PaperBanana Introduction

PaperBanana is an agentic framework designed to automate the creation of academic illustrations for AI researchers. The system employs a multi-agent workflow—Retriever, Planner, Renderer, and Critic—to transform textual descriptions or rough sketches into publication-ready methodology diagrams and statistical plots. Users can generate figures from scratch by providing context and captions, or upload hand-drawn sketches for digital polishing. The framework emphasizes academic precision, using iterative self-critique to enhance faithfulness, conciseness, and aesthetics. Benchmarking against standards from top conferences like NeurIPS, PaperBanana aims to reduce the time spent on figure generation. As an open-source project, it provides code, data, and a benchmark (PaperBananaBench) to support the research community.

PaperBanana operates as an agentic framework that automates academic illustration for researchers. Its workflow orchestrates specialized agents: a Retriever gathers source context, a Planner designs the layout, a Renderer generates the initial image using vision-language models, and a Critic performs iterative self-critique to refine outputs. The system accepts textual descriptions or rough sketches, producing publication-ready methodology diagrams and statistical plots. This process emphasizes faithfulness, conciseness, and aesthetic standards suitable for top-tier conferences. By automating the figure creation bottleneck, PaperBanana allows researchers to focus on content while ensuring vector-quality, standardized visual assets.

PaperBanana is an agentic framework designed to automate the creation of academic illustrations for AI researchers. It generates publication-ready methodology diagrams and statistical plots directly from text descriptions or rough sketches. The system employs a multi-agent workflow—Retriever, Planner, Renderer, and Critic—to iteratively refine outputs, ensuring high fidelity, conciseness, and adherence to conference standards. By handling both text-to-diagram generation and sketch polishing, PaperBanana addresses the time-intensive bottleneck of figure production. It is open-source, includes the PaperBananaBench benchmark (292 NeurIPS 2025 test cases), and integrates state-of-the-art vision-language models for reliable, vector-quality visuals.

Pros

• Automates academic illustration creation efficiently.
• Agentic framework improves diagram reliability.
• Supports both text and sketch inputs.
• Benchmarked for publication standards.

Cons

• Credit-based pricing may increase costs.
• Configuration parameters require user expertise.
• Output accuracy depends on input quality.
• Limited to methodology diagrams and plots.