ControlArt-Bench

ControlArt-Bench is a benchmark for evaluating intent-driven image retouching and visual focus enhancement, introduced with EyeControl.

Dataset Contents

The dataset contains 200 evaluation samples.

Each sample includes:

• a input image.
• a user intent masks.
• a retouching instructions (we support a long instruction and a short instruction).
• a reference image.

Directory Structure

Replace this example with the actual dataset layout:

.
├── inputs/
├── masks/
├── prompts/
├── targets/
└── README.md

Download

pip install -U huggingface_hub

hf download Dragoniss/ControlArt-Bench \
  --repo-type dataset \
  --local-dir datasets/ControlArt-Bench

Evaluation

See evaluation details in https://github.com/DragonisCV/EyeControl/blob/main/docs/evaluation.md

License

Citation

@article{qin2026dotting,
  title={Dotting the Eye: An Intent-Driven Image Retouching Agent for Visual Focus Enhancement},
  author={Qin, Chujie and Zhang, Zilong and Chang, Zewei and Guo, Chunle and Wang, Ruixing and Hu, Tao and Cheng, Ming-Ming and Li, Chongyi},
  journal={arXiv preprint arXiv:2609.01148},
  year={2026},
  url={https://arxiv.org/abs/2609.01148}
}