PaperBanana
About this tool
A multi-agent framework for academic illustrations. Five agents retrieve references, plan, set style, draw and review, learning from published figures before generating and revising candidates. It mainly produces method and workflow diagrams. The associated paper was an ICML 2026 Spotlight. Gradio, Streamlit, CLI and SKILL.md interfaces are available. Generated figures still require human checks.
- Feature 1: Retriever, Planner, Stylist, Visualizer and Critic agents.
- Feature 2: Retrieve reference figures and learn styles from leading conference papers.
- Feature 3: Generate candidates in parallel and revise through multiple review rounds.
- Feature 4: Gradio, Streamlit and CLI interfaces, plus a Hugging Face Spaces demo.
- Feature 5: skill/SKILL.md generates method diagrams from methodology text and captions.
Resource information
- Maintainer
- Maintained by first author Dawei Zhu (dwzhu-pku); the original implementation was released by Google Research as PaperVizAgent.
- Type
- Standalone app
- Category
- Scientific figures & presentations
- Runs on
- Agent application
- Who it's for
- AI, computing and related researchers drafting methods, model-architecture or workflow figures for submissions and presentations.
- License at time of listing
- Apache-2.0; skill/SKILL.md is separately marked MIT-0.
- Source
- GitHub
