dwzhu-pku/PaperBanana ? reverse-engineered prompt

Build me a Python app called PaperBanana that helps researchers turn method text and figure captions into clean academic illustrations and plots.

I want a simple web interface where I can paste a paragraph from a paper, choose the figure size and style options, enter an API key, and click generate. The app should use a reference driven workflow, first finding similar example figures, then planning the diagram, refining the style, generating the image, and running a critic step to improve it if needed. It should also support running locally from the command line, and it would be great if the web app works nicely with Gradio or Streamlit.

Please include a way to load the benchmark dataset from Hugging Face, and make the setup easy with a config file for model names and API keys. If current docs matter for the models or image generation providers, look them up online. Keep the UI polished, show the generated result clearly, and make the whole thing easy to run with minimal setup.