UniCoder

This repository hosts public resources for UniCoder: Unified Visual-to-Code Generation via Symbolic Rewards and Reference-Guided Code Optimization.

Authors

Yaozhi Zheng*, Yilei Jiang*, Manyuan Zhang, Yuxuan Wan, Kaituo Feng, Tianshuo Peng, Bo Zhang, Xiangyu Yue†

(* Equal contribution, † Corresponding author)

The Chinese University of Hong Kong  |  Shanghai Artificial Intelligence Laboratory

Overview

The current release focuses on evaluation scripts for visual-to-code benchmarks. UniCoder model weights, training data processing, and full UniCoder training scripts are planned for a later release.

Release Status

• Evaluation pipeline: available for ChartMimic and UniSVG.
• Model weights: not included in this release.
• UniCoder training scripts: not part of the current release scope.
• Benchmark data: not redistributed here; please download it from the official benchmark sources.

Repository Structure

• eval/: Evaluation release for ChartMimic and UniSVG.
  • simple_eval/: Unified evaluation entry point and model wrappers.
    • run_pipeline.py: Runs ChartMimic and UniSVG evaluation pipelines.
    • models.py: Local/API model wrappers for Qwen, OpenAI, Gemini, Claude, and DashScope Qwen APIs.
    • requirements.txt: Python dependencies for the evaluation wrapper.
• eval/ChartMimic/: Modified ChartMimic benchmark resources and evaluators.
• eval/unisvg.github.io/: Modified UniSVG benchmark resources and evaluation scripts.
• train/: Reserved for future UniCoder training-code release.

Benchmark-specific README files remain inside eval/ because they document the adapted benchmark components. The root README provides the project-level release instructions.

Requirements

• Python 3.9+
• Linux or macOS for API-based evaluation. The UniSVG evaluation uses signal.SIGALRM, so Windows is not currently supported.
• CUDA-capable GPUs are recommended for local Qwen model evaluation.

Installation

pip install -r eval/simple_eval/requirements.txt

For macOS users, SVG rendering may additionally require Cairo and Pango:

brew install cairo pango

API Keys

API-based models read credentials from environment variables. Do not hard-code keys in source files.

Set only the credentials required by the model APIs you plan to use:

export DASHSCOPE_API_KEY=your_dashscope_key
export OPENAI_API_KEY=your_openai_key
export OPENAI_BASE_URL=your_openai_base_url_optional
export GOOGLE_API_KEY=your_google_key
export ANTHROPIC_API_KEY=your_anthropic_key

Data Preparation

Download benchmark data from the official sources and place it under the expected paths:

• ChartMimic data: eval/ChartMimic/dataset/
• UniSVG data: eval/unisvg.github.io/dataset/

Please follow the original benchmark licenses and data-use requirements.

Quick Start

Run commands from the repository root.

ChartMimic

python eval/simple_eval/run_pipeline.py \
  --task chartmimic \
  --model qwen_api \
  --model_path qwen-vl-max \
  --judge_model qwen_api \
  --judge_path qwen-vl-plus

UniSVG

python eval/simple_eval/run_pipeline.py \
  --task unisvg \
  --model qwen_api \
  --model_path qwen-vl-max \
  --data_file eval/unisvg.github.io/dataset/test.json \
  --image_dir eval/unisvg.github.io/dataset/png

Use --num for a small smoke test:

python eval/simple_eval/run_pipeline.py \
  --task unisvg \
  --model qwen_api \
  --model_path qwen-vl-max \
  --data_file eval/unisvg.github.io/dataset/test.json \
  --image_dir eval/unisvg.github.io/dataset/png \
  --num 5

Citation

@article{zheng2026unicoder,
  title={Unified Visual-to-Code Generation via Symbolic Rewards and Reference-Guided Code Optimization},
  author={Zheng, Yaozhi and Jiang, Yilei and Zhang, Manyuan and Wan, Yuxuan and Feng, Kaituo and Peng, Tianshuo and Zhang, Bo and Yue, Xiangyu},
  journal={arXiv preprint},
  year={2026},
  note={Yaozhi Zheng and Yilei Jiang contributed equally; Xiangyu Yue is the corresponding author.}
}

Acknowledgements

This evaluation release builds on ChartMimic and UniSVG. Please also cite and follow the licenses of the original benchmark repositories when using their code or data.