PaperBanana

PaperBanana generates academic methodology diagrams and statistical plots from text descriptions and data. It is an unofficial community implementation of the paper "PaperBanana: Automating Academic Illustration for AI Scientists", not affiliated with the original authors or Google Research, and may differ from the system the paper describes.

Generation runs as a multi-agent pipeline of up to seven agents. An optional input optimizer structures raw methodology text and turns a vague caption into a precise visual specification. A retriever then picks reference figures from a curated set of 13 methodology diagrams, a planner writes a textual description of the target figure from those examples, and a stylist adjusts it to venue style guidelines. A visualizer renders the image and a critic compares it with the source context and revises the description, repeating for a fixed number of rounds or until the critic is satisfied. Statistical plots are rendered from model-written matplotlib code, so they need no image-generation provider.

Beyond single figures, it can plan and generate a figure package for a whole paper, run batches from a YAML or JSON manifest, stitch panels into a labeled composite figure, polish an existing figure against a venue style guide, and score a generated diagram against a human reference using a vision-language model as judge. It runs as a CLI, a Python API, a local Gradio web UI, an MCP server, and a GitHub Action that keeps an Overleaf paper's methodology figure in sync with its text.