RefCraft

Benchmarking Visual-Reference-Conditioned Agentic Generation
through Web Development

Paper · Project website · Code & evaluation · Download inputs

RefCraft provides 300 webpage-development tasks in English and Chinese. Each task combines a user request with screenshots from one to five reference webpages, together with evaluation criteria for the generated webpage.

Download

File	Contents
refcraft-inputs.tar.gz	Requests and all 3,491 reference screenshots
refcraft-evaluation.tar.gz	Evaluation criteria, core tests and supporting metadata

From the code repository root, download and extract the data:

pip install -U huggingface_hub
hf download YuchiWang/RefCraft --repo-type dataset --include 'bundles/*' --local-dir .
tar -xzf bundles/refcraft-inputs.tar.gz
tar -xzf bundles/refcraft-evaluation.tar.gz

The files are restored under data/. The code repository already includes the annotations, so only the input archive is needed with that checkout.

Browse the data

Select tasks to browse requests and reference-image previews, or keypoints to browse evaluation criteria. Both use the test split. Previews show one frame; download the input archive for the complete ordered reference images.

To load the tables, install datasets and run:

from datasets import load_dataset

tasks = load_dataset("YuchiWang/RefCraft", "tasks", split="test")
keypoints = load_dataset("YuchiWang/RefCraft", "keypoints", split="test")
print(tasks[0]["query"])

Task records contain sample_id, query, language, and ordered reference_images paths relative to data/. Evaluation criteria are for the judge; keep them out of the tested agent's workspace.

Run evaluation

In the RefCraft code repository, install the dependencies and place your agent's webpages in responses/ as <task_id>.html or <task_id>/index.html (IDs 1–300):

pip install -r requirements.txt
python -m playwright install chromium
export REFCRAFT_JUDGE_BASE_URL='https://YOUR-PROVIDER/v1'
export REFCRAFT_JUDGE_API_KEY='YOUR_API_KEY'
export REFCRAFT_JUDGE_MODEL='YOUR_VISION_MODEL'
python scripts/evaluate.py --responses responses --output runs/my-model \
  --model-name YOUR_MODEL_ID

Use Python 3.12+ and a vision-capable judge. Results are written to scores.csv and summary.json under the output directory. See the Quickstart for generating webpages.

Citation

@misc{wang2026refcraft,
  title = {RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development},
  author = {Yuchi Wang and Rongyao Fang and Weihao Yu and Xilin Wei and Tiancheng Gu and Dingkang Yang and Hang Zhang and Dayiheng Liu and Hongsheng Li and Shuai Bai},
  year = {2026},
  eprint = {2610.26826},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.26826}
}

License

Original annotation licensing is pending. See LICENSE.md and reference attribution for the recorded terms.