,256 functional test cases are provided.