Paperbanana
v0.1.1Generate publication-quality academic diagrams, methodology figures, architecture illustrations, and statistical plots from text descriptions using the Paper...
Like a lobster shell, security has layers — review code before you run it.
Runtime requirements
PaperBanana — Academic Illustration Generator
Generate publication-quality academic diagrams and statistical plots from text descriptions. Uses a multi-agent pipeline (Retriever → Planner → Stylist → Visualizer → Critic) with iterative refinement.
Quick Reference
Generate a Diagram
uv run {baseDir}/scripts/generate.py \
--context "Our framework consists of an encoder module that processes..." \
--caption "Overview of the proposed encoder-decoder architecture"
Or from a file:
uv run {baseDir}/scripts/generate.py \
--input /path/to/method_section.txt \
--caption "Overview of the proposed method"
Options:
--iterations N— refinement rounds (default: 3)--auto-refine— loop until critic is satisfied (use for final quality)--aspect RATIO— aspect ratio:1:1,2:3,3:2,3:4,4:3,9:16,16:9,21:9--provider gemini|openai|openrouter— override auto-detected provider--format png|jpeg|webp— output format (default: png)--no-optimize— disable input optimization (on by default)
Generate a Plot
uv run {baseDir}/scripts/plot.py \
--data '{"model":["GPT-4","Claude","Gemini"],"accuracy":[92.1,94.3,91.8]}' \
--intent "Bar chart comparing model accuracy across benchmarks"
Or from a CSV file:
uv run {baseDir}/scripts/plot.py \
--data-file /path/to/results.csv \
--intent "Line plot showing training loss over epochs"
Evaluate a Diagram
uv run {baseDir}/scripts/evaluate.py \
--generated /path/to/generated.png \
--reference /path/to/human_drawn.png \
--context "The methodology section text..." \
--caption "Overview of the framework"
Returns scores on: Faithfulness, Readability, Conciseness, Aesthetics.
Refine a Previous Diagram
uv run {baseDir}/scripts/generate.py \
--continue \
--feedback "Make the arrows thicker and use more distinct colors"
Or continue a specific run:
uv run {baseDir}/scripts/generate.py \
--continue-run run_20260228_143022_a1b2c3 \
--feedback "Add labels to each component box"
Setup
The skill auto-installs paperbanana on first use via uv (isolated, no global install). The package is published on PyPI by the llmsresearch team.
Required API keys: This skill requires at least one of the following API keys to function. Configure in ~/.openclaw/openclaw.json:
| Env Variable | Provider | Cost | Notes |
|---|---|---|---|
GOOGLE_API_KEY | Google Gemini | Free tier available | Recommended starting point |
OPENAI_API_KEY | OpenAI | Paid | Best quality (gpt-5.2 + gpt-image-1.5) |
OPENROUTER_API_KEY | OpenRouter | Paid | Access to any model |
{
skills: {
entries: {
"paperbanana": {
env: {
// Option A: Google Gemini (free tier — recommended)
GOOGLE_API_KEY: "AIza...",
// Option B: OpenAI (paid, best quality)
// OPENAI_API_KEY: "sk-...",
// Option C: OpenRouter (paid, access to any model)
// OPENROUTER_API_KEY: "sk-or-...",
}
}
}
}
}
Auto-detection priority: Gemini (free) → OpenAI → OpenRouter. The skill will exit with a clear error if no API key is found.
Provider Details
For provider comparison, model options, and advanced configuration:
see {baseDir}/references/providers.md
Privacy & Data Handling
This skill sends user-provided data to external third-party APIs for diagram generation and evaluation:
- Text content (context descriptions, captions, feedback) is sent to the configured LLM provider (Gemini, OpenAI, or OpenRouter) for planning and code generation.
- Generated images may be sent back to the LLM provider for VLM-based evaluation and refinement.
- CSV/JSON data provided for plot generation is sent to the LLM provider for Matplotlib code generation.
Do not use this skill with sensitive, confidential, or proprietary data unless your organization's data policies permit sending that data to the configured provider. All API calls go directly to the provider's endpoints — no intermediate servers are involved.
API keys are injected by OpenClaw from your local config (~/.openclaw/openclaw.json) and are never logged or transmitted beyond the provider's API.
Dependencies & Provenance
- PyPI package:
paperbanana(≥0.1.2, installed automatically viauv) - Source: llmsresearch/paperbanana on GitHub
- Skill source: GoatInAHat/openclaw-paperbanana on GitHub
- Transitive deps:
google-genai,openai,matplotlib,Pillow, and others (installed in an isolateduvenvironment, not globally)
Behavior Notes
- Input optimization is ON by default — enriches context and sharpens captions before generation. Disable with
--no-optimizefor speed. - Generation takes 1-5 minutes depending on iterations and provider. The script prints progress.
- Output is delivered automatically via the MEDIA: protocol — no manual file handling needed.
- Run continuation is the natural way to iterate: "make it better" →
--continue --feedback "...". - Gemini free tier has rate limits (~15 RPM). Keep iterations ≤ 3 on free tier.
Comments
Loading comments...