Question 1

What is PaperBanana?

Accepted Answer

PaperBanana is an agentic framework designed to automate the generation of publication-ready academic illustrations, including methodology diagrams, statistical plots, and educational infographics. PaperBanana bridges the gap between raw scientific text and professional visual communication by orchestrating five specialized AI agents in a closed-loop collaborative pipeline that ensures accuracy, faithfulness, and visual quality.

Question 2

How does PaperBanana's multi-agent framework work?

Accepted Answer

PaperBanana coordinates five specialized agents in sequence: the Retriever locates reference visual samples from a curated high-quality library, the Planner translates text into structured visual layouts, the Stylist synthesizes aesthetic guidelines for color and typography, the Visualizer renders the illustration using Nano-Banana-Pro or executable Python code, and the Critic conducts three rounds of VLM-powered self-reflection to catch errors. This closed-loop architecture is what makes the system uniquely faithful, precise, and reliable for generating publication-quality academic illustrations.

Question 3

How does PaperBanana ensure data accuracy in statistical plots?

Accepted Answer

Unlike general image generators that suffer from numerical hallucinations, PaperBanana uses a code-based paradigm for statistical charts. Instead of rendering charts as rasterized images, it generates executable Python Matplotlib code to ensure that bars, data points, and scales are mathematically precise and faithful to the raw data. This deterministic code-generation approach makes the framework far more reliable than pixel-based generation for any data-driven academic figure.

Question 4

What is Nano-Banana-Pro?

Accepted Answer

Nano-Banana-Pro is a state-of-the-art image generation model used by PaperBanana's Visualizer Agent. It specializes in synthesizing complex shapes, connectors, and scientific icons for methodology diagrams, excelling at following detailed layout instructions from the Planner agent while maintaining consistent visual styles throughout the entire illustration, from color palette to icon shapes.

Question 5

How is PaperBanana evaluated?

Accepted Answer

We introduced PaperBananaBench, a dedicated benchmark comprising 292 test cases curated from NeurIPS 2025 publications. PaperBanana consistently outperforms leading baselines including GPT-Image and Paper2Any across four key dimensions: faithfulness, conciseness, readability, and aesthetics. The framework achieves +17.0% overall score improvement, +37.2% conciseness improvement, and +12.9% readability improvement over the best competing system.

Question 6

Is PaperBanana open source?

Accepted Answer

Yes. PaperBanana is fully open source under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0). The complete codebase, model weights, training data, and PaperBananaBench benchmark are all publicly hosted on GitHub for the research community to freely use, reproduce, modify, and extend.

Question 7

What inputs does PaperBanana need to generate a diagram?

Accepted Answer

PaperBanana typically requires three inputs: a Visual Intent briefly describing the desired illustration, Source Context containing the relevant sections from your research paper, and a Figure Caption that will appear under the generated figure. For statistical plots, you also need to provide Raw Data in JSON or CSV format so the system can generate mathematically accurate charts through its code-based rendering pipeline.

Question 8

Can PaperBanana improve my existing hand-drawn sketches?

Accepted Answer

Yes. PaperBanana offers an Aesthetic Enhancement feature that applies auto-summarized aesthetic guidelines to elevate color schemes, typography, and the professional quality of human-drawn drafts. This capability is especially useful for researchers who already have a rough diagram they want to polish before submission to a top venue like NeurIPS, ICML, or ICLR.

Question 9

What types of academic figures can PaperBanana generate?

Accepted Answer

PaperBanana supports a wide range of academic illustration types: methodology diagrams depicting model architectures and system pipelines, statistical charts including bar charts, line graphs, and scatter plots rendered from raw data, educational infographics for teaching materials, poster and conference slide visual assets, and aesthetic refinement of existing hand-drawn sketches. The framework covers the full spectrum of visual communication needs commonly encountered in academic publishing and scientific research.

Question 10

How does PaperBanana compare to DALL-E or other image generators?

Accepted Answer

General-purpose image generators like DALL-E frequently produce errors in logical topology, connection accuracy, and numerical precision when creating academic figures. PaperBanana overcomes these limitations through its structured multi-agent pipeline: the Planner ensures logical correctness, code-based rendering guarantees data accuracy, and the Critic agent performs three rounds of iterative verification. As a result, the framework produces academic illustrations that are both visually polished and scientifically faithful — something generic image generators simply cannot reliably achieve when tasked with creating technical scientific figures.

PaperBanana: Automating Academic Illustration

Try PaperBanana Now

PaperBanana Gallery

Methodology Diagrams

Statistical Plots

Aesthetic Enhancement

Educational Infographics

Aesthetic Refinement

Loved by researchers

Simple, transparent pricing

Hobby

Basic

Pro

Frequently Asked Questions

Start Creating with PaperBanana