Google PaperBanana AI: The Agentic Framework That Automates Academic Illustrations — Complete Guide

TamilTech AI summary

Google researchers and Peking University built PaperBanana, an agentic AI framework where five specialized agents—Retriever, Planner, Stylist, Visualizer, and Critic—read a research paper and turn it into publication-ready academic illustrations in about two to five minutes instead of the usual four to eight hours. It matters because diagrams are one of the biggest time sinks for PhD students and labs, and blind evaluations gave PaperBanana a 72.7% win rate over baselines plus roughly 75% preference over diagrams made by human PhD experts, with clearer, more focused figures on the new PaperBananaBench set of 292 NeurIPS-style cases. Under the hood it uses models like Gemini-3-Pro for planning and critique, image generators for methodology diagrams, and Matplotlib code for statistical plots, with NeurIPS-style formatting baked in. You can try it through a web tool, a promised open-source GitHub repo, or a Pro API around $9.90 a month, and free or low-cost paths help a lot for Indian researchers who publish huge volumes of papers each year. One thing you should know is that content fidelity is still only about 45.8% and outputs are raster images, so the smart workflow is generate, manually review labels and arrows, then fine-tune before you submit.

• What is Google PaperBanana and how does it work?
• Is PaperBanana free to use?
• How does PaperBanana compare to DALL-E or Midjourney for academic diagrams?
• What are the main limitations of PaperBanana?

AI-assisted summary, checked by the TamilTech editorial team.