🏆 ICLR 2026

Visual Self-Refine: A Pixel-Guided Paradigm
for Accurate Chart Parsing

Jinsong Li^1,2 | Xiaoyi Dong^1,2 | Yuhang Zang² | Yuhang Cao² | Jiaqi Wang² | Dahua Lin^1,2

¹ The Chinese University of Hong Kong
² Shanghai AI Laboratory

---

Visual Self-Refine (VSR) turns chart parsing into a pixel-guided feedback loop: predict anchors, render them, inspect the rendered chart, and refine the final parsing from the verified pixels.

💻 Overview

🎈 Quick Start

From the repository root:

pip install -r requirements.txt

Chart generation

cd chart_generation
python setup_fonts.py
bash gen.sh

Data pipeline

cd data_pipeline
python resize_and_remap.py
python render_anchors.py

Benchmark evaluation

cd benchmark
python eval_api.py \
  --input-json <path/to/chartp_annotations.json> \
  --image-folder <path/to/chartp_images> \
  --output-json outputs/chartp_gpt-4o.json

📎 Citation

@article{li2026visual,
  title={Visual self-refine: A pixel-guided paradigm for accurate chart parsing},
  author={Li, Jinsong and Dong, Xiaoyi and Zang, Yuhang and Cao, Yuhang and Wang, Jiaqi and Lin, Dahua},
  journal={arXiv preprint arXiv:2602.16455},
  year={2026}
}