Google's PaperBanana uses five AI agents to auto-generate scientific diagrams

2026-02-07 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, short

Summary

Peking University and Google Cloud AI Research have developed PaperBanana, an AI system that uses five specialized agents to automatically generate publication-ready scientific diagrams from method descriptions. The system, based on Google's Nano Banana, aims to automate the manual bottleneck of creating illustrations for research papers. In evaluations, human reviewers preferred PaperBanana's diagrams over simple image generation in nearly 73% of cases, demonstrating its effectiveness in visual appeal. However, the system achieved only 45.8% content fidelity, struggling with accuracy issues like misaligned lines and arrows. For statistical plots, PaperBanana generates Python code for Matplotlib to ensure numerical accuracy, contrasting with direct image generation for other diagrams. The system also offers a feature to visually upgrade existing human-made diagrams, with refined versions preferred 56.2% of the time.

Key takeaway

For AI Researchers and Research Scientists aiming to streamline scientific illustration, PaperBanana demonstrates that multi-agent AI can significantly improve diagram aesthetics and generation efficiency. However, you should remain vigilant about content accuracy, as the system currently achieves only 45.8% fidelity, necessitating manual review for critical details like connecting lines and arrows. Consider generating multiple outputs and using code-based methods for statistical plots to mitigate accuracy risks.

Key insights

A multi-agent AI system automates scientific diagram generation, improving aesthetics but still facing content accuracy challenges.

Principles

• Specialized AI agents enhance complex task automation.
• Code-based generation ensures numerical accuracy for plots.
• Separating content and style improves visual refinement.

Method

PaperBanana employs five agents: template search, description translation, aesthetic refinement, image rendering (or code generation for plots), and a quality control critic, running a generation-criticism cycle three times.

In practice

• Generate multiple diagram versions to select the best.
• Use AI for aesthetic refinement of existing diagrams.
• Consider code-based tools for precise data visualization.

Topics

• AI Agents
• Scientific Diagram Generation
• Automated Illustration
• Content Fidelity
• Matplotlib

Best for: AI Researcher, AI Scientist, Research Scientist

Related on AIssential

• Are you still manually fighting with LaTeX and TikZ to create publication-quality figures? — Scientific figure generation benefits from multi-agent coordination addressing localized errors, not monolithic models.
• FAUNA -> FLORA’s AI agent efficiency — FAUNA is an integrated AI agent that enhances creative workflows by understanding visual context within a canvas.
• Improving the academic workflow: Introducing two AI agents for better figures and peer review — AI agents can automate complex academic tasks like figure generation and peer review, improving efficiency and quality.
• The Comic That Draws Itself: Building a Daily AI Graphic-Novel Studio — AI-generated comics demand engineered visual and narrative consistency across panels and episodes, not reliance on single prompts.
• GenClaw: Code-Driven Agentic Image Generation — GenClaw enables controllable, interpretable image generation by integrating code-driven sketching with generative models.
• Toward AI VIS Co-Scientists: A General and End-to-End Agent Harness for Solving Complex Data Visualization Tasks — An AI agent harness automates end-to-end visual analysis application design, bridging data to interactive interfaces.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.