paperbanana

What it is

Open source implementation and extension of Google Research’s PaperBanana for automated academic figures, diagrams, and research visuals, expanded to new domains like slide generation. It is described as an agentic framework for generating publication-quality diagrams and statistical plots from text descriptions, with multiple VLM and image generation providers and a CLI, Python API, and MCP server for IDE integration.

How it works

PaperBanana implements a multi-agent pipeline with up to 7 specialized agents in phases:

• Phase 0: Input Optimization (optional) using Context Enricher and Caption Sharpener to structure text for diagrams and refine captions.
• Phase 1: Linear Planning with Retriever, Planner, and Stylist to select references, describe the diagram, and refine visuals per NeurIPS-style guidelines.
• Phase 2: Iterative Refinement with Visualizer and Critic, repeating for a fixed number of iterations or until auto refinement criteria are met. It supports multiple providers for VLM and image generation, including OpenAI, Atlas Cloud, Google Gemini, and OpenRouter, with autodetection for Azure OpenAI / Foundry and Gemini gateways.

Getting started

Prerequisites include Python 3.10+, an API key for an OpenAI or Gemini provider, or a Gemini API key. Installation steps shown in the README:

• Install via PyPI: ```bash pip install paperbanana

- Or install from source and install extra dependencies: ```bash
git clone https://github.com/llmsresearch/paperbanana.git
cd paperbanana
pip install -e ".[dev,openai,google]"

Docker usage is provided to build and run with environment keys and volume mounting for inputs/outputs, e.g.:

docker build -t paperbanana .
docker run --rm -e GOOGLE_API_KEY paperbanana generate --help

Generation can be performed with a given input file and caption:

paperbanana generate \
  --input examples/sample_inputs/transformer_method.txt \
  --caption "Overview of our encoder-decoder architecture with sparse routing"

And output is saved under outputs/run_<timestamp>/final_output.png with intermediate data.

Optional: PaperBanana Studio (local web UI) can be installed and started:

pip install 'paperbanana[studio]'
paperbanana studio

Step 2: Get Your API Key

The guide shows copying .env.example to .env and filling API keys for OPENAI_API_KEY, GOOGLE_API_KEY, etc., plus optional Azure/OpenAI base URLs and Gemini overrides. If Gemini is used, there is a setup command: ```bash paperbanana setup

### Step 3: Generate a Diagram
The core generate command is demonstrated, including options for input, caption, and optional features like --optimize and --auto:
```bash
paperbanana generate \
  --input examples/sample_inputs/transformer_method.txt \
  --caption "Overview of our encoder-decoder architecture" \

Advanced usage includes --pdf-pages, --image inputs, and continuing runs with --continue and --continue-run.

Recent releases

Latest releases include:

• v0.3.0 (2026-06-12): PaperBanana v0.3.0 with bug fixes across issues and PRs.
• v0.2.0 (2026-06-11): Major refresh after 150+ commits; improved fresh installs and dev setup.
• bench-data-v1: PaperBananaBench dataset mirror (v1) from a related dataset.
• v0.1.2 (2026-02-13): Initial feature set including OpenRouter provider and sample input enhancements.

Traction

Stars: 2214 Forks: 327 Open issues: 10

License

MIT

Behind the repo

Project is described as an unofficial, community-driven open-source implementation of PaperBanana by llmsresearch, with multiple providers and an MCP server, and integration via CLI, Python API, and Gradio Studio.

Caveats

License MIT is stated; repository age started 2026-02-04, last push 2026-07-24. Prerequisites include Python 3.10+, and API keys for OpenAI, Azure/OpenAI, Gemini, or Atlas Cloud depending on provider configuration. The README notes optional Docker usage and a Colab Colab quickstart notebook for browser-based, no-local-setup usage.