Seven agents and a critic walk into a methodology diagram

What it does PaperBanana is a community-driven reimplementation of a Google Research framework that automates the creation of academic diagrams and statistical plots. You feed it a text description of your method, a PDF excerpt, or a CSV dataset, and it runs a multi-agent pipeline to produce a refined image. The system handles everything from methodology overviews to bar charts, and can batch-process entire figure manifests into composite panels.

The interesting bit The pipeline splits work across up to seven narrowly scoped agents: an input optimizer sharpens your vague caption, a retriever picks relevant examples from a curated set of thirteen NeurIPS diagrams, and a planner-stylist pair drafts the visual spec. A visualizer renders it, then a critic loops back with corrections for up to three rounds—or until satisfied. It is essentially a tiny design studio where the art director and the renderer are both language models.

Key highlights

• Supports multiple backends: OpenAI, Azure OpenAI/Foundry, Google Gemini (free tier), and OpenRouter.
• Accepts text, PDF page ranges, and CSV/JSON data as inputs.
• Offers a local Gradio web UI, an MCP server for IDE integration, and Claude Code skills.
• Batch mode can generate multiple diagrams and stitch them into composite labeled figures automatically.
• Includes an optional input-optimization phase that parallelizes context enrichment and caption sharpening.

Caveats

• The authors explicitly disclaim this as an unofficial implementation that may diverge from the original Google Research system.
• The tool is built entirely around commercial cloud vision and image-generation APIs; no local or offline generation mode is mentioned.

Verdict Researchers who dread drawing architecture diagrams or formatting plots will find this a genuine time-saver. If you need pixel-perfect control or are allergic to cloud API costs, stick with TikZ and Matplotlib.