Paperbanana

Name: Paperbanana
Author: llmsresearch

@llmsresearch

关于 Paperbanana

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.

#paperbanana #mcp #academic #diagram #plot #gemini #visualization #research #llm #claude #cursor

基本信息

分类

其他

许可证

MIT

运行时

python

传输方式

stdio

发布者

llmsresearch

提交者

Dipkumar Patel

配置

使用下面的配置,将此服务器添加到你的 MCP 客户端。

{
  "mcpServers": {
    "paperbanana": {
      "command": "uvx",
      "args": [
        "--from",
        "paperbanana[mcp]",
        "paperbanana-mcp"
      ],
      "env": {
        "GOOGLE_API_KEY": "<YOUR_GOOGLE_API_KEY>"
      }
    }
  }
}

工具

Generate a publication-quality methodology diagram from text. Args: source_context: Methodology section text or relevant paper excerpt. caption: Figure caption describing what the diagram should communicate. iterations: Number of refinement iterations (default 3). Returns: The generated diagram as a PNG image.

Generate a publication-quality statistical plot from JSON data. Args: data_json: JSON string containing the data to plot. Example: '{"x": [1,2,3], "y": [4,5,6], "labels": ["a","b","c"]}' intent: Description of the desired plot (e.g. "Bar chart comparing model accuracy"). iterations: Number of refinement iterations (default 3). Returns: The generated plot as a PNG image.

Evaluate a generated diagram against a human reference on 4 dimensions. Compares the model-generated image to a human-drawn reference using Faithfulness, Conciseness, Readability, and Aesthetics scoring with hierarchical aggregation. Args: generated_path: File path to the model-generated image. reference_path: File path to the human-drawn reference image. context: Original methodology text used to generate the diagram. caption: Figure caption describing what the diagram communicates. Returns: Formatted evaluation scores with per-dimension results and overall winner.

概览

What is Paperbanana?

PaperBanana is an agentic framework for generating publication-quality academic diagrams and statistical plots from natural language descriptions. It integrates with OpenAI, Azure OpenAI / Foundry, Google Gemini, and Atlas Cloud providers. The framework is designed for AI scientists who need automated, multi-agent pipeline support for academic illustration.

How to use Paperbanana?

Install via pip install paperbanana, set up an API key (OpenAI, Gemini, or Atlas Cloud) in .env or using paperbanana setup, then run paperbanana generate --input method.txt --caption "Your caption". Optionally use --optimize --auto for input optimization and auto-refine. A local Gradio web UI is available via paperbanana studio, and an MCP server enables IDE integration. Batch generation from manifest files is supported.

Key features of Paperbanana

Two-phase multi-agent pipeline with iterative refinement
Multiple VLM and image generation providers (OpenAI, Azure, Gemini, Atlas Cloud)
Input optimization layer for better generation quality
Auto-refine mode and run continuation with user feedback
CLI, Python API, and MCP server for IDE integration
Batch generation from manifest files (YAML/JSON)
PDF input support for methodology context
PaperBanana Studio – local Gradio web UI

Use cases of Paperbanana

Generating methodology diagrams for research papers from textual descriptions
Creating statistical plots from structured data with paperbanana plot-batch
Batch-generating multiple diagrams from a single manifest file
Iteratively refining diagrams with user feedback or auto-refine mode
Integrating with IDEs via the MCP server for inline diagram generation

FAQ from Paperbanana

What API keys are required?

You need an OpenAI API key (or Azure OpenAI / Foundry endpoint), a Google Gemini API key (free, from Google AI Studio), or an Atlas Cloud API key. Set them in .env or use paperbanana setup.

What are the runtime requirements?

Python 3.10+ and one of the supported API keys. Optional dependencies include PyMuPDF (paperbanana[pdf]) for PDF input and paperbanana[studio] for the web UI. Docker is also supported.

How does the iterative refinement loop work?

The pipeline runs up to 7 agents across two phases: input optimization (optional), linear planning with a Retriever, Planner, and Stylist, then iterative refinement between a Visualizer and Critic (default 3 iterations). With --auto, it continues until the critic is satisfied.

Can I use PDF documents as input?

Yes, install paperbanana[pdf] (or PyMuPDF) and use the --input paper.pdf --pdf-pages "3-8" flags to select specific pages for methodology context.

Is there a web interface for Paperbanana?

Yes, PaperBanana Studio is a local Gradio web UI started with paperbanana studio. It exposes the same workflows: diagram generation, statistical plots, evaluation, batch processing, and run browsing. Use --host, --port, --config, and --output-dir for configuration.