PaperBanana framework announced

Google’s PaperBanana is an agentic framework that turns raw data and text into polished academic figures, diagrams, and graphs with accurate values. Instead of manually wrestling with plotting libraries or prompting generic image models, researchers feed in data and metadata, and PaperBanana orchestrates a series of specialized agents.

The pipeline includes a retriever agent that collects relevant context and examples, a planner agent that decides how to encode and structure the visualization, and a stylist agent that optimizes colors, layout, and design. A critic loop then reviews outputs and asks for refinements across three iterations until the visualization meets quality standards.

PaperBanana is model-agnostic: it can use different image generators like GPT-Image 1.5 or NanoBanana Pro, though experiments show NanoBanana delivering better charts. Side-by-side comparisons with tools like Paper2Any and even human-made figures rate PaperBanana higher on conciseness, readability, aesthetics, and overall quality.

A GitHub repo is live, with the team promising to release code and datasets within about two weeks. That should make it easier for labs and students to standardize figures in theses and papers.