WebSight: 1-10 Million Synthetic Website Screenshots and Code Pairs

WebSight contains between 1 and 10 million pairs of synthetic website screenshots and their corresponding HTML/CSS code, released by HuggingFaceM4 in March 2024. The collection features two distinct versions covering standard HTML/CSS and modern HTML/Tailwind CSS implementations for English-language websites.

Use Cases

• Training vision-language models to generate HTML/CSS code from a screenshot
• Fine-tuning models for UI-to-code translation using the Tailwind CSS version
• Benchmarking image-to-text models on structured code generation tasks

Strengths

• Scale of 1 to 10 million records
• CC-BY-4.0 license allows for broad research and commercial use
• Tailwind CSS support in v0.2 for modern UI training

Limitations

• Synthetic nature may lack the complexity and edge cases of human-authored production code
• Restricted to English language content
• v0.1 lacks real image assets in the rendered screenshots

Provenance

Source

HuggingFaceM4 (Arxiv 2403.09029)

Collection Method

synthetic

Freshness

Last updated March 2024; static synthetic generation.