UI-Redline-bench

This dataset contains the benchmark data for the paper "UI-Redline-bench: 赤入れ指示によるWebUIコード修正ベンチマーク".

The benchmark evaluates the capability of Vision-Language Models (VLMs) to modify Web UI code (HTML/CSS) based on visual "redline" instructions (handwritten or digital) drawn on screenshots.

📄 Paper | 💻 GitHub Repository (Evaluation Code & Runnable Environment)

Usage

This Hugging Face dataset contains only the images and the corresponding metadata. To run the experiments, please follow the steps below to clone the GitHub repository for evaluation scripts and place the dataset accordingly. You can optionally clone this Hugging Face repository if you want to inspect the images manually.

mkdir ui-redline-workspace
cd ui-redline-workspace

# 1. Clone the GitHub repository (REQUIRED for running code)
git clone https://github.com/future-architect/UI-Redline-bench.git

# 2. (Optional) Clone this Hugging Face dataset
# Only needed if you want to browse instruction images manually on your local machine.
# The python script will download the dataset automatically via API.

# Initialize Git LFS (Required to download large image/parquet files)
git lfs install

# Clone into a specific directory to avoid name conflict
git clone https://huggingface.co/datasets/future-architect/UI-Redline-bench UI-Redline-bench-dataset

The resulting directory structure should look like this:

ui-redline-workspace/
├── UI-Redline-bench/           # GitHub Repo: Scripts, HTML, CSS (The execution environment)
│   ├── data/
│   ├── script/
│   └── ...
└── UI-Redline-bench-dataset/   # HF Repo: (Optional) For manual image inspection
    └── ...

Dataset Structure

Each record represents a modification task. The file paths provided (src_html_path, etc.) are relative to the root of the GitHub repository (UI-Redline-bench/).

Field	Type	Description
src_html_path	string	Relative path to the original HTML code (e.g., data/news/bootstrap/src/index.html).
src_css_path	string	Relative path to the original CSS code.
web_type	string	Type of the website (news, onlinestore, portfolio).
css_framework	string	CSS framework used (vanilla, bootstrap, tailwind).
image_instruct	image	The visual instruction (redline) image input for the VLM.
modification_category	string	Category of the modification (e.g., layout, color_contrast).
style	string	Style of the visual instruction (digital or handwritten).
image_has_arrow	bool	Whether the instruction image contains arrows.
image_has_enclosure	bool	Whether the instruction image contains enclosures/bounding boxes.
image_has_ui_sketch	bool	Whether the instruction image contains sketches of new UI elements.
ref_html_path	string	Relative path to the ground truth HTML code.
ref_css_path	string	Relative path to the ground truth CSS code.

Usage Example (Running Inference)

This example demonstrates how to load the dataset and run inference by importing the scripts directly from the cloned GitHub repository. Save the following code as run_benchmark.py in your ui-redline-workspace directory.

import os
import sys
from datasets import load_dataset

# 1. Setup Paths
# Assuming you are in the 'ui-redline-workspace' directory.
GITHUB_REPO_ROOT = os.path.abspath("./UI-Redline-bench") 
HF_DATASET_PATH = os.path.abspath("./UI-Redline-bench-dataset")
OUTPUT_DIR = os.path.abspath("./output_results")

# 2. Add GitHub script directory to sys.path to allow imports
sys.path.append(os.path.join(GITHUB_REPO_ROOT, "script"))

try:
    # -------------------------------------------------------------------------
    # IMPORT THE TARGET MODEL SCRIPT HERE
    # Change this line depending on the model you want to evaluate:
    #   from prediction_based_on_image_gpt5 import process_sample
    #   from prediction_based_on_image_claude import process_sample
    #   from prediction_based_on_image_gemini import process_sample
    #   from prediction_based_on_image_qwen import process_sample
    # -------------------------------------------------------------------------
    from prediction_based_on_image_gemini import process_sample
except ImportError as e:
    print("Error importing scripts. Make sure you are running this script with the correct environment (e.g., via 'uv run').")
    raise e

# 3. Load Dataset
if os.path.exists(HF_DATASET_PATH):
    print(f"Loading dataset locally from: {HF_DATASET_PATH}")
    ds = load_dataset(HF_DATASET_PATH, split="test")
else:
    print("Local dataset not found. Downloading from Hugging Face Hub...")
    ds = load_dataset("future-architect/UI-Redline-bench", split="test")

# 4. Iterate and Run Inference
for example in ds:
    # The dataset returns a PIL.Image object, which can be passed directly to the scripts.
    img_input = example['image_instruct']

    # Construct absolute paths for HTML/CSS
    html_path = os.path.join(GITHUB_REPO_ROOT, example['src_html_path'])
    css_path = os.path.join(GITHUB_REPO_ROOT, example['src_css_path'])
    
    # Construct output directory for this case
    case_output_dir = os.path.join(OUTPUT_DIR, os.path.dirname(example['ref_html_path']))
    
    print(f"Processing: {html_path}")
    
    # Call the imported function directly
    process_sample(
        html_path=html_path,
        css_path=css_path,
        image_path=img_input,
        output_dir=case_output_dir
    )

print("Inference completed.")

How to execute the script

Since we use uv for dependency management, you must run the script using the correct environment defined in the GitHub repository.

For GPT, Claude, and Gemini (API-based models): Use the cpu-env.

uv run --project UI-Redline-bench/cpu-env python run_benchmark.py

For Qwen (Local vLLM model): Use the gpu-env. Make sure you have started the vLLM server beforehand.

# 1. Start the server (in a separate terminal)
uv run --project UI-Redline-bench/gpu-env bash UI-Redline-bench/script/launch_vllm_server.sh

# 2. Run the benchmark
uv run --project UI-Redline-bench/gpu-env python run_benchmark.py

Citation

If you use this dataset, please cite our paper:

@inproceedings{hiai2026uiredline,
  title={UI-Redline-bench: 赤入れ指示によるWebUIコード修正ベンチマーク},
  author={肥合智史 and 藤井諒 and 岸波洋介 and 森下睦},
  booktitle={Proceedings of the 32nd Annual Meeting of the Association for Natural Language Processing (NLP2026)},
  year={2026}
}