7.2K ·GitHub
What is this?
PaperBanana is an open-source framework that turns raw scientific content into publication-quality diagrams and plots, acting like a creative team of five specialized agents: Retriever, Planner, Stylist, Visualizer and Critic.
PaperBanana is an open-source framework that turns raw scientific content into publication-quality diagrams and plots, acting like a creative team of five specialized agents: Retriever, Planner, Stylist, Visualizer and Critic.Is it a framework, a web app, or a command-line tool?It's a Python framework you can use three ways: a Gradio web app (recommended), an interactive Streamlit demo, or a CLI. There's also a hosted version on Hugging Face Spaces you can try with no setup at all.
It's a Python framework you can use three ways: a Gradio web app (recommended), an interactive Streamlit demo, or a CLI. There's also a hosted version on Hugging Face Spaces you can try with no setup at all.How do I set it up locally, and do I need to download any dataset?Clone the repo, copy configs/modelconfig.template.yaml to configs/modelconfig.yaml with your API key, install the Python 3.12 packages via uv, then launch Gradio or Streamlit. The PaperBananaBench dataset is optional; without it the Retriever just skips its few-shot examples.
Clone the repo, copy configs/modelconfig.template.yaml to configs/modelconfig.yaml with your API key, install the Python 3.12 packages via uv, then launch Gradio or Streamlit. The PaperBananaBench dataset is optional; without it the Retriever just skips its few-shot examples.Do I need my own API key to run it?Yes — at least one of Google Gemini or OpenRouter; either alone is enough, and OpenRouter is preferred if you configure both. For generating many candidates in parallel you will need a key that supports high concurrency.
Yes — at least one of Google Gemini or OpenRouter; either alone is enough, and OpenRouter is preferred if you configure both. For generating many candidates in parallel you will need a key that supports high concurrency.What are its current limitations?The authors admit generation is not yet fully reliable for diverse, complex scenarios. The curated reference set covers mainly computer science, and code for generating statistical plots is still on the TODO list, so diagrams are the mature path today.
The authors admit generation is not yet fully reliable for diverse, complex scenarios. The curated reference set covers mainly computer science, and code for generating statistical plots is still on the TODO list, so diagrams are the mature path today.How does it make figures look academically polished?The Retriever finds similar examples from a curated reference collection for in-context learning, the Planner drafts a textual figure description, and automatically synthesized style guidelines shape the output; a Critic-Visualizer loop then refines the image over multiple rounds.
The Retriever finds similar examples from a curated reference collection for in-context learning, the Planner drafts a textual figure description, and automatically synthesized style guidelines shape the output; a Critic-Visualizer loop then refines the image over multiple rounds.How many figure candidates can it generate, and can I improve one I like?Up to 20 candidates in parallel, shown in a grid with evolution timelines and exportable as PNGs or a batch ZIP. The Refine tab lets you upload any candidate (or your own diagram), describe changes, and upscale the result to 2K or 4K.
Up to 20 candidates in parallel, shown in a grid with evolution timelines and exportable as PNGs or a batch ZIP. The Refine tab lets you upload any candidate (or your own diagram), describe changes, and upscale the result to 2K or 4K.Can I run only part of the agent pipeline?Yes. Experiment modes range from vanilla (direct generation, no planning or refinement) to devfull with all five agents, with intermediate stops like devplanner and devplannercritic; demo modes run the same pipelines without evaluation.
Yes. Experiment modes range from vanilla (direct generation, no planning or refinement) to devfull with all five agents, with intermediate stops like devplanner and devplannercritic; demo modes run the same pipelines without evaluation.Where did PaperBanana come from?It forked Google-Research's PaperVizAgent and keeps evolving it. The project is fully open source under the Apache-2.0 license with no commercial plans — the authors' stated goal is simply to benefit the research community.
It forked Google-Research's PaperVizAgent and keeps evolving it. The project is fully open source under the Apache-2.0 license with no commercial plans — the authors' stated goal is simply to benefit the research community.What are similar projects?CodeVideoModel, latent-diffusion, InstantID, WanGP, RSIGame, AutoFigure-Edit, MOSS-Image, RedInk
CodeVideoModel, latent-diffusion, InstantID, WanGP, RSIGame, AutoFigure-Edit, MOSS-Image, RedInkWhich topics is it in?streamlit, gradio, openrouter, text-to-image, data-visualization, academic-research, agent-orchestration, image-generation, gemini, generative-ai, huggingface, multi-agent
streamlit, gradio, openrouter, text-to-image, data-visualization, academic-research, agent-orchestration, image-generation, gemini, generative-ai, huggingface, multi-agent