PaperBanana is a reference-driven multi-agent framework for automated academic illustration generation that transforms scientific content into publication-quality diagrams and plots.
The tool addresses the time-consuming challenge of creating high-quality figures for academic papers by automating the illustration pipeline. It works through an orchestrated sequence of specialized agents: a Retriever that finds relevant reference examples, a Planner that designs the illustration structure, a Stylist that applies visual conventions, a Visualizer that generates the actual diagrams, and a Critic that refines the output. The framework uses in-context learning from reference examples and iterative refinement to produce both aesthetically pleasing and semantically accurate scientific illustrations.
Researchers writing academic papers in computer science and related fields should consider this tool if they spend significant time manually creating or refining figures. The project is designed to support diverse illustration needs, though the README indicates the reference set currently focuses on computer science domains. The tool supports flexible model selection, allowing users to choose different vision-language models and image generation backends, including options from OpenAI, Anthropic, and other providers via OpenRouter. A web interface is available through Hugging Face Spaces for those who prefer not to run code locally.
The project maintains active development with recent additions including model selection in the user interface, OpenRouter integration for flexible API provider choice, and availability as a ClawHub skill. The README lists several planned features still under development, including support for manually selected examples, code for generating statistical plots, diagram improvement based on style guidelines, and expansion beyond computer science to other research domains. The project explicitly states its commitment to remaining fully open-source and community-focused, with no commercial plans.
