What it is
PaperBanana is a text-to-figure AI tool that converts academic text and simple sketches into publication-ready scientific illustrations. It runs a closed-loop five-agent architecture that coordinates planning, visualization, code generation, and aesthetic refinement. Users supply natural-language descriptions and choose a visual style through a web interface; the system returns methodology diagrams, statistical charts, system architectures, flow charts, and poster assets. For statistical figures PaperBanana generates executable Python Matplotlib code rather than static pixels to preserve numerical accuracy. The architecture includes a Planner agent that structures visual layouts and a Visualizer agent (using the Nano-Banana-Pro renderer) that produces precise shapes, connectors, and icons. The project is accompanied by a research paper and a GitHub project page, and the website provides a gallery of examples produced directly from text.
Key features
PaperBanana's core features center on an agentic, closed-loop five-agent pipeline that divides tasks among specialized components. The Planner agent interprets and sequences textual descriptions into structured layouts; the Visualizer agent renders diagrams using the Nano-Banana-Pro renderer to produce precise shapes, connectors, and scientific icons. For statistical plots the system outputs executable Python Matplotlib code to ensure mathematical fidelity and avoid numerical hallucination. The framework includes aesthetic refinement based on auto-summarized guidelines that adjust color schemes, typography, spacing, and iconography. The web interface exposes templates, style presets, and parameters such as aspect ratio, resolution, and output format for user control. PaperBanana also accepts rough sketches or whiteboard images as input and can convert them into polished figures while preserving original structure and scientific intent.
Use cases
PaperBanana is presented for tasks that require precise, publication-quality academic visuals. Typical use cases described on the site include automated generation of methodology and model-architecture diagrams (encoder–decoder frameworks, training pipelines, algorithm flows), creation of statistically accurate charts via executable Matplotlib code, and production of system and pipeline illustrations. The tool is also positioned for converting hand-drawn sketches or whiteboard notes into refined figures and for producing educational infographics in biology, chemistry, and physics. Additional uses include preparing poster assets, supplementary materials, and lecture-slide graphics. Researchers and educators can iterate on prompts and templates via the web interface to refine visualizations and adapt figures to publication or teaching formats.
