JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence
JanusCode-800K is a large-scale multimodal code corpus designed to establish a visual-programmatic interface for code intelligence. It contains 800,000 samples spanning heterogeneous domains including standard charts, complex interactive web UIs, SVG graphics, and code-driven animations (e.g., Manim, Wolfram).
The dataset powers the training of the JanusCoder and JanusCoderV models, enabling them to handle both text-centric and vision-centric coding tasks—from chart-to-code generation to precise, program-driven editing of visualizations.
The corpus includes the following specialized splits:
web_generate: Web UI generation from visual/textual specifications.web_edit: Code-driven editing of web interfaces.svg: Visual programming for vector graphics.manim: Mathematical animations and dynamic theorem visualizations.python_chart: Programmatic data-to-chart generation.wolfram: Symbolic and visual computation code.demo: Web artifacts and demonstration.🫶 If you are interested in our work or find the repository / checkpoints / benchmark / data helpful, please consider using the following citation format when referencing our papers:
@inproceedings{sun2026januscoder,
title={JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence},
author={Qiushi Sun and Jingyang Gong and Yang Liu and Qiaosheng Chen and Lei Li and Kai Chen and Qipeng Guo and Ben Kao and Fei Yuan},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026}
}
JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence
JanusCode-800K is a large-scale multimodal code corpus designed to establish a visual-programmatic interface for code intelligence. It contains 800,000 samples spanning heterogeneous domains including standard charts, complex interactive web UIs, SVG graphics, and code-driven animations (e.g., Manim, Wolfram).
The dataset powers the training of the JanusCoder and JanusCoderV models, enabling them to handle both text-centric and vision-centric coding tasks—from chart-to-code generation to precise, program-driven editing of visualizations.
The corpus includes the following specialized splits:
web_generate: Web UI generation from visual/textual specifications.web_edit: Code-driven editing of web interfaces.svg: Visual programming for vector graphics.manim: Mathematical animations and dynamic theorem visualizations.python_chart: Programmatic data-to-chart generation.wolfram: Symbolic and visual computation code.demo: Web artifacts and demonstration.🫶 If you are interested in our work or find the repository / checkpoints / benchmark / data helpful, please consider using the following citation format when referencing our papers:
@inproceedings{sun2026januscoder,
title={JanusCoder: Towards a Foundational Visual-Programmatic Interface for Code Intelligence},
author={Qiushi Sun and Jingyang Gong and Yang Liu and Qiaosheng Chen and Lei Li and Kai Chen and Qipeng Guo and Ben Kao and Fei Yuan},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026}
}