Community Builds Open-Source PaperBanana to Auto-Generate ML Paper Diagrams
A community team shipped an MIT-licensed reimplementation of Google's PaperBanana paper, wrapping the five-agent diagram pipeline in a CLI, Python API, Gradio studio, and MCP server.
Takeaways
- Community team llmsresearch shipped an MIT-licensed open-source implementation of Google's PaperBanana paper.
- Two-phase pipeline: Retriever, Planner, Stylist plan the figure; Visualizer and Critic refine it iteratively.
- Supports OpenAI, Azure, Google Gemini, Atlas Cloud, and OpenRouter as swappable providers.
- Ships a CLI, Python API, Gradio Studio, MCP server, and Overleaf GitHub Action.
- Includes batch generation, full-paper orchestration, composite figures, and VLM-as-a-Judge evaluation.
- Original paper is arXiv:2601.23265 by Zhu et al. from Peking University and Google Cloud AI Research.
Turning a methodology section into a clean architecture diagram is still one of the most manual parts of writing an ML paper. A community group called llmsresearch has released an open source reimplementation of Google's PaperBanana paper that tries to close that gap, and it has already crossed 2,000 stars on GitHub.
The project is explicitly a fan port. It is an unofficial, community-driven open-source implementation of the paper "PaperBanana: Automating Academic Illustration for AI Scientists" by Dawei Zhu, Rui Meng, Yale Song, Xiyu Wei, Sujian Li, Tomas Pfister, and Jinsung Yoon (arXiv:2601.23265), not affiliated with or endorsed by the original authors or Google Research, and the implementation may differ from the original system. The official codebase is now branded PaperVizAgent at google-research/papervizagent; the paper name PaperBanana remains the common label.
What the paper actually proposed
The PaperBanana framework, from Peking University and Google Cloud AI Research and based on Google's Nano Banana, uses five specialized AI agents to generate publication-ready diagrams from method descriptions. Pure image generation models look good but rarely meet academic publication standards, so PaperBanana divides tasks among specialized agents: the first searches a reference database for similar diagrams to use as templates, the second translates the paper's method description into a detailed image description, and the third refines it using aesthetics guidelines extracted from NeurIPS publications.