Google AI has introduced PaperBanana, a new agentic AI framework designed to automate the creation of publication-ready methodology diagrams and statistical plots for academic papers. The framework aims to solve a persistent challenge in scientific publishing: translating complex methods and data into clear, professional visuals that meet strict conference and journal standards.

Creating figures is often a time-consuming process for researchers, requiring manual design, coding, and repeated revisions. PaperBanana addresses this gap by converting textual descriptions and experimental data directly into structured diagrams and accurate plots, significantly reducing the effort required to prepare visuals for submission.

Why Research Visuals Matter

Methodology diagrams and statistical plots play a critical role in communicating research ideas. While text explains concepts, visuals help reviewers and readers quickly grasp workflows, model architectures, and experimental results. However, producing these figures usually demands expertise in design tools, LaTeX environments, or plotting libraries.

PaperBanana is designed to streamline this process by automating visual generation while preserving scientific accuracy and clarity. The goal is not only speed, but also consistency with academic formatting and visual conventions.

How PaperBanana Works

PaperBanana is built as a multi-agent system, where each agent handles a specific part of the visual creation pipeline. Instead of relying on a single model, the framework coordinates several specialised agents to produce higher-quality results.

This agentic workflow allows PaperBanana to improve outputs over multiple steps, resulting in clearer and more faithful representations of the original research content.

Key components include:

  • A retrieval agent that identifies relevant reference figures and styles.
  • A planning agent that converts textual method descriptions into structured visual plans.
  • A styling agent that ensures figures align with publication norms.
  • A visualisation agent that generates diagrams or executable plotting code.
  • A critical agent that evaluates outputs and guides iterative refinement.

Focus on Accuracy for Statistical Plots

One notable feature of PaperBanana is its approach to statistical plots. Rather than generating charts purely as images, the system can produce executable code for plots. This ensures numerical correctness and avoids common issues such as distorted scales or incorrect data representation, which can occur in purely image-based generation.

This design choice is crucial for research papers, where even minor numerical inaccuracies can undermine credibility.

How Well Does PaperBanana Perform in Real-World Benchmarks?

To evaluate performance, the researchers introduced a benchmark based on hundreds of real methodology diagrams from top-tier AI conferences. Across key criteria such as faithfulness to text, readability, conciseness, and visual quality, PaperBanana outperformed existing baseline approaches.

The results suggest that agentic collaboration and reference-driven generation can significantly improve the quality of automatically generated academic visuals.

What This Means for Researchers

PaperBanana highlights a broader shift toward automating supporting tasks in research workflows. By reducing the time spent on figure creation, researchers can focus more on experimentation, analysis, and writing. While initially targeted at AI research papers, the framework could be extended to other scientific disciplines that rely heavily on structured diagrams and data visualisation.

As AI-assisted research tools mature, PaperBanana represents an important step toward making high-quality scientific communication faster, more accessible, and less dependent on manual design effort.