WebUI95 : UI-to-Code Generation and Analysis

This repository contains the code and data for the Web++ project, which evaluates UI-to-code generation models.

Repository Structure

web++/
├── WebUI95/                    # UI-to-Code Experiment Pipeline
│   ├── output/                 # Final generated HTML (63,824 samples)
│   │   └── final_html/         # Generated static HTML files
│   ├── final_results_64k.zip   # Compressed final results
│   ├── slurm/                  # SLURM job scripts
│   └── *.py                    # Processing scripts
│
├── data_analysis/              # Quality and Diversity Analysis
│   ├── analysis/               # Analysis results
│   │   ├── *.json              # Embedding and metric files
│   │   └── plots/              # Generated figures
│   ├── webgen_bench/           # WebGen-Bench baseline
│   │   └── fixed_samples.zip   # Fixed prompt experiments (475 samples)
│   ├── scripts/                # Analysis scripts
│   └── slurm/                  # SLURM job scripts
│
└── venv/                       # Python virtual environment

Datasets

WebUI95 Dataset

• Location: WebUI95/output/final_html/ or WebUI95/final_results_64k.zip
• Size: 63,824 UI samples
• Format: Each sample contains generated static HTML/CSS

WebGen-Bench Fixed Prompt Samples

• Location: data_analysis/webgen_bench/fixed_samples.zip
• Size: 475 samples
• Purpose: Control experiment for diversity analysis (same prompt, varying seeds)

Key Scripts

WebUI95 Pipeline

Script	Description
run_inference_multigpu.py	Multi-GPU inference for UI-to-code generation
process_unified_data.py	Data preprocessing
merge_outputs.py	Merge outputs from multiple workers

Data Analysis

Script	Description
scripts/uiclip_analysis.py	UIClip quality and similarity scoring
scripts/compute_code_embeddings.py	Qwen code embeddings for diversity
scripts/generate_figures.py	Generate analysis figures
scripts/html_statistics.py	DOM complexity analysis

Results

Visual Quality (UIClip)

Metric	Original	Generated
Mean Quality Score	0.487	0.518
Mean Similarity	-	0.938

Diversity Analysis

Type	Original	Generated	Synthetic (Fixed)
Visual	0.150	0.137	0.140
Code	0.119	0.049	0.009

Key Finding: The synthetic fixed-prompt experiment shows near-zero code diversity (0.009), confirming mode collapse when using identical prompts. Our generated approach maintains significantly better diversity.

Figures

Generated figures are in data_analysis/analysis/plots/:

• figure2_html_statistics.png - HTML structure comparison
• figure3_combined_diversity_comparative.png - Diversity analysis

Requirements

• Python 3.10+
• PyTorch 2.0+
• Transformers
• Playwright (for screenshot generation)
• UIClip model: biglab/uiclip_jitteredwebsites-2-224-paraphrased_webpairs_humanpairs
• Qwen models: Qwen/Qwen3-Coder-30B-A3B-Instruct, Qwen/Qwen2.5-Coder-1.5B

Compute Environment

• GPU: NVIDIA A40 (48GB VRAM)
• Partitions: gpu_a40, cpu_sapphire

Citation

If you use this code or data, please cite:

@inproceedings{webuininetyfive2026,
  title={WebUI-95: A Large-Scale Dataset of Normalized Web Interfaces via UI-to-Code Generation},
  author={...},
  booktitle={CHI Posyets},
  year={2026}
}

Acknowledgments

Computational resources provided by institutional HPC cluster.