# PaperBanana ## Automating Academic Illustration for AI Scientists An agentic framework that orchestrates five specialized AI agents to transform raw scientific content into **publication-ready** methodology diagrams and statistical plots — no design skills required. Dawei Zhu· PKU\*Rui Meng· GoogleYale Song· GoogleXiyu Wei· PKUSujian Li· PKUTomas Pfister· GoogleJinsung Yoon· Google [🍌 Create Illustration](/content/paper-banana-create/index.html) [arXiv Paper](https://arxiv.org/abs/2601.23265) [View Code](https://github.com/dwzhu-pku/PaperBanana) 292 ## Test Cases from NeurIPS 2025 5 ## Agents specialized AI agents 4/4 ## Benchmarks Won faithfulness, conciseness, readability, aesthetics ### Abstract Despite rapid advances in autonomous AI scientists powered by language models, generating publication-ready illustrations remains a labor-intensive bottleneck in the research workflow. To lift this burden, we introduce PaperBanana, an agentic framework for automated generation of publication-ready academic illustrations. Powered by state-of-the-art VLMs and image generation models, PaperBanana orchestrates specialized agents to retrieve references, plan content and style, render images, and iteratively refine via self-critique. To rigorously evaluate our framework, we introduce **PaperBananaBench**, comprising **292 test cases** for methodology diagrams curated from NeurIPS 2025 publications. ### VLMs State-of-the-art vision-language models ### Multi-Agent 5 specialized collaborative agents ### Self-Critique Iterative quality refinement loop ### NeurIPS 2025 Benchmark from top venue ## Multi-Agent Architecture ### Five Agents, One Mission Each agent specializes in a critical stage of the illustration pipeline, working in concert to deliver publication-quality results. #### 🔍 Retriever Identifies relevant reference examples from academic databases to guide style and content alignment. #### → #### 📐 Planner Translates scientific content into a detailed visual plan, decomposing structure and layout. #### → #### 🎨 Stylist Enforces academic aesthetic standards — color palettes, typography, line weights, and visual hierarchy. #### → #### 🖼️ Visualizer Renders the initial image or generates Python/Matplotlib plotting code from the visual plan. #### → #### 🔬 Critic Performs iterative self-critique, inspecting generated results against source content and triggering refinement. ### Generated Illustrations #### MethodologyStatistical Plot InputLLM CorePlannerStylistOutputself-critique loopMulti-Agent Pipeline ✓Generated by PaperBanana · Multi-agent pipeline with iterative Critic refinement ## Capabilities ### What PaperBanana Can Do 🗂️ NeurIPS-ready #### Methodology Diagrams Neural networks, flowcharts, multi-agent pipelines, and complex system architectures — all rendered to publication standards. 📊 Data-exact #### Statistical Plots Accurate data visualization via Matplotlib code generation. Bar charts, ablation studies, accuracy comparisons — zero hallucination guaranteed. ✏️ Sketch input #### Sketch-to-Pro Transform rough hand-drawn sketches into clean, harmonious academic figures with consistent fonts and styling. ✨ Polish mode #### Aesthetic Refinement Upload existing diagrams to upgrade fonts, colors, and spacing without altering underlying content or structure. 📚 Context-aware #### Reference-Driven Retrieves relevant papers to align style with academic conventions — your figures will match the venue aesthetic. 🔄 Auto-refine #### Iterative Refinement Self-critique loop ensures publication-quality output. The Critic agent reviews and forces regeneration until quality passes. ### PaperBananaBench ## Consistently Outperforms All Baselines Evaluated on 292 test cases from NeurIPS 2025, PaperBanana leads across all four dimensions: faithfulness to source content, conciseness, readability, and visual aesthetics. ### Faithfulness Baseline: 61%PaperBanana: 87% ### Conciseness Baseline: 58%PaperBanana: 82% ### Readability Baseline: 70%PaperBanana: 91% ### Aesthetics Baseline: 63%PaperBanana: 85% ### PaperBananaBench 292 ## Test Cases NeurIPS ### Source Venue 2025 ### Publications 4 Dims ### Evaluation Axes The benchmark covers diverse research domains and illustration styles, representing the breadth of modern AI research publications. ## How It Works ### From Text to Publication Figure in Seconds Describe your methodology or paste your data. PaperBanana's agents collaborate to deliver a figure that passes peer review. ### STEP 01 ✍️ Describe Input your methodology, data, or sketch. PaperBanana accepts text, PDFs, or images. ### STEP 02 🔍 Retrieve The Retriever agent scans reference databases to find style-aligned academic examples. ### STEP 03 📐 Plan & Style Planner and Stylist agents create a detailed visual plan with academic-grade aesthetics. ### STEP 04 ✨ Render & Refine Visualizer renders the figure; Critic reviews and iterates until quality is publication-ready. ## Citation ### Cite This Work ```bibtex @article{zhu2026paperbanana, title={PaperBanana: Automating Academic Illustration for AI Scientists}, author={Zhu, Dawei and Meng, Rui and Song, Yale and Wei, Xiyu and Li, Sujian and Pfister, Tomas and Yoon, Jinsung}, journal={arXiv preprint arXiv:2601.23265}, year={2026} } ``` 🏛️ Peking University ☁️ Google Cloud AI Research Corresponding authors: [dwzhu@pku.edu.cn](mailto:dwzhu@pku.edu.cn) · [lisujian@pku.edu.cn](mailto:lisujian@pku.edu.cn) · [jinsungyoon@google.com](mailto:jinsungyoon@google.com) 🍌 PaperBanana arXiv:2601.23265 · Peking University × Google Cloud AI [arXiv](https://arxiv.org/abs/2601.23265) [GitHub](https://github.com/dwzhu-pku/PaperBanana)