PaperBanana Is Now PaperVizAgent—but Still a Research Prototype

PaperBanana is now called PaperVizAgent, and its original research framework has gained an official runnable implementation. On April 8, 2026, Google Research’s PaperVizAgent introduction presented the new name and classified the system as an experimental research prototype rather than a production-ready publishing tool.

The central idea has not changed: specialized AI agents turn manuscript context and a figure caption into methodology diagrams or statistical plots. What has changed is the project’s practical status—researchers can inspect and run the code, while the explicit prototype designation makes clear that generated figures still require scientific and editorial verification.

From PaperBanana paper to PaperVizAgent code

The original PaperBanana paper was posted on January 30, 2026. Its arXiv research record describes a reference-driven framework evaluated on 292 methodology-diagram test cases drawn from NeurIPS 2025 publications, with assessments covering faithfulness, conciseness, readability and aesthetics.

The rename separates the continuing software project from the title under which the research first circulated. PaperBanana remains the paper’s name, while PaperVizAgent is the current name of the authors’ implementation. That distinction also helps separate the research framework from unrelated websites and community projects using similar PaperBanana branding.

The official PaperVizAgent repository documents local installation with Python 3.12, model credentials, a Streamlit interface and command-line operation. It supports candidate generation, inspection of intermediate revisions and refinement of an uploaded diagram; the repository also marks PaperBananaBench for a later release and allows execution without it by bypassing retrieval-based few-shot examples.

Public code therefore makes the framework testable, but not self-contained. Users must supply access to external language, vision and image-generation models, and the selected providers and configuration can affect cost, output quality and reproducibility.

How the five-agent pipeline constructs a figure

PaperVizAgent divides illustration generation among five roles instead of asking one model to produce the final image in a single pass. The inputs are normally a focused passage from a manuscript and a detailed caption expressing what the figure should communicate.

• Retriever: selects related academic diagrams to provide visual and structural references.
• Planner: converts the source material and caption into a detailed description of components and relationships.
• Stylist: refines that description using aesthetic guidance synthesized from reference figures.
• Visualizer: renders a methodology diagram or generates executable Python code for a statistical plot.
• Critic: compares the output with the source context and returns targeted changes to the visualizer.

The division addresses two different problems: deciding what the figure must contain and deciding how to display it. Retrieval, planning and styling happen before rendering, while the critic creates a revision loop after an initial result exists.

That loop should not be mistaken for a correctness guarantee. A critic model can request useful revisions while still overlooking a reversed arrow, duplicated component, missing dependency or inaccurate label. Its presence reduces neither the author’s responsibility for the figure nor the need to compare every visible claim with the manuscript.

What the benchmark result actually supports

The reported benchmark advantage applies to the complete tested configuration, not to every image model or to the visualizer component in isolation. It shows that the multi-agent arrangement performed better than the study’s selected baselines under its evaluation method; it does not establish universal accuracy across scientific fields, journals or model configurations.

The benchmark is also concentrated on methodology diagrams from one computer-science conference. Those examples can test complicated pipelines and model architectures, but they do not represent every visual convention used in medicine, chemistry, engineering or the social sciences. A strong score on that collection cannot guarantee that a generated figure follows a different discipline’s notation or reporting requirements.

Statistical plots create an additional distinction between appearance and numerical fidelity. An image generator may produce an attractive chart without preserving every value or repeated element correctly. Code-based plotting offers a more inspectable path when exact data matter because the inputs, calculations and rendering instructions remain available for review.

What researchers can realistically expect

PaperVizAgent is best understood as a drafting and revision framework for research visuals. It can organize dense method descriptions, produce alternative compositions and shorten the distance between a caption and a usable first figure, especially when the input clearly defines components, relationships and direction of flow.

Its output is not automatically ready for submission simply because the project targets publication-quality illustrations. For a methodology diagram, labels, arrows, ordering and visual grouping must be checked against the paper. For a plot, the underlying data, units, axes, aggregation rules and code require their own review.

Editability is another practical boundary. The documented workflows primarily generate or refine raster images, while many publication processes benefit from vector objects, editable text and reusable chart code. Researchers may still need conventional graphics or plotting software to make precise final changes and satisfy a venue’s file requirements.

The meaningful update, then, is not that AI has eliminated figure design. PaperBanana has evolved into a named, inspectable implementation that coordinates multiple models around the task, but PaperVizAgent remains a research prototype whose polished results can conceal substantive errors. Its value lies in accelerating a carefully supervised workflow—not replacing scientific judgment.

Also read:

• Not Microdramas, But TikTok Dramas: The Platform Is Now Actively Producing Its Own Mini-Series
• Blog Link Building Still Works—But Its Old Shortcuts Now Count as Spam
• AI Is Scaling Video Ads—But Humans Still Own the Claims