Paperbanana

PaperBanana — Academic Diagram Generation

Generate publication-quality methodology diagrams, architecture overviews, and statistical plots from text descriptions using a multi-agent VLM pipeline with iterative refinement.

When to Use

• Creating methodology diagrams for research papers
• Generating architecture overview figures from text descriptions
• Building statistical plots from CSV data for publications
• Producing NeurIPS/ICML-style figures from plain-language descriptions
• Upgrading rough diagrams-as-code into polished academic illustrations
• Batch generating figures for a multi-figure paper or report

When Not To Use

• For interactive or web-based diagrams — use the mermaid-diagrams skill instead
• For Wardley maps or custom TikZ diagrams — use report-builder or latex-documents
• For photo editing or raster manipulation — use the imagemagick skill
• For simple bar/line charts from data — use Python matplotlib directly
• For 3D visualisations — use the blender skill
• For accessible, re-editable SVG/HTML diagrams (architecture, flowcharts, org charts) — use the diagram-design skill instead. Both claim "publication-quality" but render differently: paperbanana produces a VLM-rendered raster figure matching a specific journal's visual convention; diagram-design produces re-editable, accessible SVG/HTML.

Prerequisites

Unlike most of the diagrams/documents skills, paperbanana is not baked into this image — it is an external PyPI package installed on first use, gated on a provisioned API key. Do not assume the zero-install parity that e.g. docs-alignment or mermaid-diagrams have.

# Check installation
paperbanana --help

# If not installed:
pip install "paperbanana[google]"  # For Gemini (free tier)
pip install "paperbanana[openai]"  # For OpenAI
pip install "paperbanana[google,openai]"  # Both

API Keys (at least one required):

• GOOGLE_API_KEY — Gemini (free tier available, recommended for cost)
• OPENAI_API_KEY — OpenAI gpt-image-1

Default image model is nano-banana-2 (gemini-3.1-flash-image-preview); nano-banana-pro (gemini-3-pro-image-preview) for maximum quality. See references/providers-and-config.md.

Quick Start

Generate a methodology diagram

# From a text file describing your method
paperbanana generate \
  --input method_description.txt \
  --caption "Overview of the proposed framework" \
  --optimize --auto

# From inline text
echo "Our system uses a hierarchical mesh of specialist agents..." | \
  paperbanana generate --input - --caption "Agent Architecture" --auto

Generate a statistical plot

paperbanana plot \
  --data results.csv \
  --intent "Bar chart comparing F1 scores across models with error bars"

Interactive setup (first time)

paperbanana setup  # Guided wizard, works with free Gemini API

References

Deep content lives in references/ — load on demand:

• references/cli-reference.md — full flags for generate, plot, evaluate, batch, and the batch manifest YAML format.
• references/pipeline-architecture.md — the two-phase multi-agent pipeline (optimise → plan → refine) and the critic's four evaluation dimensions.
• references/providers-and-config.md — provider/model table (nano-banana-2 / Gemini 3.x era), env vars, and .env setup.
• references/python-and-mcp.md — Python API usage and the MCP server integration (tools, settings block).
• references/integration.md — how PaperBanana slots into the report-builder pipeline for LaTeX figures.
• references/troubleshooting.md — API-key, quality, rate-limit, and description-mismatch fixes.