llmsresearch/paperbanana

What it solves

PaperBanana automates the creation of publication-quality academic diagrams and statistical plots. It removes the manual effort of designing complex methodology figures and plotting data for AI research papers, ensuring they follow professional aesthetic guidelines (like those of NeurIPS).

How it works

The system uses a multi-agent pipeline consisting of up to seven specialized agents:

1. Input Optimization: An Input Optimizer uses a Context Enricher and Caption Sharpener to structure raw text and refine captions into precise visual specifications.
2. Linear Planning: A Retriever finds relevant reference examples, a Planner creates a detailed textual description of the diagram, and a Stylist applies aesthetic guidelines (color, layout, typography).
3. Iterative Refinement: A Visualizer renders the image, and a Critic evaluates it against the source context, providing feedback for revisions. This loop repeats until the image is satisfactory or a limit is reached.

For statistical plots, the system generates matplotlib code via a VLM to render data from CSV or JSON files.

Who it’s for

AI scientists and researchers who need to generate professional figures, methodology diagrams, and data plots for academic publications.

Highlights

• Multi-Provider Support: Works with OpenAI, Google Gemini, Azure OpenAI, and Atlas Cloud.
• Iterative Refinement: Features an "auto-refine" mode where a Critic agent loops until the output meets quality standards.
• Flexible Inputs: Supports raw text, PDF files (with page selection), and reference sketch images to guide the layout.
• Batch Processing: Can generate multiple diagrams or plots from a manifest file and automatically stitch them into composite figures.
• PaperBanana Studio: Includes a local Gradio web UI for managing workflows and browsing outputs.
• Venue Style Packs: Built-in and customizable style guides for venues like NeurIPS, ICML, ACL, and IEEE.