Quick Verdict
PaperBanana is useful for researchers who can verify every node in a scientific figure and want to shorten the first-draft stage. This page represents the active community fork at dwzhu-pku/PaperBanana; it does not create a separate listing for the upstream implementation. The repository is licensed under Apache 2.0 and organizes illustration generation into Retriever, Planner, Stylist, Visualizer, and Critic roles. It offers Gradio, Streamlit, and command-line entry points. As of July 21, 2026, the repository has no formal GitHub release, so it should be treated as moving research code rather than a versioned production application.
Visual polish is not the main decision criterion. A generated method figure can omit a branch, reverse an arrow, rename a variable, invent a component, or make a weak causal relationship look certain. The Critic stage is another model-assisted review, not proof of scientific correctness. Inputs may also be sent to external providers selected by the user, including Google, OpenAI, Anthropic, or OpenRouter. Unpublished papers, reviewer responses, invention material, and restricted datasets require a confidentiality review before submission. The repository disclaimer says Google has filed patents covering specific workflows and warns about similar commercial applications. That statement is not a legal determination here; commercial users should review the actual implementation, Apache 2.0 patent language, ownership, and relevant claims with qualified counsel.
Best For
- Researchers who understand their method and want several layout drafts before rebuilding the final figure.
- Paper authors who can compare candidates and inspect every label, edge, symbol, and visual encoding.
- Developers studying reference retrieval, multi-agent planning, visual generation, and model-based critique.
- Labs able to manage API credentials, provider terms, confidential-data classification, and usage budgets.
- Teams that plan to keep an editable human-authored final asset after using generated output as a sketch.
- Not ideal for anyone expecting a one-click publication figure or unable to send material to external model services.
Key Features
The pipeline divides work among five roles. Retriever selects relevant examples from a curated reference collection. Planner converts the method description and caption into a detailed visual plan. Stylist applies academic presentation guidance. Visualizer calls an image-generation model, and Critic proposes another round of changes. Several experiment modes allow direct generation, a Planner-only path, combinations with Stylist or Critic, or the full sequence. PaperBananaBench can support reference-driven runs; the application can also bypass retrieval when the dataset is absent.
The active community fork provides a local Gradio app, a Streamlit interface, and CLI experiments. Users can configure the main vision-language model separately from the image-generation model, generate multiple candidates, inspect intermediate stages, refine an uploaded image, and export files. “Local app” describes the interface and orchestration process, not necessarily inference. The configured API keys determine which external services receive prompts and images.
The project is flexible but not formally released. Cloning main gives a rolling snapshot whose dependencies, default model names, UI behavior, and prompts may change. Reproducible research requires a commit hash, environment lock, provider, exact model identifiers, generation parameters, input text, reference dataset state, and original outputs.
Use Cases
A careful paper workflow starts with a redacted method summary and a short caption. Generate low-resolution composition options first, then compare them against a written fact checklist: component names, edge directions, optional branches, parallel stages, color meanings, and all symbols. Only after the structure is correct should the team request higher-resolution refinement. The final figure should remain editable so spelling, typography, spacing, and scientific notation can be corrected without asking an image model to redraw the whole asset.
Statistical figures need a stricter boundary. Bars, points, confidence intervals, sample sizes, and significance markers should come from the underlying data and reproducible plotting code. PaperBanana may suggest a visual arrangement or help produce a conceptual companion diagram, but an image generator should not redraw measured values. For conceptual figures, compare every relationship with the manuscript and ask a subject expert to inspect claims that a reader could interpret as causal.
Confidentiality depends on provider routing. A configuration file can keep API keys out of Git, but it does not keep the paper text on the device. OpenRouter adds another routing layer, while direct provider calls remain governed by the selected account and model terms. Before uploading unpublished results, patent drafts, partner data, or ethics-controlled material, document who receives it, whether prompts are retained or used for training, where processing occurs, and how deletion works.
Pricing
The repository code is free under Apache 2.0. Operational use is not necessarily free because the workflow calls external text and image models. Cost varies with the provider, main model, image model, number of parallel candidates, resolution, critique rounds, and failed retries. Generating twenty candidates can multiply spend quickly. Run one redacted example first and record requests, latency, output quality, and human correction time before starting a batch.
There is no formal release as of July 21, 2026. Pinning a commit is therefore more meaningful than saying “latest PaperBanana.” Apache 2.0 covers the repository under its terms, but it does not automatically settle rights for external model outputs, reference images, fonts, uploaded material, datasets, or every patent question. The repository’s patent statement deserves a separate commercial review rather than an assumption in either direction.
Pros
- The five-role pipeline separates content planning, style guidance, generation, and critique.
- The active community fork includes Gradio, Streamlit, and CLI paths for different experiments.
- Main and image models can be selected separately, enabling provider and quality comparisons.
- Intermediate results make it easier to see where a factual or visual error entered the process.
- Apache-2.0 licensing gives a clear baseline for repository code.
- Multiple candidates are useful when the goal is exploring compositions rather than accepting the first image.
- PaperBananaBench and retrieval modes support research into academic-illustration generation.
Cons
- There is no formal release or stable compatibility promise.
- A visually convincing figure may still be scientifically wrong, incomplete, or misleading.
- The Critic is model-based and cannot certify accuracy.
- Local orchestration may send manuscript text and images to several external providers.
- Parallel candidates, critique rounds, and high resolution can raise cost and latency substantially.
- Reference images may introduce provenance, imitation, or publication-rights questions.
- Text, equations, symbols, and quantitative marks often need manual reconstruction in an editable tool.
- The repository disclaimer raises a patent caveat for similar commercial workflows that Apache licensing alone should not be assumed to resolve.
Alternatives
| Tool | Best fit | Main difference |
|---|---|---|
| Recraft | Designers who need consistent styles and vector assets | More mature design production, but no understanding of research-method logic |
| Adobe Firefly | Teams already using Adobe applications | Strong editing and asset workflow; academic multi-agent planning is not the focus |
| Gamma | Researchers creating talks and decks quickly | Better at pages and narrative flow than exact method diagrams |
| Beautiful.ai | Users who want automated presentation layout | Presentation-oriented and not a scientific fact-checking system |
| Ideogram | Users generating creative images containing text | Text rendering is central, but it lacks the academic retrieval and Critic pipeline |
FAQ
Which PaperBanana repository does this page cover?
It covers the active community fork at https://github.com/dwzhu-pku/PaperBanana. The page intentionally does not create another tool listing for the upstream implementation.
Does PaperBanana have a stable release?
No formal GitHub release existed on July 21, 2026. Pin a commit and record every model and environment detail when reproducibility matters.
Does running the interface locally keep a manuscript local?
No. The local UI orchestrates calls to configured providers such as Google, OpenAI, Anthropic, or OpenRouter. Unless the workflow is deliberately replaced with compatible local inference, prompts and images leave the machine.
Can a generated figure be submitted without revision?
It should not be. Verify labels, variables, values, arrows, sequence, citations, and caption alignment. Check the venue’s policy on generated content, and produce quantitative charts from data and code.
Does Apache 2.0 remove the patent concern?
That conclusion cannot be made from the license name alone. Apache 2.0 contains a patent grant tied to licensable contributor claims, while the repository disclaimer says Google filed patents for specific workflows. Commercial applicability requires legal analysis of the claims and implementation.
Bottom Line
PaperBanana is best understood as an academic-figure drafting framework and an active research project. Use the dwzhu-pku community fork, pin a commit because there is no formal release, and treat each external model provider as a real confidentiality and cost boundary. Every output needs a line-by-line scientific review, with quantitative graphics regenerated from source data where possible. Before commercial deployment, review model terms, reference provenance, Apache obligations, and the repository’s patent statement independently.